Import ClosureBench dataset and evaluation artifacts (#1)

- Import ClosureBench dataset and evaluation artifacts (8eab1ca05de983b623e309f13e397ec915947139)

.gitignore

@@ -0,0 +1,4 @@
+.DS_Store
+__pycache__/
+*.pyc
+.cache/

LICENSE

@@ -0,0 +1,7 @@
+ClosureBench dataset
+
+This dataset is released under the Creative Commons Attribution 4.0
+International License (CC BY 4.0).
+
+License summary: https://creativecommons.org/licenses/by/4.0/
+Legal code: https://creativecommons.org/licenses/by/4.0/legalcode

README.md

@@ -1,3 +1,168 @@
 ---
-license: mit
+pretty_name: ClosureBench
+language:
+- en
+license: cc-by-4.0
+task_categories:
+- text-classification
+- question-answering
+tags:
+- benchmark
+- llm-evaluation
+- reasoning
+- knowledge-representation
+- open-world-assumption
+- closed-world-assumption
+- local-closed-world-assumption
+- agent-evaluation
+- symbolic-reasoning
+configs:
+- config_name: base
+  data_files:
+  - split: full
+    path: data/base/full.jsonl
+  - split: train
+    path: data/base/train.jsonl
+  - split: validation
+    path: data/base/validation.jsonl
+  - split: test
+    path: data/base/test.jsonl
+- config_name: ask_act
+  data_files:
+  - split: full
+    path: data/ask_act/full.jsonl
+  - split: train
+    path: data/ask_act/train.jsonl
+  - split: validation
+    path: data/ask_act/validation.jsonl
+  - split: test
+    path: data/ask_act/test.jsonl
+- config_name: multi_agent
+  data_files:
+  - split: full
+    path: data/multi_agent/full.jsonl
+  - split: train
+    path: data/multi_agent/train.jsonl
+  - split: validation
+    path: data/multi_agent/validation.jsonl
+  - split: test
+    path: data/multi_agent/test.jsonl
+- config_name: dynamic_dialogue
+  data_files:
+  - split: full
+    path: data/dynamic_dialogue/full.jsonl
+  - split: train
+    path: data/dynamic_dialogue/train.jsonl
+  - split: validation
+    path: data/dynamic_dialogue/validation.jsonl
+  - split: test
+    path: data/dynamic_dialogue/test.jsonl
 ---
+
+# ClosureBench
+
+ClosureBench is a controlled benchmark for evaluating if LLMs
+respect explicit semantic contracts about missing information. It tests whether
+models distinguish absence-as-unknown, absence-as-false, and
+absence-as-false-only-in-complete-scopes under explicit open-world,
+closed-world, and locally closed-world contracts.
+
+The dataset includes the base benchmark and three extensions:
+
+| Config | `full` rows | Description |
+|---|---:|---|
+| `base` | 960 | Main OWA/CWA/LCWA benchmark with fixed facts, rules, and query across semantic variants. |
+| `ask_act` | 960 | Maps truth values to operational actions: `approve`, `deny`, or `request_information`. |
+| `multi_agent` | 360 | Tests whether a coordinator preserves source-scoped closure. |
+| `dynamic_dialogue` | 100 | Tests whether models update conclusions when complete predicates change across turns. |
+
+Each config exposes four splits:
+
+| Split | Meaning |
+|---|---|
+| `full` | Full split used for the paper's reported metrics. |
+| `train` | Internal benchmark train partition. |
+| `validation` | Internal benchmark development partition. |
+| `test` | Internal benchmark test partition. |
+
+## Loading
+
+```python
+from datasets import load_dataset
+
+base = load_dataset("ML0037/ClosureBench", "base", split="full")
+ask_act = load_dataset("ML0037/ClosureBench", "ask_act", split="full")
+multi_agent = load_dataset("ML0037/ClosureBench", "multi_agent", split="full")
+dynamic = load_dataset("ML0037/ClosureBench", "dynamic_dialogue", split="full")
+```
+
+For the held-out partition only:
+
+```python
+base_test = load_dataset("ML0037/ClosureBench", "base", split="test")
+```
+
+## Base Results
+
+Values are three-run mean +/- sample standard deviation on the `base` config,
+`full` split.
+
+| Model | Semantic switch | Core switch | LCWA closed | LCWA open | Overall |
+|---|---:|---:|---:|---:|---:|
+| Mistral Small | 55.73 +/- 2.43 | 27.60 +/- 3.65 | 31.60 +/- 4.34 | 31.25 +/- 1.80 | 80.38 +/- 1.02 |
+| DeepSeek Flash | 81.25 +/- 0.83 | 68.75 +/- 1.38 | 49.65 +/- 3.18 | 100.00 +/- 0.00 | 90.31 +/- 0.47 |
+| DeepSeek Pro | 86.46 +/- 0.18 | 77.43 +/- 0.30 | 62.15 +/- 1.20 | 100.00 +/- 0.00 | 93.61 +/- 0.21 |
+| Mistral Medium | 87.08 +/- 0.65 | 78.47 +/- 1.09 | 56.95 +/- 2.17 | 100.00 +/- 0.00 | 95.59 +/- 0.37 |
+| Llama Scout | 50.73 +/- 0.48 | 18.23 +/- 1.04 | 4.51 +/- 1.59 | 52.43 +/- 3.01 | 71.18 +/- 0.53 |
+| Llama Maverick | 99.59 +/- 0.18 | 99.31 +/- 0.30 | 99.65 +/- 0.60 | 100.00 +/- 0.00 | 99.86 +/- 0.06 |
+
+Semantic switch accuracy is the primary metric: a base scenario is correct only
+when all semantic variants of that scenario are answered correctly.
+
+## Result Artifacts
+
+The `results/scored/` directory contains final `*_scored.jsonl` files used to
+compute reported metrics. Raw provider response dumps are intentionally not
+included.
+
+The `results/reports/` directory contains JSON manifests and aggregate
+summaries, including:
+
+| File | Purpose |
+|---|---|
+| `results/reports/closurebench_replicate_summary.json` | Base benchmark three-run summary. |
+| `results/reports/closurebench_replicate_manifest.json` | Base benchmark scored-run manifest. |
+| `results/reports/closurebench_model_comparison.json` | Single-run base comparison. |
+| `results/reports/closurebench_ask_act_summary.json` | Ask/Act extension summary. |
+| `results/reports/closurebench_multi_agent_summary.json` | Multi-Agent extension summary. |
+| `results/reports/closurebench_dynamic_dialogue_summary.json` | Dynamic Dialogue extension summary. |
+
+## Data Fields
+
+Common fields include:
+
+- `id`: item identifier.
+- `base_id`: contrastive scenario identifier.
+- `split`: original benchmark partition (`train`, `dev`, or `test`).
+- `domain`, `family`, `subset`: item grouping metadata.
+- `semantics`: semantic contract for base-style items (`owa`, `cwa`, `lcwa`).
+- `closed_predicates`: predicates declared complete for the item.
+- `facts_positive`, `facts_negative`, `rules_natural`: natural-language KB.
+- `symbolic`: symbolic atoms, rules, query atom, and closure atoms.
+- `prompt`: exact prompt used for evaluation.
+- `gold_answer`, `gold_truth_value`, or extension-specific gold fields.
+
+Extension configs add task-specific fields, such as `gold_action` for Ask/Act,
+`gold_source_used` for Multi-Agent, and turn-level gold labels for Dynamic
+Dialogue.
+
+## Limitations
+
+ClosureBench is a targeted diagnostic benchmark. It isolates closure-contract
+compliance under controlled prompts but it is not a broad measure of general agent
+performance, factual knowledge, or end-to-end tool-use reliability.
+
+
+## License
+
+The dataset is released under CC BY 4.0.

data/dynamic_dialogue/test.jsonl

@@ -0,0 +1,12 @@
+{"id": "dynamic_dialogue_v3_0003", "base_id": "cloud_deployment_0250_closed_missing_with_open_distractor", "split": "test", "domain": "cloud_deployment", "family": "closed_missing_with_open_distractor", "subset": "core_contrastive", "gold_pattern": "unknown|false|false", "dialogue_type": "persist_after_narrowing", "target_statement": "Service E has deployment approval.", "query_atom": "has_deployment_approval::Service E", "facts_positive": ["Service F has an incident note.", "Service E passed unit tests."], "facts_negative": [], "rules_natural": [], "predicate_glossary": {"passed_unit_tests": "passed unit tests", "has_deployment_approval": "has deployment approval", "has_incident_note": "has an incident note", "may_deploy": "may deploy to production", "needs_sre_review": "needs SRE review"}, "vocabulary_predicates": ["has_deployment_approval", "has_incident_note", "may_deploy", "needs_sre_review", "passed_unit_tests"], "symbolic": {"positive_atoms": ["has_incident_note::Service F", "passed_unit_tests::Service E"], "negative_atoms": [], "rules": [], "query_atom": "has_deployment_approval::Service E", "closure_atoms": ["has_incident_note::Service F", "passed_unit_tests::Service E"]}, "turns": [{"turn_index": 1, "source_semantics": "owa", "update_operation": "SET_COMPLETE_PREDICATES", "update_predicates": [], "complete_predicates_after_update": [], "gold_answer": "unknown", "expected_changed_from_previous": null, "expected_applied_update": "open", "prompt": "This is turn 1 of a 3-turn dynamic completeness dialogue.\nYou are given an initial knowledge base and an initial completeness state.\nDo not use outside knowledge.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly open. For turn 1, changed_from_previous must be null.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nKnowledge base:\n\nPredicate glossary:\n- has_deployment_approval: has deployment approval\n- has_incident_note: has an incident note\n- may_deploy: may deploy to production\n- needs_sre_review: needs SRE review\n- passed_unit_tests: passed unit tests\n\nPositive facts:\n- Service F has an incident note.\n- Service E passed unit tests.\n\nNegative facts:\n- none\n\nRules:\n- none\n\nTarget statement: Service E has deployment approval.\nTarget atom: has_deployment_approval::Service E\n\nCompleteness-state update for this turn:\n- Operation: SET_COMPLETE_PREDICATES\n- New complete predicates: none"}, {"turn_index": 2, "source_semantics": "cwa", "update_operation": "ADD_COMPLETE_PREDICATES", "update_predicates": ["has_deployment_approval", "has_incident_note", "may_deploy", "needs_sre_review", "passed_unit_tests"], "complete_predicates_after_update": ["has_deployment_approval", "has_incident_note", "may_deploy", "needs_sre_review", "passed_unit_tests"], "gold_answer": "false", "expected_changed_from_previous": true, "expected_applied_update": "global_complete", "prompt": "This is turn 2 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly global_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: ADD_COMPLETE_PREDICATES\n- Predicates to add as complete: has_deployment_approval, has_incident_note, may_deploy, needs_sre_review, passed_unit_tests\n- Complete predicates after this update: has_deployment_approval, has_incident_note, may_deploy, needs_sre_review, passed_unit_tests\n\nTarget statement: Service E has deployment approval.\nTarget atom: has_deployment_approval::Service E"}, {"turn_index": 3, "source_semantics": "lcwa", "update_operation": "REPLACE_COMPLETE_PREDICATES", "update_predicates": ["has_deployment_approval"], "complete_predicates_after_update": ["has_deployment_approval"], "gold_answer": "false", "expected_changed_from_previous": false, "expected_applied_update": "local_complete", "prompt": "This is turn 3 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly local_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: REPLACE_COMPLETE_PREDICATES\n- Discard the previous complete-predicate set.\n- Replacement complete predicates: has_deployment_approval\n\nTarget statement: Service E has deployment approval.\nTarget atom: has_deployment_approval::Service E"}]}
+{"id": "dynamic_dialogue_v3_0005", "base_id": "cloud_deployment_0254_open_missing_with_closed_distractor", "split": "test", "domain": "cloud_deployment", "family": "open_missing_with_closed_distractor", "subset": "core_contrastive", "gold_pattern": "unknown|false|unknown", "dialogue_type": "reopen", "target_statement": "Service C has an incident note.", "query_atom": "has_incident_note::Service C", "facts_positive": ["Service D has deployment approval.", "Service D passed unit tests."], "facts_negative": [], "rules_natural": [], "predicate_glossary": {"passed_unit_tests": "passed unit tests", "has_deployment_approval": "has deployment approval", "has_incident_note": "has an incident note", "may_deploy": "may deploy to production", "needs_sre_review": "needs SRE review"}, "vocabulary_predicates": ["has_deployment_approval", "has_incident_note", "may_deploy", "needs_sre_review", "passed_unit_tests"], "symbolic": {"positive_atoms": ["has_deployment_approval::Service D", "passed_unit_tests::Service D"], "negative_atoms": [], "rules": [], "query_atom": "has_incident_note::Service C", "closure_atoms": ["has_deployment_approval::Service D", "passed_unit_tests::Service D"]}, "turns": [{"turn_index": 1, "source_semantics": "owa", "update_operation": "SET_COMPLETE_PREDICATES", "update_predicates": [], "complete_predicates_after_update": [], "gold_answer": "unknown", "expected_changed_from_previous": null, "expected_applied_update": "open", "prompt": "This is turn 1 of a 3-turn dynamic completeness dialogue.\nYou are given an initial knowledge base and an initial completeness state.\nDo not use outside knowledge.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly open. For turn 1, changed_from_previous must be null.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nKnowledge base:\n\nPredicate glossary:\n- has_deployment_approval: has deployment approval\n- has_incident_note: has an incident note\n- may_deploy: may deploy to production\n- needs_sre_review: needs SRE review\n- passed_unit_tests: passed unit tests\n\nPositive facts:\n- Service D has deployment approval.\n- Service D passed unit tests.\n\nNegative facts:\n- none\n\nRules:\n- none\n\nTarget statement: Service C has an incident note.\nTarget atom: has_incident_note::Service C\n\nCompleteness-state update for this turn:\n- Operation: SET_COMPLETE_PREDICATES\n- New complete predicates: none"}, {"turn_index": 2, "source_semantics": "cwa", "update_operation": "ADD_COMPLETE_PREDICATES", "update_predicates": ["has_deployment_approval", "has_incident_note", "may_deploy", "needs_sre_review", "passed_unit_tests"], "complete_predicates_after_update": ["has_deployment_approval", "has_incident_note", "may_deploy", "needs_sre_review", "passed_unit_tests"], "gold_answer": "false", "expected_changed_from_previous": true, "expected_applied_update": "global_complete", "prompt": "This is turn 2 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly global_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: ADD_COMPLETE_PREDICATES\n- Predicates to add as complete: has_deployment_approval, has_incident_note, may_deploy, needs_sre_review, passed_unit_tests\n- Complete predicates after this update: has_deployment_approval, has_incident_note, may_deploy, needs_sre_review, passed_unit_tests\n\nTarget statement: Service C has an incident note.\nTarget atom: has_incident_note::Service C"}, {"turn_index": 3, "source_semantics": "lcwa", "update_operation": "REPLACE_COMPLETE_PREDICATES", "update_predicates": ["has_deployment_approval", "may_deploy"], "complete_predicates_after_update": ["has_deployment_approval", "may_deploy"], "gold_answer": "unknown", "expected_changed_from_previous": true, "expected_applied_update": "local_complete", "prompt": "This is turn 3 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly local_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: REPLACE_COMPLETE_PREDICATES\n- Discard the previous complete-predicate set.\n- Replacement complete predicates: has_deployment_approval, may_deploy\n\nTarget statement: Service C has an incident note.\nTarget atom: has_incident_note::Service C"}]}
+{"id": "dynamic_dialogue_v3_0016", "base_id": "finance_controls_0136_closed_derived_missing_antecedent", "split": "test", "domain": "finance_controls", "family": "closed_derived_missing_antecedent", "subset": "core_contrastive", "gold_pattern": "unknown|false|false", "dialogue_type": "persist_after_narrowing", "target_statement": "Account 59 may receive a wire transfer.", "query_atom": "may_receive_wire::Account 59", "facts_positive": ["Account 59 passed KYC checks."], "facts_negative": [], "rules_natural": ["If an account passed KYC checks and has compliance clearance, then that account may receive a wire transfer."], "predicate_glossary": {"passed_kyc": "passed KYC checks", "has_compliance_clearance": "has compliance clearance", "has_manual_exception": "has a manual exception", "may_receive_wire": "may receive a wire transfer", "needs_compliance_review": "needs compliance review"}, "vocabulary_predicates": ["has_compliance_clearance", "has_manual_exception", "may_receive_wire", "needs_compliance_review", "passed_kyc"], "symbolic": {"positive_atoms": ["passed_kyc::Account 59"], "negative_atoms": [], "rules": [{"antecedents": ["passed_kyc::Account 59", "has_compliance_clearance::Account 59"], "conclusion": "may_receive_wire::Account 59", "text": "If an account passed KYC checks and has compliance clearance, then that account may receive a wire transfer."}], "query_atom": "may_receive_wire::Account 59", "closure_atoms": ["passed_kyc::Account 59"]}, "turns": [{"turn_index": 1, "source_semantics": "owa", "update_operation": "SET_COMPLETE_PREDICATES", "update_predicates": [], "complete_predicates_after_update": [], "gold_answer": "unknown", "expected_changed_from_previous": null, "expected_applied_update": "open", "prompt": "This is turn 1 of a 3-turn dynamic completeness dialogue.\nYou are given an initial knowledge base and an initial completeness state.\nDo not use outside knowledge.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly open. For turn 1, changed_from_previous must be null.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nKnowledge base:\n\nPredicate glossary:\n- has_compliance_clearance: has compliance clearance\n- has_manual_exception: has a manual exception\n- may_receive_wire: may receive a wire transfer\n- needs_compliance_review: needs compliance review\n- passed_kyc: passed KYC checks\n\nPositive facts:\n- Account 59 passed KYC checks.\n\nNegative facts:\n- none\n\nRules:\n- If an account passed KYC checks and has compliance clearance, then that account may receive a wire transfer.\n\nTarget statement: Account 59 may receive a wire transfer.\nTarget atom: may_receive_wire::Account 59\n\nCompleteness-state update for this turn:\n- Operation: SET_COMPLETE_PREDICATES\n- New complete predicates: none"}, {"turn_index": 2, "source_semantics": "cwa", "update_operation": "ADD_COMPLETE_PREDICATES", "update_predicates": ["has_compliance_clearance", "has_manual_exception", "may_receive_wire", "needs_compliance_review", "passed_kyc"], "complete_predicates_after_update": ["has_compliance_clearance", "has_manual_exception", "may_receive_wire", "needs_compliance_review", "passed_kyc"], "gold_answer": "false", "expected_changed_from_previous": true, "expected_applied_update": "global_complete", "prompt": "This is turn 2 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly global_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: ADD_COMPLETE_PREDICATES\n- Predicates to add as complete: has_compliance_clearance, has_manual_exception, may_receive_wire, needs_compliance_review, passed_kyc\n- Complete predicates after this update: has_compliance_clearance, has_manual_exception, may_receive_wire, needs_compliance_review, passed_kyc\n\nTarget statement: Account 59 may receive a wire transfer.\nTarget atom: may_receive_wire::Account 59"}, {"turn_index": 3, "source_semantics": "lcwa", "update_operation": "REPLACE_COMPLETE_PREDICATES", "update_predicates": ["has_compliance_clearance", "may_receive_wire"], "complete_predicates_after_update": ["has_compliance_clearance", "may_receive_wire"], "gold_answer": "false", "expected_changed_from_previous": false, "expected_applied_update": "local_complete", "prompt": "This is turn 3 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly local_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: REPLACE_COMPLETE_PREDICATES\n- Discard the previous complete-predicate set.\n- Replacement complete predicates: has_compliance_clearance, may_receive_wire\n\nTarget statement: Account 59 may receive a wire transfer.\nTarget atom: may_receive_wire::Account 59"}]}
+{"id": "dynamic_dialogue_v3_0026", "base_id": "hospital_access_0008_closed_missing_with_open_distractor", "split": "test", "domain": "hospital_access", "family": "closed_missing_with_open_distractor", "subset": "core_contrastive", "gold_pattern": "unknown|false|false", "dialogue_type": "persist_after_narrowing", "target_statement": "Cora has security clearance.", "query_atom": "has_clearance::Cora", "facts_positive": ["Dylan is assigned to the clinical trial.", "Cora has completed safety training."], "facts_negative": [], "rules_natural": [], "predicate_glossary": {"completed_training": "has completed safety training", "has_clearance": "has security clearance", "assigned_to_trial": "is assigned to the clinical trial", "may_enter_lab": "may enter Lab A", "needs_supervisor_review": "needs supervisor review"}, "vocabulary_predicates": ["assigned_to_trial", "completed_training", "has_clearance", "may_enter_lab", "needs_supervisor_review"], "symbolic": {"positive_atoms": ["assigned_to_trial::Dylan", "completed_training::Cora"], "negative_atoms": [], "rules": [], "query_atom": "has_clearance::Cora", "closure_atoms": ["assigned_to_trial::Dylan", "completed_training::Cora"]}, "turns": [{"turn_index": 1, "source_semantics": "owa", "update_operation": "SET_COMPLETE_PREDICATES", "update_predicates": [], "complete_predicates_after_update": [], "gold_answer": "unknown", "expected_changed_from_previous": null, "expected_applied_update": "open", "prompt": "This is turn 1 of a 3-turn dynamic completeness dialogue.\nYou are given an initial knowledge base and an initial completeness state.\nDo not use outside knowledge.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly open. For turn 1, changed_from_previous must be null.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nKnowledge base:\n\nPredicate glossary:\n- assigned_to_trial: is assigned to the clinical trial\n- completed_training: has completed safety training\n- has_clearance: has security clearance\n- may_enter_lab: may enter Lab A\n- needs_supervisor_review: needs supervisor review\n\nPositive facts:\n- Dylan is assigned to the clinical trial.\n- Cora has completed safety training.\n\nNegative facts:\n- none\n\nRules:\n- none\n\nTarget statement: Cora has security clearance.\nTarget atom: has_clearance::Cora\n\nCompleteness-state update for this turn:\n- Operation: SET_COMPLETE_PREDICATES\n- New complete predicates: none"}, {"turn_index": 2, "source_semantics": "cwa", "update_operation": "ADD_COMPLETE_PREDICATES", "update_predicates": ["assigned_to_trial", "completed_training", "has_clearance", "may_enter_lab", "needs_supervisor_review"], "complete_predicates_after_update": ["assigned_to_trial", "completed_training", "has_clearance", "may_enter_lab", "needs_supervisor_review"], "gold_answer": "false", "expected_changed_from_previous": true, "expected_applied_update": "global_complete", "prompt": "This is turn 2 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly global_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: ADD_COMPLETE_PREDICATES\n- Predicates to add as complete: assigned_to_trial, completed_training, has_clearance, may_enter_lab, needs_supervisor_review\n- Complete predicates after this update: assigned_to_trial, completed_training, has_clearance, may_enter_lab, needs_supervisor_review\n\nTarget statement: Cora has security clearance.\nTarget atom: has_clearance::Cora"}, {"turn_index": 3, "source_semantics": "lcwa", "update_operation": "REPLACE_COMPLETE_PREDICATES", "update_predicates": ["has_clearance"], "complete_predicates_after_update": ["has_clearance"], "gold_answer": "false", "expected_changed_from_previous": false, "expected_applied_update": "local_complete", "prompt": "This is turn 3 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly local_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: REPLACE_COMPLETE_PREDICATES\n- Discard the previous complete-predicate set.\n- Replacement complete predicates: has_clearance\n\nTarget statement: Cora has security clearance.\nTarget atom: has_clearance::Cora"}]}
+{"id": "dynamic_dialogue_v3_0027", "base_id": "hospital_access_0009_closed_missing_with_open_distractor", "split": "test", "domain": "hospital_access", "family": "closed_missing_with_open_distractor", "subset": "core_contrastive", "gold_pattern": "unknown|false|false", "dialogue_type": "persist_after_narrowing", "target_statement": "Dylan has security clearance.", "query_atom": "has_clearance::Dylan", "facts_positive": ["Elena is assigned to the clinical trial.", "Dylan has completed safety training."], "facts_negative": [], "rules_natural": [], "predicate_glossary": {"completed_training": "has completed safety training", "has_clearance": "has security clearance", "assigned_to_trial": "is assigned to the clinical trial", "may_enter_lab": "may enter Lab A", "needs_supervisor_review": "needs supervisor review"}, "vocabulary_predicates": ["assigned_to_trial", "completed_training", "has_clearance", "may_enter_lab", "needs_supervisor_review"], "symbolic": {"positive_atoms": ["assigned_to_trial::Elena", "completed_training::Dylan"], "negative_atoms": [], "rules": [], "query_atom": "has_clearance::Dylan", "closure_atoms": ["assigned_to_trial::Elena", "completed_training::Dylan"]}, "turns": [{"turn_index": 1, "source_semantics": "owa", "update_operation": "SET_COMPLETE_PREDICATES", "update_predicates": [], "complete_predicates_after_update": [], "gold_answer": "unknown", "expected_changed_from_previous": null, "expected_applied_update": "open", "prompt": "This is turn 1 of a 3-turn dynamic completeness dialogue.\nYou are given an initial knowledge base and an initial completeness state.\nDo not use outside knowledge.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly open. For turn 1, changed_from_previous must be null.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nKnowledge base:\n\nPredicate glossary:\n- assigned_to_trial: is assigned to the clinical trial\n- completed_training: has completed safety training\n- has_clearance: has security clearance\n- may_enter_lab: may enter Lab A\n- needs_supervisor_review: needs supervisor review\n\nPositive facts:\n- Elena is assigned to the clinical trial.\n- Dylan has completed safety training.\n\nNegative facts:\n- none\n\nRules:\n- none\n\nTarget statement: Dylan has security clearance.\nTarget atom: has_clearance::Dylan\n\nCompleteness-state update for this turn:\n- Operation: SET_COMPLETE_PREDICATES\n- New complete predicates: none"}, {"turn_index": 2, "source_semantics": "cwa", "update_operation": "ADD_COMPLETE_PREDICATES", "update_predicates": ["assigned_to_trial", "completed_training", "has_clearance", "may_enter_lab", "needs_supervisor_review"], "complete_predicates_after_update": ["assigned_to_trial", "completed_training", "has_clearance", "may_enter_lab", "needs_supervisor_review"], "gold_answer": "false", "expected_changed_from_previous": true, "expected_applied_update": "global_complete", "prompt": "This is turn 2 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly global_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: ADD_COMPLETE_PREDICATES\n- Predicates to add as complete: assigned_to_trial, completed_training, has_clearance, may_enter_lab, needs_supervisor_review\n- Complete predicates after this update: assigned_to_trial, completed_training, has_clearance, may_enter_lab, needs_supervisor_review\n\nTarget statement: Dylan has security clearance.\nTarget atom: has_clearance::Dylan"}, {"turn_index": 3, "source_semantics": "lcwa", "update_operation": "REPLACE_COMPLETE_PREDICATES", "update_predicates": ["has_clearance"], "complete_predicates_after_update": ["has_clearance"], "gold_answer": "false", "expected_changed_from_previous": false, "expected_applied_update": "local_complete", "prompt": "This is turn 3 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly local_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: REPLACE_COMPLETE_PREDICATES\n- Discard the previous complete-predicate set.\n- Replacement complete predicates: has_clearance\n\nTarget statement: Dylan has security clearance.\nTarget atom: has_clearance::Dylan"}]}
+{"id": "dynamic_dialogue_v3_0033", "base_id": "hospital_access_0019_closed_derived_missing_antecedent", "split": "test", "domain": "hospital_access", "family": "closed_derived_missing_antecedent", "subset": "core_contrastive", "gold_pattern": "unknown|false|false", "dialogue_type": "persist_after_narrowing", "target_statement": "Ben may enter Lab A.", "query_atom": "may_enter_lab::Ben", "facts_positive": ["Ben has completed safety training."], "facts_negative": [], "rules_natural": ["If a staff member has security clearance and has completed safety training, then that staff member may enter Lab A."], "predicate_glossary": {"completed_training": "has completed safety training", "has_clearance": "has security clearance", "assigned_to_trial": "is assigned to the clinical trial", "may_enter_lab": "may enter Lab A", "needs_supervisor_review": "needs supervisor review"}, "vocabulary_predicates": ["assigned_to_trial", "completed_training", "has_clearance", "may_enter_lab", "needs_supervisor_review"], "symbolic": {"positive_atoms": ["completed_training::Ben"], "negative_atoms": [], "rules": [{"antecedents": ["completed_training::Ben", "has_clearance::Ben"], "conclusion": "may_enter_lab::Ben", "text": "If a staff member has security clearance and has completed safety training, then that staff member may enter Lab A."}], "query_atom": "may_enter_lab::Ben", "closure_atoms": ["completed_training::Ben"]}, "turns": [{"turn_index": 1, "source_semantics": "owa", "update_operation": "SET_COMPLETE_PREDICATES", "update_predicates": [], "complete_predicates_after_update": [], "gold_answer": "unknown", "expected_changed_from_previous": null, "expected_applied_update": "open", "prompt": "This is turn 1 of a 3-turn dynamic completeness dialogue.\nYou are given an initial knowledge base and an initial completeness state.\nDo not use outside knowledge.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly open. For turn 1, changed_from_previous must be null.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nKnowledge base:\n\nPredicate glossary:\n- assigned_to_trial: is assigned to the clinical trial\n- completed_training: has completed safety training\n- has_clearance: has security clearance\n- may_enter_lab: may enter Lab A\n- needs_supervisor_review: needs supervisor review\n\nPositive facts:\n- Ben has completed safety training.\n\nNegative facts:\n- none\n\nRules:\n- If a staff member has security clearance and has completed safety training, then that staff member may enter Lab A.\n\nTarget statement: Ben may enter Lab A.\nTarget atom: may_enter_lab::Ben\n\nCompleteness-state update for this turn:\n- Operation: SET_COMPLETE_PREDICATES\n- New complete predicates: none"}, {"turn_index": 2, "source_semantics": "cwa", "update_operation": "ADD_COMPLETE_PREDICATES", "update_predicates": ["assigned_to_trial", "completed_training", "has_clearance", "may_enter_lab", "needs_supervisor_review"], "complete_predicates_after_update": ["assigned_to_trial", "completed_training", "has_clearance", "may_enter_lab", "needs_supervisor_review"], "gold_answer": "false", "expected_changed_from_previous": true, "expected_applied_update": "global_complete", "prompt": "This is turn 2 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly global_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: ADD_COMPLETE_PREDICATES\n- Predicates to add as complete: assigned_to_trial, completed_training, has_clearance, may_enter_lab, needs_supervisor_review\n- Complete predicates after this update: assigned_to_trial, completed_training, has_clearance, may_enter_lab, needs_supervisor_review\n\nTarget statement: Ben may enter Lab A.\nTarget atom: may_enter_lab::Ben"}, {"turn_index": 3, "source_semantics": "lcwa", "update_operation": "REPLACE_COMPLETE_PREDICATES", "update_predicates": ["has_clearance", "may_enter_lab"], "complete_predicates_after_update": ["has_clearance", "may_enter_lab"], "gold_answer": "false", "expected_changed_from_previous": false, "expected_applied_update": "local_complete", "prompt": "This is turn 3 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly local_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: REPLACE_COMPLETE_PREDICATES\n- Discard the previous complete-predicate set.\n- Replacement complete predicates: has_clearance, may_enter_lab\n\nTarget statement: Ben may enter Lab A.\nTarget atom: may_enter_lab::Ben"}]}
+{"id": "dynamic_dialogue_v3_0040", "base_id": "library_services_0206_open_missing_direct", "split": "test", "domain": "library_services", "family": "open_missing_direct", "subset": "core_contrastive", "gold_pattern": "unknown|false|unknown", "dialogue_type": "reopen", "target_statement": "Reader C has a special request.", "query_atom": "has_special_request::Reader C", "facts_positive": ["Reader C has an active membership."], "facts_negative": [], "rules_natural": [], "predicate_glossary": {"has_active_membership": "has an active membership", "has_borrowing_clearance": "has borrowing clearance", "has_special_request": "has a special request", "may_borrow_archive_item": "may borrow an archive item", "needs_librarian_review": "needs librarian review"}, "vocabulary_predicates": ["has_active_membership", "has_borrowing_clearance", "has_special_request", "may_borrow_archive_item", "needs_librarian_review"], "symbolic": {"positive_atoms": ["has_active_membership::Reader C"], "negative_atoms": [], "rules": [], "query_atom": "has_special_request::Reader C", "closure_atoms": ["has_active_membership::Reader C"]}, "turns": [{"turn_index": 1, "source_semantics": "owa", "update_operation": "SET_COMPLETE_PREDICATES", "update_predicates": [], "complete_predicates_after_update": [], "gold_answer": "unknown", "expected_changed_from_previous": null, "expected_applied_update": "open", "prompt": "This is turn 1 of a 3-turn dynamic completeness dialogue.\nYou are given an initial knowledge base and an initial completeness state.\nDo not use outside knowledge.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly open. For turn 1, changed_from_previous must be null.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nKnowledge base:\n\nPredicate glossary:\n- has_active_membership: has an active membership\n- has_borrowing_clearance: has borrowing clearance\n- has_special_request: has a special request\n- may_borrow_archive_item: may borrow an archive item\n- needs_librarian_review: needs librarian review\n\nPositive facts:\n- Reader C has an active membership.\n\nNegative facts:\n- none\n\nRules:\n- none\n\nTarget statement: Reader C has a special request.\nTarget atom: has_special_request::Reader C\n\nCompleteness-state update for this turn:\n- Operation: SET_COMPLETE_PREDICATES\n- New complete predicates: none"}, {"turn_index": 2, "source_semantics": "cwa", "update_operation": "ADD_COMPLETE_PREDICATES", "update_predicates": ["has_active_membership", "has_borrowing_clearance", "has_special_request", "may_borrow_archive_item", "needs_librarian_review"], "complete_predicates_after_update": ["has_active_membership", "has_borrowing_clearance", "has_special_request", "may_borrow_archive_item", "needs_librarian_review"], "gold_answer": "false", "expected_changed_from_previous": true, "expected_applied_update": "global_complete", "prompt": "This is turn 2 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly global_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: ADD_COMPLETE_PREDICATES\n- Predicates to add as complete: has_active_membership, has_borrowing_clearance, has_special_request, may_borrow_archive_item, needs_librarian_review\n- Complete predicates after this update: has_active_membership, has_borrowing_clearance, has_special_request, may_borrow_archive_item, needs_librarian_review\n\nTarget statement: Reader C has a special request.\nTarget atom: has_special_request::Reader C"}, {"turn_index": 3, "source_semantics": "lcwa", "update_operation": "REPLACE_COMPLETE_PREDICATES", "update_predicates": ["has_borrowing_clearance", "may_borrow_archive_item"], "complete_predicates_after_update": ["has_borrowing_clearance", "may_borrow_archive_item"], "gold_answer": "unknown", "expected_changed_from_previous": true, "expected_applied_update": "local_complete", "prompt": "This is turn 3 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly local_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: REPLACE_COMPLETE_PREDICATES\n- Discard the previous complete-predicate set.\n- Replacement complete predicates: has_borrowing_clearance, may_borrow_archive_item\n\nTarget statement: Reader C has a special request.\nTarget atom: has_special_request::Reader C"}]}
+{"id": "dynamic_dialogue_v3_0045", "base_id": "library_services_0218_closed_derived_missing_antecedent", "split": "test", "domain": "library_services", "family": "closed_derived_missing_antecedent", "subset": "core_contrastive", "gold_pattern": "unknown|false|false", "dialogue_type": "persist_after_narrowing", "target_statement": "Reader C may borrow an archive item.", "query_atom": "may_borrow_archive_item::Reader C", "facts_positive": ["Reader C has an active membership."], "facts_negative": [], "rules_natural": ["If a reader has an active membership and has borrowing clearance, then that reader may borrow an archive item."], "predicate_glossary": {"has_active_membership": "has an active membership", "has_borrowing_clearance": "has borrowing clearance", "has_special_request": "has a special request", "may_borrow_archive_item": "may borrow an archive item", "needs_librarian_review": "needs librarian review"}, "vocabulary_predicates": ["has_active_membership", "has_borrowing_clearance", "has_special_request", "may_borrow_archive_item", "needs_librarian_review"], "symbolic": {"positive_atoms": ["has_active_membership::Reader C"], "negative_atoms": [], "rules": [{"antecedents": ["has_active_membership::Reader C", "has_borrowing_clearance::Reader C"], "conclusion": "may_borrow_archive_item::Reader C", "text": "If a reader has an active membership and has borrowing clearance, then that reader may borrow an archive item."}], "query_atom": "may_borrow_archive_item::Reader C", "closure_atoms": ["has_active_membership::Reader C"]}, "turns": [{"turn_index": 1, "source_semantics": "owa", "update_operation": "SET_COMPLETE_PREDICATES", "update_predicates": [], "complete_predicates_after_update": [], "gold_answer": "unknown", "expected_changed_from_previous": null, "expected_applied_update": "open", "prompt": "This is turn 1 of a 3-turn dynamic completeness dialogue.\nYou are given an initial knowledge base and an initial completeness state.\nDo not use outside knowledge.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly open. For turn 1, changed_from_previous must be null.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nKnowledge base:\n\nPredicate glossary:\n- has_active_membership: has an active membership\n- has_borrowing_clearance: has borrowing clearance\n- has_special_request: has a special request\n- may_borrow_archive_item: may borrow an archive item\n- needs_librarian_review: needs librarian review\n\nPositive facts:\n- Reader C has an active membership.\n\nNegative facts:\n- none\n\nRules:\n- If a reader has an active membership and has borrowing clearance, then that reader may borrow an archive item.\n\nTarget statement: Reader C may borrow an archive item.\nTarget atom: may_borrow_archive_item::Reader C\n\nCompleteness-state update for this turn:\n- Operation: SET_COMPLETE_PREDICATES\n- New complete predicates: none"}, {"turn_index": 2, "source_semantics": "cwa", "update_operation": "ADD_COMPLETE_PREDICATES", "update_predicates": ["has_active_membership", "has_borrowing_clearance", "has_special_request", "may_borrow_archive_item", "needs_librarian_review"], "complete_predicates_after_update": ["has_active_membership", "has_borrowing_clearance", "has_special_request", "may_borrow_archive_item", "needs_librarian_review"], "gold_answer": "false", "expected_changed_from_previous": true, "expected_applied_update": "global_complete", "prompt": "This is turn 2 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly global_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: ADD_COMPLETE_PREDICATES\n- Predicates to add as complete: has_active_membership, has_borrowing_clearance, has_special_request, may_borrow_archive_item, needs_librarian_review\n- Complete predicates after this update: has_active_membership, has_borrowing_clearance, has_special_request, may_borrow_archive_item, needs_librarian_review\n\nTarget statement: Reader C may borrow an archive item.\nTarget atom: may_borrow_archive_item::Reader C"}, {"turn_index": 3, "source_semantics": "lcwa", "update_operation": "REPLACE_COMPLETE_PREDICATES", "update_predicates": ["has_borrowing_clearance", "may_borrow_archive_item"], "complete_predicates_after_update": ["has_borrowing_clearance", "may_borrow_archive_item"], "gold_answer": "false", "expected_changed_from_previous": false, "expected_applied_update": "local_complete", "prompt": "This is turn 3 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly local_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: REPLACE_COMPLETE_PREDICATES\n- Discard the previous complete-predicate set.\n- Replacement complete predicates: has_borrowing_clearance, may_borrow_archive_item\n\nTarget statement: Reader C may borrow an archive item.\nTarget atom: may_borrow_archive_item::Reader C"}]}
+{"id": "dynamic_dialogue_v3_0047", "base_id": "manufacturing_quality_0080_closed_missing_direct", "split": "test", "domain": "manufacturing_quality", "family": "closed_missing_direct", "subset": "core_contrastive", "gold_pattern": "unknown|false|false", "dialogue_type": "persist_after_narrowing", "target_statement": "Part C has release approval.", "query_atom": "has_release_approval::Part C", "facts_positive": ["Part C passed the visual check."], "facts_negative": [], "rules_natural": [], "predicate_glossary": {"passed_visual_check": "passed the visual check", "has_release_approval": "has release approval", "has_supplier_note": "has a supplier note", "may_ship": "may ship to customers", "needs_engineer_review": "needs engineer review"}, "vocabulary_predicates": ["has_release_approval", "has_supplier_note", "may_ship", "needs_engineer_review", "passed_visual_check"], "symbolic": {"positive_atoms": ["passed_visual_check::Part C"], "negative_atoms": [], "rules": [], "query_atom": "has_release_approval::Part C", "closure_atoms": ["passed_visual_check::Part C"]}, "turns": [{"turn_index": 1, "source_semantics": "owa", "update_operation": "SET_COMPLETE_PREDICATES", "update_predicates": [], "complete_predicates_after_update": [], "gold_answer": "unknown", "expected_changed_from_previous": null, "expected_applied_update": "open", "prompt": "This is turn 1 of a 3-turn dynamic completeness dialogue.\nYou are given an initial knowledge base and an initial completeness state.\nDo not use outside knowledge.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly open. For turn 1, changed_from_previous must be null.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nKnowledge base:\n\nPredicate glossary:\n- has_release_approval: has release approval\n- has_supplier_note: has a supplier note\n- may_ship: may ship to customers\n- needs_engineer_review: needs engineer review\n- passed_visual_check: passed the visual check\n\nPositive facts:\n- Part C passed the visual check.\n\nNegative facts:\n- none\n\nRules:\n- none\n\nTarget statement: Part C has release approval.\nTarget atom: has_release_approval::Part C\n\nCompleteness-state update for this turn:\n- Operation: SET_COMPLETE_PREDICATES\n- New complete predicates: none"}, {"turn_index": 2, "source_semantics": "cwa", "update_operation": "ADD_COMPLETE_PREDICATES", "update_predicates": ["has_release_approval", "has_supplier_note", "may_ship", "needs_engineer_review", "passed_visual_check"], "complete_predicates_after_update": ["has_release_approval", "has_supplier_note", "may_ship", "needs_engineer_review", "passed_visual_check"], "gold_answer": "false", "expected_changed_from_previous": true, "expected_applied_update": "global_complete", "prompt": "This is turn 2 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly global_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: ADD_COMPLETE_PREDICATES\n- Predicates to add as complete: has_release_approval, has_supplier_note, may_ship, needs_engineer_review, passed_visual_check\n- Complete predicates after this update: has_release_approval, has_supplier_note, may_ship, needs_engineer_review, passed_visual_check\n\nTarget statement: Part C has release approval.\nTarget atom: has_release_approval::Part C"}, {"turn_index": 3, "source_semantics": "lcwa", "update_operation": "REPLACE_COMPLETE_PREDICATES", "update_predicates": ["has_release_approval"], "complete_predicates_after_update": ["has_release_approval"], "gold_answer": "false", "expected_changed_from_previous": false, "expected_applied_update": "local_complete", "prompt": "This is turn 3 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly local_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: REPLACE_COMPLETE_PREDICATES\n- Discard the previous complete-predicate set.\n- Replacement complete predicates: has_release_approval\n\nTarget statement: Part C has release approval.\nTarget atom: has_release_approval::Part C"}]}
+{"id": "dynamic_dialogue_v3_0057", "base_id": "manufacturing_quality_0101_open_derived_missing_antecedent", "split": "test", "domain": "manufacturing_quality", "family": "open_derived_missing_antecedent", "subset": "core_contrastive", "gold_pattern": "unknown|false|unknown", "dialogue_type": "reopen", "target_statement": "Part F needs engineer review.", "query_atom": "needs_engineer_review::Part F", "facts_positive": ["Part F passed the visual check."], "facts_negative": [], "rules_natural": ["If a part has a supplier note, then that part needs engineer review."], "predicate_glossary": {"passed_visual_check": "passed the visual check", "has_release_approval": "has release approval", "has_supplier_note": "has a supplier note", "may_ship": "may ship to customers", "needs_engineer_review": "needs engineer review"}, "vocabulary_predicates": ["has_release_approval", "has_supplier_note", "may_ship", "needs_engineer_review", "passed_visual_check"], "symbolic": {"positive_atoms": ["passed_visual_check::Part F"], "negative_atoms": [], "rules": [{"antecedents": ["has_supplier_note::Part F"], "conclusion": "needs_engineer_review::Part F", "text": "If a part has a supplier note, then that part needs engineer review."}], "query_atom": "needs_engineer_review::Part F", "closure_atoms": ["passed_visual_check::Part F"]}, "turns": [{"turn_index": 1, "source_semantics": "owa", "update_operation": "SET_COMPLETE_PREDICATES", "update_predicates": [], "complete_predicates_after_update": [], "gold_answer": "unknown", "expected_changed_from_previous": null, "expected_applied_update": "open", "prompt": "This is turn 1 of a 3-turn dynamic completeness dialogue.\nYou are given an initial knowledge base and an initial completeness state.\nDo not use outside knowledge.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly open. For turn 1, changed_from_previous must be null.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nKnowledge base:\n\nPredicate glossary:\n- has_release_approval: has release approval\n- has_supplier_note: has a supplier note\n- may_ship: may ship to customers\n- needs_engineer_review: needs engineer review\n- passed_visual_check: passed the visual check\n\nPositive facts:\n- Part F passed the visual check.\n\nNegative facts:\n- none\n\nRules:\n- If a part has a supplier note, then that part needs engineer review.\n\nTarget statement: Part F needs engineer review.\nTarget atom: needs_engineer_review::Part F\n\nCompleteness-state update for this turn:\n- Operation: SET_COMPLETE_PREDICATES\n- New complete predicates: none"}, {"turn_index": 2, "source_semantics": "cwa", "update_operation": "ADD_COMPLETE_PREDICATES", "update_predicates": ["has_release_approval", "has_supplier_note", "may_ship", "needs_engineer_review", "passed_visual_check"], "complete_predicates_after_update": ["has_release_approval", "has_supplier_note", "may_ship", "needs_engineer_review", "passed_visual_check"], "gold_answer": "false", "expected_changed_from_previous": true, "expected_applied_update": "global_complete", "prompt": "This is turn 2 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly global_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: ADD_COMPLETE_PREDICATES\n- Predicates to add as complete: has_release_approval, has_supplier_note, may_ship, needs_engineer_review, passed_visual_check\n- Complete predicates after this update: has_release_approval, has_supplier_note, may_ship, needs_engineer_review, passed_visual_check\n\nTarget statement: Part F needs engineer review.\nTarget atom: needs_engineer_review::Part F"}, {"turn_index": 3, "source_semantics": "lcwa", "update_operation": "REPLACE_COMPLETE_PREDICATES", "update_predicates": ["has_release_approval", "may_ship"], "complete_predicates_after_update": ["has_release_approval", "may_ship"], "gold_answer": "unknown", "expected_changed_from_previous": true, "expected_applied_update": "local_complete", "prompt": "This is turn 3 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly local_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: REPLACE_COMPLETE_PREDICATES\n- Discard the previous complete-predicate set.\n- Replacement complete predicates: has_release_approval, may_ship\n\nTarget statement: Part F needs engineer review.\nTarget atom: needs_engineer_review::Part F"}]}
+{"id": "dynamic_dialogue_v3_0090", "base_id": "university_registration_0045_open_missing_direct", "split": "test", "domain": "university_registration", "family": "open_missing_direct", "subset": "core_contrastive", "gold_pattern": "unknown|false|unknown", "dialogue_type": "reopen", "target_statement": "Omar has an external scholarship.", "query_atom": "has_external_scholarship::Omar", "facts_positive": ["Omar passed the qualifying exam."], "facts_negative": [], "rules_natural": [], "predicate_glossary": {"passed_exam": "passed the qualifying exam", "has_advising_clearance": "has advising clearance", "has_external_scholarship": "has an external scholarship", "may_register": "may register for the seminar", "needs_manual_audit": "needs a manual audit"}, "vocabulary_predicates": ["has_advising_clearance", "has_external_scholarship", "may_register", "needs_manual_audit", "passed_exam"], "symbolic": {"positive_atoms": ["passed_exam::Omar"], "negative_atoms": [], "rules": [], "query_atom": "has_external_scholarship::Omar", "closure_atoms": ["passed_exam::Omar"]}, "turns": [{"turn_index": 1, "source_semantics": "owa", "update_operation": "SET_COMPLETE_PREDICATES", "update_predicates": [], "complete_predicates_after_update": [], "gold_answer": "unknown", "expected_changed_from_previous": null, "expected_applied_update": "open", "prompt": "This is turn 1 of a 3-turn dynamic completeness dialogue.\nYou are given an initial knowledge base and an initial completeness state.\nDo not use outside knowledge.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly open. For turn 1, changed_from_previous must be null.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nKnowledge base:\n\nPredicate glossary:\n- has_advising_clearance: has advising clearance\n- has_external_scholarship: has an external scholarship\n- may_register: may register for the seminar\n- needs_manual_audit: needs a manual audit\n- passed_exam: passed the qualifying exam\n\nPositive facts:\n- Omar passed the qualifying exam.\n\nNegative facts:\n- none\n\nRules:\n- none\n\nTarget statement: Omar has an external scholarship.\nTarget atom: has_external_scholarship::Omar\n\nCompleteness-state update for this turn:\n- Operation: SET_COMPLETE_PREDICATES\n- New complete predicates: none"}, {"turn_index": 2, "source_semantics": "cwa", "update_operation": "ADD_COMPLETE_PREDICATES", "update_predicates": ["has_advising_clearance", "has_external_scholarship", "may_register", "needs_manual_audit", "passed_exam"], "complete_predicates_after_update": ["has_advising_clearance", "has_external_scholarship", "may_register", "needs_manual_audit", "passed_exam"], "gold_answer": "false", "expected_changed_from_previous": true, "expected_applied_update": "global_complete", "prompt": "This is turn 2 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly global_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: ADD_COMPLETE_PREDICATES\n- Predicates to add as complete: has_advising_clearance, has_external_scholarship, may_register, needs_manual_audit, passed_exam\n- Complete predicates after this update: has_advising_clearance, has_external_scholarship, may_register, needs_manual_audit, passed_exam\n\nTarget statement: Omar has an external scholarship.\nTarget atom: has_external_scholarship::Omar"}, {"turn_index": 3, "source_semantics": "lcwa", "update_operation": "REPLACE_COMPLETE_PREDICATES", "update_predicates": ["has_advising_clearance", "may_register"], "complete_predicates_after_update": ["has_advising_clearance", "may_register"], "gold_answer": "unknown", "expected_changed_from_previous": true, "expected_applied_update": "local_complete", "prompt": "This is turn 3 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly local_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: REPLACE_COMPLETE_PREDICATES\n- Discard the previous complete-predicate set.\n- Replacement complete predicates: has_advising_clearance, may_register\n\nTarget statement: Omar has an external scholarship.\nTarget atom: has_external_scholarship::Omar"}]}
+{"id": "dynamic_dialogue_v3_0095", "base_id": "university_registration_0050_closed_missing_with_open_distractor", "split": "test", "domain": "university_registration", "family": "closed_missing_with_open_distractor", "subset": "core_contrastive", "gold_pattern": "unknown|false|false", "dialogue_type": "persist_after_narrowing", "target_statement": "Lena has advising clearance.", "query_atom": "has_advising_clearance::Lena", "facts_positive": ["Omar has an external scholarship.", "Lena passed the qualifying exam."], "facts_negative": [], "rules_natural": [], "predicate_glossary": {"passed_exam": "passed the qualifying exam", "has_advising_clearance": "has advising clearance", "has_external_scholarship": "has an external scholarship", "may_register": "may register for the seminar", "needs_manual_audit": "needs a manual audit"}, "vocabulary_predicates": ["has_advising_clearance", "has_external_scholarship", "may_register", "needs_manual_audit", "passed_exam"], "symbolic": {"positive_atoms": ["has_external_scholarship::Omar", "passed_exam::Lena"], "negative_atoms": [], "rules": [], "query_atom": "has_advising_clearance::Lena", "closure_atoms": ["has_external_scholarship::Omar", "passed_exam::Lena"]}, "turns": [{"turn_index": 1, "source_semantics": "owa", "update_operation": "SET_COMPLETE_PREDICATES", "update_predicates": [], "complete_predicates_after_update": [], "gold_answer": "unknown", "expected_changed_from_previous": null, "expected_applied_update": "open", "prompt": "This is turn 1 of a 3-turn dynamic completeness dialogue.\nYou are given an initial knowledge base and an initial completeness state.\nDo not use outside knowledge.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly open. For turn 1, changed_from_previous must be null.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nKnowledge base:\n\nPredicate glossary:\n- has_advising_clearance: has advising clearance\n- has_external_scholarship: has an external scholarship\n- may_register: may register for the seminar\n- needs_manual_audit: needs a manual audit\n- passed_exam: passed the qualifying exam\n\nPositive facts:\n- Omar has an external scholarship.\n- Lena passed the qualifying exam.\n\nNegative facts:\n- none\n\nRules:\n- none\n\nTarget statement: Lena has advising clearance.\nTarget atom: has_advising_clearance::Lena\n\nCompleteness-state update for this turn:\n- Operation: SET_COMPLETE_PREDICATES\n- New complete predicates: none"}, {"turn_index": 2, "source_semantics": "cwa", "update_operation": "ADD_COMPLETE_PREDICATES", "update_predicates": ["has_advising_clearance", "has_external_scholarship", "may_register", "needs_manual_audit", "passed_exam"], "complete_predicates_after_update": ["has_advising_clearance", "has_external_scholarship", "may_register", "needs_manual_audit", "passed_exam"], "gold_answer": "false", "expected_changed_from_previous": true, "expected_applied_update": "global_complete", "prompt": "This is turn 2 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly global_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: ADD_COMPLETE_PREDICATES\n- Predicates to add as complete: has_advising_clearance, has_external_scholarship, may_register, needs_manual_audit, passed_exam\n- Complete predicates after this update: has_advising_clearance, has_external_scholarship, may_register, needs_manual_audit, passed_exam\n\nTarget statement: Lena has advising clearance.\nTarget atom: has_advising_clearance::Lena"}, {"turn_index": 3, "source_semantics": "lcwa", "update_operation": "REPLACE_COMPLETE_PREDICATES", "update_predicates": ["has_advising_clearance"], "complete_predicates_after_update": ["has_advising_clearance"], "gold_answer": "false", "expected_changed_from_previous": false, "expected_applied_update": "local_complete", "prompt": "This is turn 3 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly local_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: REPLACE_COMPLETE_PREDICATES\n- Discard the previous complete-predicate set.\n- Replacement complete predicates: has_advising_clearance\n\nTarget statement: Lena has advising clearance.\nTarget atom: has_advising_clearance::Lena"}]}

data/dynamic_dialogue/validation.jsonl

@@ -0,0 +1,10 @@
+{"id": "dynamic_dialogue_v3_0012", "base_id": "finance_controls_0130_closed_missing_with_open_distractor", "split": "dev", "domain": "finance_controls", "family": "closed_missing_with_open_distractor", "subset": "core_contrastive", "gold_pattern": "unknown|false|false", "dialogue_type": "persist_after_narrowing", "target_statement": "Account 59 has compliance clearance.", "query_atom": "has_compliance_clearance::Account 59", "facts_positive": ["Account 61 has a manual exception.", "Account 59 passed KYC checks."], "facts_negative": [], "rules_natural": [], "predicate_glossary": {"passed_kyc": "passed KYC checks", "has_compliance_clearance": "has compliance clearance", "has_manual_exception": "has a manual exception", "may_receive_wire": "may receive a wire transfer", "needs_compliance_review": "needs compliance review"}, "vocabulary_predicates": ["has_compliance_clearance", "has_manual_exception", "may_receive_wire", "needs_compliance_review", "passed_kyc"], "symbolic": {"positive_atoms": ["has_manual_exception::Account 61", "passed_kyc::Account 59"], "negative_atoms": [], "rules": [], "query_atom": "has_compliance_clearance::Account 59", "closure_atoms": ["has_manual_exception::Account 61", "passed_kyc::Account 59"]}, "turns": [{"turn_index": 1, "source_semantics": "owa", "update_operation": "SET_COMPLETE_PREDICATES", "update_predicates": [], "complete_predicates_after_update": [], "gold_answer": "unknown", "expected_changed_from_previous": null, "expected_applied_update": "open", "prompt": "This is turn 1 of a 3-turn dynamic completeness dialogue.\nYou are given an initial knowledge base and an initial completeness state.\nDo not use outside knowledge.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly open. For turn 1, changed_from_previous must be null.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nKnowledge base:\n\nPredicate glossary:\n- has_compliance_clearance: has compliance clearance\n- has_manual_exception: has a manual exception\n- may_receive_wire: may receive a wire transfer\n- needs_compliance_review: needs compliance review\n- passed_kyc: passed KYC checks\n\nPositive facts:\n- Account 61 has a manual exception.\n- Account 59 passed KYC checks.\n\nNegative facts:\n- none\n\nRules:\n- none\n\nTarget statement: Account 59 has compliance clearance.\nTarget atom: has_compliance_clearance::Account 59\n\nCompleteness-state update for this turn:\n- Operation: SET_COMPLETE_PREDICATES\n- New complete predicates: none"}, {"turn_index": 2, "source_semantics": "cwa", "update_operation": "ADD_COMPLETE_PREDICATES", "update_predicates": ["has_compliance_clearance", "has_manual_exception", "may_receive_wire", "needs_compliance_review", "passed_kyc"], "complete_predicates_after_update": ["has_compliance_clearance", "has_manual_exception", "may_receive_wire", "needs_compliance_review", "passed_kyc"], "gold_answer": "false", "expected_changed_from_previous": true, "expected_applied_update": "global_complete", "prompt": "This is turn 2 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly global_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: ADD_COMPLETE_PREDICATES\n- Predicates to add as complete: has_compliance_clearance, has_manual_exception, may_receive_wire, needs_compliance_review, passed_kyc\n- Complete predicates after this update: has_compliance_clearance, has_manual_exception, may_receive_wire, needs_compliance_review, passed_kyc\n\nTarget statement: Account 59 has compliance clearance.\nTarget atom: has_compliance_clearance::Account 59"}, {"turn_index": 3, "source_semantics": "lcwa", "update_operation": "REPLACE_COMPLETE_PREDICATES", "update_predicates": ["has_compliance_clearance"], "complete_predicates_after_update": ["has_compliance_clearance"], "gold_answer": "false", "expected_changed_from_previous": false, "expected_applied_update": "local_complete", "prompt": "This is turn 3 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly local_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: REPLACE_COMPLETE_PREDICATES\n- Discard the previous complete-predicate set.\n- Replacement complete predicates: has_compliance_clearance\n\nTarget statement: Account 59 has compliance clearance.\nTarget atom: has_compliance_clearance::Account 59"}]}
+{"id": "dynamic_dialogue_v3_0018", "base_id": "finance_controls_0138_closed_derived_missing_antecedent", "split": "dev", "domain": "finance_controls", "family": "closed_derived_missing_antecedent", "subset": "core_contrastive", "gold_pattern": "unknown|false|false", "dialogue_type": "persist_after_narrowing", "target_statement": "Account 12 may receive a wire transfer.", "query_atom": "may_receive_wire::Account 12", "facts_positive": ["Account 12 passed KYC checks."], "facts_negative": [], "rules_natural": ["If an account passed KYC checks and has compliance clearance, then that account may receive a wire transfer."], "predicate_glossary": {"passed_kyc": "passed KYC checks", "has_compliance_clearance": "has compliance clearance", "has_manual_exception": "has a manual exception", "may_receive_wire": "may receive a wire transfer", "needs_compliance_review": "needs compliance review"}, "vocabulary_predicates": ["has_compliance_clearance", "has_manual_exception", "may_receive_wire", "needs_compliance_review", "passed_kyc"], "symbolic": {"positive_atoms": ["passed_kyc::Account 12"], "negative_atoms": [], "rules": [{"antecedents": ["passed_kyc::Account 12", "has_compliance_clearance::Account 12"], "conclusion": "may_receive_wire::Account 12", "text": "If an account passed KYC checks and has compliance clearance, then that account may receive a wire transfer."}], "query_atom": "may_receive_wire::Account 12", "closure_atoms": ["passed_kyc::Account 12"]}, "turns": [{"turn_index": 1, "source_semantics": "owa", "update_operation": "SET_COMPLETE_PREDICATES", "update_predicates": [], "complete_predicates_after_update": [], "gold_answer": "unknown", "expected_changed_from_previous": null, "expected_applied_update": "open", "prompt": "This is turn 1 of a 3-turn dynamic completeness dialogue.\nYou are given an initial knowledge base and an initial completeness state.\nDo not use outside knowledge.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly open. For turn 1, changed_from_previous must be null.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nKnowledge base:\n\nPredicate glossary:\n- has_compliance_clearance: has compliance clearance\n- has_manual_exception: has a manual exception\n- may_receive_wire: may receive a wire transfer\n- needs_compliance_review: needs compliance review\n- passed_kyc: passed KYC checks\n\nPositive facts:\n- Account 12 passed KYC checks.\n\nNegative facts:\n- none\n\nRules:\n- If an account passed KYC checks and has compliance clearance, then that account may receive a wire transfer.\n\nTarget statement: Account 12 may receive a wire transfer.\nTarget atom: may_receive_wire::Account 12\n\nCompleteness-state update for this turn:\n- Operation: SET_COMPLETE_PREDICATES\n- New complete predicates: none"}, {"turn_index": 2, "source_semantics": "cwa", "update_operation": "ADD_COMPLETE_PREDICATES", "update_predicates": ["has_compliance_clearance", "has_manual_exception", "may_receive_wire", "needs_compliance_review", "passed_kyc"], "complete_predicates_after_update": ["has_compliance_clearance", "has_manual_exception", "may_receive_wire", "needs_compliance_review", "passed_kyc"], "gold_answer": "false", "expected_changed_from_previous": true, "expected_applied_update": "global_complete", "prompt": "This is turn 2 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly global_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: ADD_COMPLETE_PREDICATES\n- Predicates to add as complete: has_compliance_clearance, has_manual_exception, may_receive_wire, needs_compliance_review, passed_kyc\n- Complete predicates after this update: has_compliance_clearance, has_manual_exception, may_receive_wire, needs_compliance_review, passed_kyc\n\nTarget statement: Account 12 may receive a wire transfer.\nTarget atom: may_receive_wire::Account 12"}, {"turn_index": 3, "source_semantics": "lcwa", "update_operation": "REPLACE_COMPLETE_PREDICATES", "update_predicates": ["has_compliance_clearance", "may_receive_wire"], "complete_predicates_after_update": ["has_compliance_clearance", "may_receive_wire"], "gold_answer": "false", "expected_changed_from_previous": false, "expected_applied_update": "local_complete", "prompt": "This is turn 3 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly local_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: REPLACE_COMPLETE_PREDICATES\n- Discard the previous complete-predicate set.\n- Replacement complete predicates: has_compliance_clearance, may_receive_wire\n\nTarget statement: Account 12 may receive a wire transfer.\nTarget atom: may_receive_wire::Account 12"}]}
+{"id": "dynamic_dialogue_v3_0022", "base_id": "finance_controls_0142_open_derived_missing_antecedent", "split": "dev", "domain": "finance_controls", "family": "open_derived_missing_antecedent", "subset": "core_contrastive", "gold_pattern": "unknown|false|unknown", "dialogue_type": "reopen", "target_statement": "Account 59 needs compliance review.", "query_atom": "needs_compliance_review::Account 59", "facts_positive": ["Account 59 passed KYC checks."], "facts_negative": [], "rules_natural": ["If an account has a manual exception, then that account needs compliance review."], "predicate_glossary": {"passed_kyc": "passed KYC checks", "has_compliance_clearance": "has compliance clearance", "has_manual_exception": "has a manual exception", "may_receive_wire": "may receive a wire transfer", "needs_compliance_review": "needs compliance review"}, "vocabulary_predicates": ["has_compliance_clearance", "has_manual_exception", "may_receive_wire", "needs_compliance_review", "passed_kyc"], "symbolic": {"positive_atoms": ["passed_kyc::Account 59"], "negative_atoms": [], "rules": [{"antecedents": ["has_manual_exception::Account 59"], "conclusion": "needs_compliance_review::Account 59", "text": "If an account has a manual exception, then that account needs compliance review."}], "query_atom": "needs_compliance_review::Account 59", "closure_atoms": ["passed_kyc::Account 59"]}, "turns": [{"turn_index": 1, "source_semantics": "owa", "update_operation": "SET_COMPLETE_PREDICATES", "update_predicates": [], "complete_predicates_after_update": [], "gold_answer": "unknown", "expected_changed_from_previous": null, "expected_applied_update": "open", "prompt": "This is turn 1 of a 3-turn dynamic completeness dialogue.\nYou are given an initial knowledge base and an initial completeness state.\nDo not use outside knowledge.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly open. For turn 1, changed_from_previous must be null.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nKnowledge base:\n\nPredicate glossary:\n- has_compliance_clearance: has compliance clearance\n- has_manual_exception: has a manual exception\n- may_receive_wire: may receive a wire transfer\n- needs_compliance_review: needs compliance review\n- passed_kyc: passed KYC checks\n\nPositive facts:\n- Account 59 passed KYC checks.\n\nNegative facts:\n- none\n\nRules:\n- If an account has a manual exception, then that account needs compliance review.\n\nTarget statement: Account 59 needs compliance review.\nTarget atom: needs_compliance_review::Account 59\n\nCompleteness-state update for this turn:\n- Operation: SET_COMPLETE_PREDICATES\n- New complete predicates: none"}, {"turn_index": 2, "source_semantics": "cwa", "update_operation": "ADD_COMPLETE_PREDICATES", "update_predicates": ["has_compliance_clearance", "has_manual_exception", "may_receive_wire", "needs_compliance_review", "passed_kyc"], "complete_predicates_after_update": ["has_compliance_clearance", "has_manual_exception", "may_receive_wire", "needs_compliance_review", "passed_kyc"], "gold_answer": "false", "expected_changed_from_previous": true, "expected_applied_update": "global_complete", "prompt": "This is turn 2 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly global_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: ADD_COMPLETE_PREDICATES\n- Predicates to add as complete: has_compliance_clearance, has_manual_exception, may_receive_wire, needs_compliance_review, passed_kyc\n- Complete predicates after this update: has_compliance_clearance, has_manual_exception, may_receive_wire, needs_compliance_review, passed_kyc\n\nTarget statement: Account 59 needs compliance review.\nTarget atom: needs_compliance_review::Account 59"}, {"turn_index": 3, "source_semantics": "lcwa", "update_operation": "REPLACE_COMPLETE_PREDICATES", "update_predicates": ["has_compliance_clearance", "may_receive_wire"], "complete_predicates_after_update": ["has_compliance_clearance", "may_receive_wire"], "gold_answer": "unknown", "expected_changed_from_previous": true, "expected_applied_update": "local_complete", "prompt": "This is turn 3 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly local_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: REPLACE_COMPLETE_PREDICATES\n- Discard the previous complete-predicate set.\n- Replacement complete predicates: has_compliance_clearance, may_receive_wire\n\nTarget statement: Account 59 needs compliance review.\nTarget atom: needs_compliance_review::Account 59"}]}
+{"id": "dynamic_dialogue_v3_0024", "base_id": "hospital_access_0004_open_missing_direct", "split": "dev", "domain": "hospital_access", "family": "open_missing_direct", "subset": "core_contrastive", "gold_pattern": "unknown|false|unknown", "dialogue_type": "reopen", "target_statement": "Elena is assigned to the clinical trial.", "query_atom": "assigned_to_trial::Elena", "facts_positive": ["Elena has completed safety training."], "facts_negative": [], "rules_natural": [], "predicate_glossary": {"completed_training": "has completed safety training", "has_clearance": "has security clearance", "assigned_to_trial": "is assigned to the clinical trial", "may_enter_lab": "may enter Lab A", "needs_supervisor_review": "needs supervisor review"}, "vocabulary_predicates": ["assigned_to_trial", "completed_training", "has_clearance", "may_enter_lab", "needs_supervisor_review"], "symbolic": {"positive_atoms": ["completed_training::Elena"], "negative_atoms": [], "rules": [], "query_atom": "assigned_to_trial::Elena", "closure_atoms": ["completed_training::Elena"]}, "turns": [{"turn_index": 1, "source_semantics": "owa", "update_operation": "SET_COMPLETE_PREDICATES", "update_predicates": [], "complete_predicates_after_update": [], "gold_answer": "unknown", "expected_changed_from_previous": null, "expected_applied_update": "open", "prompt": "This is turn 1 of a 3-turn dynamic completeness dialogue.\nYou are given an initial knowledge base and an initial completeness state.\nDo not use outside knowledge.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly open. For turn 1, changed_from_previous must be null.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nKnowledge base:\n\nPredicate glossary:\n- assigned_to_trial: is assigned to the clinical trial\n- completed_training: has completed safety training\n- has_clearance: has security clearance\n- may_enter_lab: may enter Lab A\n- needs_supervisor_review: needs supervisor review\n\nPositive facts:\n- Elena has completed safety training.\n\nNegative facts:\n- none\n\nRules:\n- none\n\nTarget statement: Elena is assigned to the clinical trial.\nTarget atom: assigned_to_trial::Elena\n\nCompleteness-state update for this turn:\n- Operation: SET_COMPLETE_PREDICATES\n- New complete predicates: none"}, {"turn_index": 2, "source_semantics": "cwa", "update_operation": "ADD_COMPLETE_PREDICATES", "update_predicates": ["assigned_to_trial", "completed_training", "has_clearance", "may_enter_lab", "needs_supervisor_review"], "complete_predicates_after_update": ["assigned_to_trial", "completed_training", "has_clearance", "may_enter_lab", "needs_supervisor_review"], "gold_answer": "false", "expected_changed_from_previous": true, "expected_applied_update": "global_complete", "prompt": "This is turn 2 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly global_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: ADD_COMPLETE_PREDICATES\n- Predicates to add as complete: assigned_to_trial, completed_training, has_clearance, may_enter_lab, needs_supervisor_review\n- Complete predicates after this update: assigned_to_trial, completed_training, has_clearance, may_enter_lab, needs_supervisor_review\n\nTarget statement: Elena is assigned to the clinical trial.\nTarget atom: assigned_to_trial::Elena"}, {"turn_index": 3, "source_semantics": "lcwa", "update_operation": "REPLACE_COMPLETE_PREDICATES", "update_predicates": ["has_clearance", "may_enter_lab"], "complete_predicates_after_update": ["has_clearance", "may_enter_lab"], "gold_answer": "unknown", "expected_changed_from_previous": true, "expected_applied_update": "local_complete", "prompt": "This is turn 3 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly local_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: REPLACE_COMPLETE_PREDICATES\n- Discard the previous complete-predicate set.\n- Replacement complete predicates: has_clearance, may_enter_lab\n\nTarget statement: Elena is assigned to the clinical trial.\nTarget atom: assigned_to_trial::Elena"}]}
+{"id": "dynamic_dialogue_v3_0031", "base_id": "hospital_access_0017_closed_derived_missing_antecedent", "split": "dev", "domain": "hospital_access", "family": "closed_derived_missing_antecedent", "subset": "core_contrastive", "gold_pattern": "unknown|false|false", "dialogue_type": "persist_after_narrowing", "target_statement": "Farid may enter Lab A.", "query_atom": "may_enter_lab::Farid", "facts_positive": ["Farid has completed safety training."], "facts_negative": [], "rules_natural": ["If a staff member has security clearance and has completed safety training, then that staff member may enter Lab A."], "predicate_glossary": {"completed_training": "has completed safety training", "has_clearance": "has security clearance", "assigned_to_trial": "is assigned to the clinical trial", "may_enter_lab": "may enter Lab A", "needs_supervisor_review": "needs supervisor review"}, "vocabulary_predicates": ["assigned_to_trial", "completed_training", "has_clearance", "may_enter_lab", "needs_supervisor_review"], "symbolic": {"positive_atoms": ["completed_training::Farid"], "negative_atoms": [], "rules": [{"antecedents": ["completed_training::Farid", "has_clearance::Farid"], "conclusion": "may_enter_lab::Farid", "text": "If a staff member has security clearance and has completed safety training, then that staff member may enter Lab A."}], "query_atom": "may_enter_lab::Farid", "closure_atoms": ["completed_training::Farid"]}, "turns": [{"turn_index": 1, "source_semantics": "owa", "update_operation": "SET_COMPLETE_PREDICATES", "update_predicates": [], "complete_predicates_after_update": [], "gold_answer": "unknown", "expected_changed_from_previous": null, "expected_applied_update": "open", "prompt": "This is turn 1 of a 3-turn dynamic completeness dialogue.\nYou are given an initial knowledge base and an initial completeness state.\nDo not use outside knowledge.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly open. For turn 1, changed_from_previous must be null.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nKnowledge base:\n\nPredicate glossary:\n- assigned_to_trial: is assigned to the clinical trial\n- completed_training: has completed safety training\n- has_clearance: has security clearance\n- may_enter_lab: may enter Lab A\n- needs_supervisor_review: needs supervisor review\n\nPositive facts:\n- Farid has completed safety training.\n\nNegative facts:\n- none\n\nRules:\n- If a staff member has security clearance and has completed safety training, then that staff member may enter Lab A.\n\nTarget statement: Farid may enter Lab A.\nTarget atom: may_enter_lab::Farid\n\nCompleteness-state update for this turn:\n- Operation: SET_COMPLETE_PREDICATES\n- New complete predicates: none"}, {"turn_index": 2, "source_semantics": "cwa", "update_operation": "ADD_COMPLETE_PREDICATES", "update_predicates": ["assigned_to_trial", "completed_training", "has_clearance", "may_enter_lab", "needs_supervisor_review"], "complete_predicates_after_update": ["assigned_to_trial", "completed_training", "has_clearance", "may_enter_lab", "needs_supervisor_review"], "gold_answer": "false", "expected_changed_from_previous": true, "expected_applied_update": "global_complete", "prompt": "This is turn 2 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly global_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: ADD_COMPLETE_PREDICATES\n- Predicates to add as complete: assigned_to_trial, completed_training, has_clearance, may_enter_lab, needs_supervisor_review\n- Complete predicates after this update: assigned_to_trial, completed_training, has_clearance, may_enter_lab, needs_supervisor_review\n\nTarget statement: Farid may enter Lab A.\nTarget atom: may_enter_lab::Farid"}, {"turn_index": 3, "source_semantics": "lcwa", "update_operation": "REPLACE_COMPLETE_PREDICATES", "update_predicates": ["has_clearance", "may_enter_lab"], "complete_predicates_after_update": ["has_clearance", "may_enter_lab"], "gold_answer": "false", "expected_changed_from_previous": false, "expected_applied_update": "local_complete", "prompt": "This is turn 3 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly local_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: REPLACE_COMPLETE_PREDICATES\n- Discard the previous complete-predicate set.\n- Replacement complete predicates: has_clearance, may_enter_lab\n\nTarget statement: Farid may enter Lab A.\nTarget atom: may_enter_lab::Farid"}]}
+{"id": "dynamic_dialogue_v3_0042", "base_id": "library_services_0209_closed_missing_with_open_distractor", "split": "dev", "domain": "library_services", "family": "closed_missing_with_open_distractor", "subset": "core_contrastive", "gold_pattern": "unknown|false|false", "dialogue_type": "persist_after_narrowing", "target_statement": "Reader F has borrowing clearance.", "query_atom": "has_borrowing_clearance::Reader F", "facts_positive": ["Reader F has an active membership.", "Reader A has a special request."], "facts_negative": [], "rules_natural": [], "predicate_glossary": {"has_active_membership": "has an active membership", "has_borrowing_clearance": "has borrowing clearance", "has_special_request": "has a special request", "may_borrow_archive_item": "may borrow an archive item", "needs_librarian_review": "needs librarian review"}, "vocabulary_predicates": ["has_active_membership", "has_borrowing_clearance", "has_special_request", "may_borrow_archive_item", "needs_librarian_review"], "symbolic": {"positive_atoms": ["has_active_membership::Reader F", "has_special_request::Reader A"], "negative_atoms": [], "rules": [], "query_atom": "has_borrowing_clearance::Reader F", "closure_atoms": ["has_active_membership::Reader F", "has_special_request::Reader A"]}, "turns": [{"turn_index": 1, "source_semantics": "owa", "update_operation": "SET_COMPLETE_PREDICATES", "update_predicates": [], "complete_predicates_after_update": [], "gold_answer": "unknown", "expected_changed_from_previous": null, "expected_applied_update": "open", "prompt": "This is turn 1 of a 3-turn dynamic completeness dialogue.\nYou are given an initial knowledge base and an initial completeness state.\nDo not use outside knowledge.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly open. For turn 1, changed_from_previous must be null.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nKnowledge base:\n\nPredicate glossary:\n- has_active_membership: has an active membership\n- has_borrowing_clearance: has borrowing clearance\n- has_special_request: has a special request\n- may_borrow_archive_item: may borrow an archive item\n- needs_librarian_review: needs librarian review\n\nPositive facts:\n- Reader F has an active membership.\n- Reader A has a special request.\n\nNegative facts:\n- none\n\nRules:\n- none\n\nTarget statement: Reader F has borrowing clearance.\nTarget atom: has_borrowing_clearance::Reader F\n\nCompleteness-state update for this turn:\n- Operation: SET_COMPLETE_PREDICATES\n- New complete predicates: none"}, {"turn_index": 2, "source_semantics": "cwa", "update_operation": "ADD_COMPLETE_PREDICATES", "update_predicates": ["has_active_membership", "has_borrowing_clearance", "has_special_request", "may_borrow_archive_item", "needs_librarian_review"], "complete_predicates_after_update": ["has_active_membership", "has_borrowing_clearance", "has_special_request", "may_borrow_archive_item", "needs_librarian_review"], "gold_answer": "false", "expected_changed_from_previous": true, "expected_applied_update": "global_complete", "prompt": "This is turn 2 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly global_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: ADD_COMPLETE_PREDICATES\n- Predicates to add as complete: has_active_membership, has_borrowing_clearance, has_special_request, may_borrow_archive_item, needs_librarian_review\n- Complete predicates after this update: has_active_membership, has_borrowing_clearance, has_special_request, may_borrow_archive_item, needs_librarian_review\n\nTarget statement: Reader F has borrowing clearance.\nTarget atom: has_borrowing_clearance::Reader F"}, {"turn_index": 3, "source_semantics": "lcwa", "update_operation": "REPLACE_COMPLETE_PREDICATES", "update_predicates": ["has_borrowing_clearance"], "complete_predicates_after_update": ["has_borrowing_clearance"], "gold_answer": "false", "expected_changed_from_previous": false, "expected_applied_update": "local_complete", "prompt": "This is turn 3 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly local_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: REPLACE_COMPLETE_PREDICATES\n- Discard the previous complete-predicate set.\n- Replacement complete predicates: has_borrowing_clearance\n\nTarget statement: Reader F has borrowing clearance.\nTarget atom: has_borrowing_clearance::Reader F"}]}
+{"id": "dynamic_dialogue_v3_0055", "base_id": "manufacturing_quality_0099_closed_derived_missing_antecedent", "split": "dev", "domain": "manufacturing_quality", "family": "closed_derived_missing_antecedent", "subset": "core_contrastive", "gold_pattern": "unknown|false|false", "dialogue_type": "persist_after_narrowing", "target_statement": "Part D may ship to customers.", "query_atom": "may_ship::Part D", "facts_positive": ["Part D passed the visual check."], "facts_negative": [], "rules_natural": ["If a part passed the visual check and has release approval, then that part may ship to customers."], "predicate_glossary": {"passed_visual_check": "passed the visual check", "has_release_approval": "has release approval", "has_supplier_note": "has a supplier note", "may_ship": "may ship to customers", "needs_engineer_review": "needs engineer review"}, "vocabulary_predicates": ["has_release_approval", "has_supplier_note", "may_ship", "needs_engineer_review", "passed_visual_check"], "symbolic": {"positive_atoms": ["passed_visual_check::Part D"], "negative_atoms": [], "rules": [{"antecedents": ["passed_visual_check::Part D", "has_release_approval::Part D"], "conclusion": "may_ship::Part D", "text": "If a part passed the visual check and has release approval, then that part may ship to customers."}], "query_atom": "may_ship::Part D", "closure_atoms": ["passed_visual_check::Part D"]}, "turns": [{"turn_index": 1, "source_semantics": "owa", "update_operation": "SET_COMPLETE_PREDICATES", "update_predicates": [], "complete_predicates_after_update": [], "gold_answer": "unknown", "expected_changed_from_previous": null, "expected_applied_update": "open", "prompt": "This is turn 1 of a 3-turn dynamic completeness dialogue.\nYou are given an initial knowledge base and an initial completeness state.\nDo not use outside knowledge.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly open. For turn 1, changed_from_previous must be null.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nKnowledge base:\n\nPredicate glossary:\n- has_release_approval: has release approval\n- has_supplier_note: has a supplier note\n- may_ship: may ship to customers\n- needs_engineer_review: needs engineer review\n- passed_visual_check: passed the visual check\n\nPositive facts:\n- Part D passed the visual check.\n\nNegative facts:\n- none\n\nRules:\n- If a part passed the visual check and has release approval, then that part may ship to customers.\n\nTarget statement: Part D may ship to customers.\nTarget atom: may_ship::Part D\n\nCompleteness-state update for this turn:\n- Operation: SET_COMPLETE_PREDICATES\n- New complete predicates: none"}, {"turn_index": 2, "source_semantics": "cwa", "update_operation": "ADD_COMPLETE_PREDICATES", "update_predicates": ["has_release_approval", "has_supplier_note", "may_ship", "needs_engineer_review", "passed_visual_check"], "complete_predicates_after_update": ["has_release_approval", "has_supplier_note", "may_ship", "needs_engineer_review", "passed_visual_check"], "gold_answer": "false", "expected_changed_from_previous": true, "expected_applied_update": "global_complete", "prompt": "This is turn 2 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly global_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: ADD_COMPLETE_PREDICATES\n- Predicates to add as complete: has_release_approval, has_supplier_note, may_ship, needs_engineer_review, passed_visual_check\n- Complete predicates after this update: has_release_approval, has_supplier_note, may_ship, needs_engineer_review, passed_visual_check\n\nTarget statement: Part D may ship to customers.\nTarget atom: may_ship::Part D"}, {"turn_index": 3, "source_semantics": "lcwa", "update_operation": "REPLACE_COMPLETE_PREDICATES", "update_predicates": ["has_release_approval", "may_ship"], "complete_predicates_after_update": ["has_release_approval", "may_ship"], "gold_answer": "false", "expected_changed_from_previous": false, "expected_applied_update": "local_complete", "prompt": "This is turn 3 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly local_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: REPLACE_COMPLETE_PREDICATES\n- Discard the previous complete-predicate set.\n- Replacement complete predicates: has_release_approval, may_ship\n\nTarget statement: Part D may ship to customers.\nTarget atom: may_ship::Part D"}]}
+{"id": "dynamic_dialogue_v3_0077", "base_id": "robotics_operations_0169_closed_missing_with_open_distractor", "split": "dev", "domain": "robotics_operations", "family": "closed_missing_with_open_distractor", "subset": "core_contrastive", "gold_pattern": "unknown|false|false", "dialogue_type": "persist_after_narrowing", "target_statement": "Rover 2 has shift clearance.", "query_atom": "has_shift_clearance::Rover 2", "facts_positive": ["Rover 3 has a field note.", "Rover 2 passed diagnostics."], "facts_negative": [], "rules_natural": [], "predicate_glossary": {"passed_diagnostics": "passed diagnostics", "has_shift_clearance": "has shift clearance", "has_field_note": "has a field note", "may_start_shift": "may start its shift", "needs_operator_review": "needs operator review"}, "vocabulary_predicates": ["has_field_note", "has_shift_clearance", "may_start_shift", "needs_operator_review", "passed_diagnostics"], "symbolic": {"positive_atoms": ["has_field_note::Rover 3", "passed_diagnostics::Rover 2"], "negative_atoms": [], "rules": [], "query_atom": "has_shift_clearance::Rover 2", "closure_atoms": ["has_field_note::Rover 3", "passed_diagnostics::Rover 2"]}, "turns": [{"turn_index": 1, "source_semantics": "owa", "update_operation": "SET_COMPLETE_PREDICATES", "update_predicates": [], "complete_predicates_after_update": [], "gold_answer": "unknown", "expected_changed_from_previous": null, "expected_applied_update": "open", "prompt": "This is turn 1 of a 3-turn dynamic completeness dialogue.\nYou are given an initial knowledge base and an initial completeness state.\nDo not use outside knowledge.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly open. For turn 1, changed_from_previous must be null.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nKnowledge base:\n\nPredicate glossary:\n- has_field_note: has a field note\n- has_shift_clearance: has shift clearance\n- may_start_shift: may start its shift\n- needs_operator_review: needs operator review\n- passed_diagnostics: passed diagnostics\n\nPositive facts:\n- Rover 3 has a field note.\n- Rover 2 passed diagnostics.\n\nNegative facts:\n- none\n\nRules:\n- none\n\nTarget statement: Rover 2 has shift clearance.\nTarget atom: has_shift_clearance::Rover 2\n\nCompleteness-state update for this turn:\n- Operation: SET_COMPLETE_PREDICATES\n- New complete predicates: none"}, {"turn_index": 2, "source_semantics": "cwa", "update_operation": "ADD_COMPLETE_PREDICATES", "update_predicates": ["has_field_note", "has_shift_clearance", "may_start_shift", "needs_operator_review", "passed_diagnostics"], "complete_predicates_after_update": ["has_field_note", "has_shift_clearance", "may_start_shift", "needs_operator_review", "passed_diagnostics"], "gold_answer": "false", "expected_changed_from_previous": true, "expected_applied_update": "global_complete", "prompt": "This is turn 2 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly global_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: ADD_COMPLETE_PREDICATES\n- Predicates to add as complete: has_field_note, has_shift_clearance, may_start_shift, needs_operator_review, passed_diagnostics\n- Complete predicates after this update: has_field_note, has_shift_clearance, may_start_shift, needs_operator_review, passed_diagnostics\n\nTarget statement: Rover 2 has shift clearance.\nTarget atom: has_shift_clearance::Rover 2"}, {"turn_index": 3, "source_semantics": "lcwa", "update_operation": "REPLACE_COMPLETE_PREDICATES", "update_predicates": ["has_shift_clearance"], "complete_predicates_after_update": ["has_shift_clearance"], "gold_answer": "false", "expected_changed_from_previous": false, "expected_applied_update": "local_complete", "prompt": "This is turn 3 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly local_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: REPLACE_COMPLETE_PREDICATES\n- Discard the previous complete-predicate set.\n- Replacement complete predicates: has_shift_clearance\n\nTarget statement: Rover 2 has shift clearance.\nTarget atom: has_shift_clearance::Rover 2"}]}
+{"id": "dynamic_dialogue_v3_0093", "base_id": "university_registration_0048_closed_missing_with_open_distractor", "split": "dev", "domain": "university_registration", "family": "closed_missing_with_open_distractor", "subset": "core_contrastive", "gold_pattern": "unknown|false|false", "dialogue_type": "persist_after_narrowing", "target_statement": "Mia has advising clearance.", "query_atom": "has_advising_clearance::Mia", "facts_positive": ["Noah has an external scholarship.", "Mia passed the qualifying exam."], "facts_negative": [], "rules_natural": [], "predicate_glossary": {"passed_exam": "passed the qualifying exam", "has_advising_clearance": "has advising clearance", "has_external_scholarship": "has an external scholarship", "may_register": "may register for the seminar", "needs_manual_audit": "needs a manual audit"}, "vocabulary_predicates": ["has_advising_clearance", "has_external_scholarship", "may_register", "needs_manual_audit", "passed_exam"], "symbolic": {"positive_atoms": ["has_external_scholarship::Noah", "passed_exam::Mia"], "negative_atoms": [], "rules": [], "query_atom": "has_advising_clearance::Mia", "closure_atoms": ["has_external_scholarship::Noah", "passed_exam::Mia"]}, "turns": [{"turn_index": 1, "source_semantics": "owa", "update_operation": "SET_COMPLETE_PREDICATES", "update_predicates": [], "complete_predicates_after_update": [], "gold_answer": "unknown", "expected_changed_from_previous": null, "expected_applied_update": "open", "prompt": "This is turn 1 of a 3-turn dynamic completeness dialogue.\nYou are given an initial knowledge base and an initial completeness state.\nDo not use outside knowledge.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly open. For turn 1, changed_from_previous must be null.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nKnowledge base:\n\nPredicate glossary:\n- has_advising_clearance: has advising clearance\n- has_external_scholarship: has an external scholarship\n- may_register: may register for the seminar\n- needs_manual_audit: needs a manual audit\n- passed_exam: passed the qualifying exam\n\nPositive facts:\n- Noah has an external scholarship.\n- Mia passed the qualifying exam.\n\nNegative facts:\n- none\n\nRules:\n- none\n\nTarget statement: Mia has advising clearance.\nTarget atom: has_advising_clearance::Mia\n\nCompleteness-state update for this turn:\n- Operation: SET_COMPLETE_PREDICATES\n- New complete predicates: none"}, {"turn_index": 2, "source_semantics": "cwa", "update_operation": "ADD_COMPLETE_PREDICATES", "update_predicates": ["has_advising_clearance", "has_external_scholarship", "may_register", "needs_manual_audit", "passed_exam"], "complete_predicates_after_update": ["has_advising_clearance", "has_external_scholarship", "may_register", "needs_manual_audit", "passed_exam"], "gold_answer": "false", "expected_changed_from_previous": true, "expected_applied_update": "global_complete", "prompt": "This is turn 2 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly global_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: ADD_COMPLETE_PREDICATES\n- Predicates to add as complete: has_advising_clearance, has_external_scholarship, may_register, needs_manual_audit, passed_exam\n- Complete predicates after this update: has_advising_clearance, has_external_scholarship, may_register, needs_manual_audit, passed_exam\n\nTarget statement: Mia has advising clearance.\nTarget atom: has_advising_clearance::Mia"}, {"turn_index": 3, "source_semantics": "lcwa", "update_operation": "REPLACE_COMPLETE_PREDICATES", "update_predicates": ["has_advising_clearance"], "complete_predicates_after_update": ["has_advising_clearance"], "gold_answer": "false", "expected_changed_from_previous": false, "expected_applied_update": "local_complete", "prompt": "This is turn 3 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly local_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: REPLACE_COMPLETE_PREDICATES\n- Discard the previous complete-predicate set.\n- Replacement complete predicates: has_advising_clearance\n\nTarget statement: Mia has advising clearance.\nTarget atom: has_advising_clearance::Mia"}]}
+{"id": "dynamic_dialogue_v3_0098", "base_id": "university_registration_0056_closed_derived_missing_antecedent", "split": "dev", "domain": "university_registration", "family": "closed_derived_missing_antecedent", "subset": "core_contrastive", "gold_pattern": "unknown|false|false", "dialogue_type": "persist_after_narrowing", "target_statement": "Lena may register for the seminar.", "query_atom": "may_register::Lena", "facts_positive": ["Lena passed the qualifying exam."], "facts_negative": [], "rules_natural": ["If a student passed the qualifying exam and has advising clearance, then that student may register for the seminar."], "predicate_glossary": {"passed_exam": "passed the qualifying exam", "has_advising_clearance": "has advising clearance", "has_external_scholarship": "has an external scholarship", "may_register": "may register for the seminar", "needs_manual_audit": "needs a manual audit"}, "vocabulary_predicates": ["has_advising_clearance", "has_external_scholarship", "may_register", "needs_manual_audit", "passed_exam"], "symbolic": {"positive_atoms": ["passed_exam::Lena"], "negative_atoms": [], "rules": [{"antecedents": ["passed_exam::Lena", "has_advising_clearance::Lena"], "conclusion": "may_register::Lena", "text": "If a student passed the qualifying exam and has advising clearance, then that student may register for the seminar."}], "query_atom": "may_register::Lena", "closure_atoms": ["passed_exam::Lena"]}, "turns": [{"turn_index": 1, "source_semantics": "owa", "update_operation": "SET_COMPLETE_PREDICATES", "update_predicates": [], "complete_predicates_after_update": [], "gold_answer": "unknown", "expected_changed_from_previous": null, "expected_applied_update": "open", "prompt": "This is turn 1 of a 3-turn dynamic completeness dialogue.\nYou are given an initial knowledge base and an initial completeness state.\nDo not use outside knowledge.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly open. For turn 1, changed_from_previous must be null.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nKnowledge base:\n\nPredicate glossary:\n- has_advising_clearance: has advising clearance\n- has_external_scholarship: has an external scholarship\n- may_register: may register for the seminar\n- needs_manual_audit: needs a manual audit\n- passed_exam: passed the qualifying exam\n\nPositive facts:\n- Lena passed the qualifying exam.\n\nNegative facts:\n- none\n\nRules:\n- If a student passed the qualifying exam and has advising clearance, then that student may register for the seminar.\n\nTarget statement: Lena may register for the seminar.\nTarget atom: may_register::Lena\n\nCompleteness-state update for this turn:\n- Operation: SET_COMPLETE_PREDICATES\n- New complete predicates: none"}, {"turn_index": 2, "source_semantics": "cwa", "update_operation": "ADD_COMPLETE_PREDICATES", "update_predicates": ["has_advising_clearance", "has_external_scholarship", "may_register", "needs_manual_audit", "passed_exam"], "complete_predicates_after_update": ["has_advising_clearance", "has_external_scholarship", "may_register", "needs_manual_audit", "passed_exam"], "gold_answer": "false", "expected_changed_from_previous": true, "expected_applied_update": "global_complete", "prompt": "This is turn 2 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly global_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: ADD_COMPLETE_PREDICATES\n- Predicates to add as complete: has_advising_clearance, has_external_scholarship, may_register, needs_manual_audit, passed_exam\n- Complete predicates after this update: has_advising_clearance, has_external_scholarship, may_register, needs_manual_audit, passed_exam\n\nTarget statement: Lena may register for the seminar.\nTarget atom: may_register::Lena"}, {"turn_index": 3, "source_semantics": "lcwa", "update_operation": "REPLACE_COMPLETE_PREDICATES", "update_predicates": ["has_advising_clearance", "may_register"], "complete_predicates_after_update": ["has_advising_clearance", "may_register"], "gold_answer": "false", "expected_changed_from_previous": false, "expected_applied_update": "local_complete", "prompt": "This is turn 3 of a 3-turn dynamic completeness dialogue.\nKeep the exact same knowledge base, facts, rules, glossary, and target from turn 1.\nApply only the completeness-state update below.\nRecompute the answer from the updated state. Do not treat a previous false/unknown conclusion as a new fact.\nReturn exactly one compact JSON object with keys answer, changed_from_previous, applied_update, and rationale. The answer value must be one of true, false, unknown. The changed_from_previous value must be true, false, or null. The applied_update value must be one of open, global_complete, local_complete. For this turn, applied_update must be exactly local_complete. For this turn, changed_from_previous must be true iff your current answer differs from your immediately previous answer.\n\nReasoning contract:\n- Positive facts are true.\n- Negative facts make the corresponding target atom false.\n- A rule derives its conclusion only when every antecedent is true.\n- First apply all rules until no new facts can be derived.\n- Then evaluate unstated and underivable atoms using the current completeness state.\n- If a predicate is complete, an unstated and underivable atom using that predicate is false.\n- If a predicate is not complete, an unstated and underivable atom using that predicate is unknown.\n- A false answer derived from completeness is not a permanent negative fact; it can be withdrawn if the completeness state changes.\n\nCompleteness-state update for this turn:\n- Operation: REPLACE_COMPLETE_PREDICATES\n- Discard the previous complete-predicate set.\n- Replacement complete predicates: has_advising_clearance, may_register\n\nTarget statement: Lena may register for the seminar.\nTarget atom: may_register::Lena"}]}

metadata/closurebench_ask_act_metadata.json

@@ -0,0 +1,74 @@
+{
+  "source_dataset": "data/base/full.jsonl",
+  "dataset": "data/ask_act/full.jsonl",
+  "version": "closurebench_ask_act",
+  "num_items": 960,
+  "num_base_scenarios": 320,
+  "semantics": {
+    "owa": 320,
+    "cwa": 320,
+    "lcwa": 320
+  },
+  "domains": {
+    "cloud_deployment": 120,
+    "hospital_access": 120,
+    "library_services": 120,
+    "university_registration": 120,
+    "finance_controls": 120,
+    "procurement_review": 120,
+    "manufacturing_quality": 120,
+    "robotics_operations": 120
+  },
+  "families": {
+    "entailed_open_conclusion": 96,
+    "explicit_negative_closed": 96,
+    "explicit_negative_open": 96,
+    "closed_missing_with_open_distractor": 96,
+    "open_missing_with_closed_distractor": 96,
+    "open_missing_direct": 96,
+    "entailed_closed_conclusion": 96,
+    "closed_missing_direct": 96,
+    "closed_derived_missing_antecedent": 96,
+    "open_derived_missing_antecedent": 96
+  },
+  "subsets": {
+    "control_entailed": 192,
+    "control_explicit_negative": 192,
+    "core_contrastive": 576
+  },
+  "gold_truth_values": {
+    "true": 192,
+    "false": 480,
+    "unknown": 288
+  },
+  "gold_actions": {
+    "approve": 192,
+    "deny": 480,
+    "request_information": 288
+  },
+  "gold_reason_types": {
+    "entailed_by_fact_or_rule": 192,
+    "explicit_negative_fact": 192,
+    "global_closed_world_underivable": 192,
+    "local_open_world_underivable": 96,
+    "open_world_underivable": 192,
+    "local_closed_world_underivable": 96
+  },
+  "primary_metrics": [
+    "action_accuracy",
+    "truth_accuracy",
+    "full_decision_accuracy",
+    "ask_precision_recall_f1",
+    "targeted_information_accuracy",
+    "premature_denial_rate",
+    "excessive_ask_rate",
+    "closure_denial_accuracy",
+    "semantic_switch_action_accuracy"
+  ],
+  "notes": [
+    "This file is derived deterministically from ClosureBench V3.",
+    "Unknown direct targets ask for the target atom.",
+    "Unknown derived targets ask for the missing rule antecedent.",
+    "Approve/deny cases require null missing fields to avoid unnecessary escalation."
+  ]
+}

metadata/closurebench_base_metadata.json

@@ -0,0 +1,76 @@
+{
+  "dataset": "closurebench_base.jsonl",
+  "version": "closurebench-v1",
+  "num_items": 960,
+  "num_base_scenarios": 320,
+  "domains": {
+    "cloud_deployment": 120,
+    "hospital_access": 120,
+    "library_services": 120,
+    "university_registration": 120,
+    "finance_controls": 120,
+    "procurement_review": 120,
+    "manufacturing_quality": 120,
+    "robotics_operations": 120
+  },
+  "families": {
+    "entailed_open_conclusion": 96,
+    "explicit_negative_closed": 96,
+    "explicit_negative_open": 96,
+    "closed_missing_with_open_distractor": 96,
+    "open_missing_with_closed_distractor": 96,
+    "open_missing_direct": 96,
+    "entailed_closed_conclusion": 96,
+    "closed_missing_direct": 96,
+    "closed_derived_missing_antecedent": 96,
+    "open_derived_missing_antecedent": 96
+  },
+  "subsets": {
+    "control_entailed": 192,
+    "control_explicit_negative": 192,
+    "core_contrastive": 576
+  },
+  "splits": {
+    "dev": 93,
+    "train": 648,
+    "test": 219
+  },
+  "semantics": {
+    "owa": 320,
+    "cwa": 320,
+    "lcwa": 320
+  },
+  "gold_answers": {
+    "true": 192,
+    "false": 480,
+    "unknown": 288
+  },
+  "gold_reason_types": {
+    "entailed_by_fact_or_rule": 192,
+    "explicit_negative_fact": 192,
+    "global_closed_world_underivable": 192,
+    "local_open_world_underivable": 96,
+    "open_world_underivable": 192,
+    "local_closed_world_underivable": 96
+  },
+  "semantic_switch_patterns": {
+    "false|false|false": 64,
+    "true|true|true": 64,
+    "unknown|false|false": 96,
+    "unknown|false|unknown": 96
+  },
+  "primary_metrics": [
+    "semantic_switch_accuracy",
+    "lcwa_closed_scope_accuracy",
+    "lcwa_open_scope_accuracy",
+    "cwa_accuracy",
+    "owa_accuracy"
+  ],
+  "notes": [
+    "V3 preserves pilot v1/v2 files and should be treated as the publication candidate.",
+    "Each base scenario has OWA, CWA, and LCWA variants with identical facts, rules, and target atom.",
+    "Each item contains natural-language text plus symbolic atoms/rules and a gold reason type.",
+    "Closure is explicit: closed predicates make unstated and underivable atoms false after rule closure.",
+    "Controls are retained but marked separately from core contrastive cases."
+  ]
+}

metadata/closurebench_dynamic_dialogue_metadata.json

@@ -0,0 +1,53 @@
+{
+  "source_dataset": "data/base/full.jsonl",
+  "dataset": "data/dynamic_dialogue/full.jsonl",
+  "n_dialogues": 100,
+  "n_turns": 300,
+  "turns_per_dialogue": 3,
+  "seed": 11,
+  "gold_patterns": {
+    "unknown|false|false": 50,
+    "unknown|false|unknown": 50
+  },
+  "dialogue_types": {
+    "persist_after_narrowing": 50,
+    "reopen": 50
+  },
+  "domains": {
+    "cloud_deployment": 8,
+    "finance_controls": 15,
+    "hospital_access": 13,
+    "library_services": 11,
+    "manufacturing_quality": 13,
+    "procurement_review": 12,
+    "robotics_operations": 15,
+    "university_registration": 13
+  },
+  "families": {
+    "closed_derived_missing_antecedent": 17,
+    "closed_missing_direct": 15,
+    "closed_missing_with_open_distractor": 18,
+    "open_derived_missing_antecedent": 16,
+    "open_missing_direct": 18,
+    "open_missing_with_closed_distractor": 16
+  },
+  "turn_update_sequence": [
+    "SET_COMPLETE_PREDICATES([])",
+    "ADD_COMPLETE_PREDICATES(vocabulary)",
+    "REPLACE_COMPLETE_PREDICATES(local_subset)"
+  ],
+  "primary_metrics": [
+    "dialogue_all_answer_accuracy",
+    "revision_accuracy",
+    "persistence_accuracy",
+    "dialogue_all_turn_accuracy",
+    "completeness_addition_accuracy",
+    "reopening_accuracy",
+    "narrowing_persistence_accuracy",
+    "closure_inertia_rate",
+    "closure_retention_failure_rate",
+    "changed_flag_accuracy",
+    "applied_update_accuracy",
+    "anchoring_error_rate"
+  ]
+}

metadata/closurebench_multi_agent_metadata.json

@@ -0,0 +1,42 @@
+{
+  "name": "ClosureBench-MA",
+  "source_dataset": "data/base/full.jsonl",
+  "num_items": 360,
+  "num_base_scenarios": 120,
+  "semantics": {
+    "cwa": 120,
+    "lcwa": 120,
+    "owa": 120
+  },
+  "subsets": {
+    "core_contrastive": 216,
+    "control_entailed": 72,
+    "control_explicit_negative": 72
+  },
+  "gold_truth_values": {
+    "false": 178,
+    "unknown": 110,
+    "true": 72
+  },
+  "gold_source_used": {
+    "agent_a": 166,
+    "agent_b": 179,
+    "both": 15
+  },
+  "gold_closure_handling": {
+    "closed_absence": 106,
+    "open_absence": 110,
+    "entailed_or_explicit": 72,
+    "explicit_negative": 72
+  },
+  "domains": {
+    "cloud_deployment": 45,
+    "finance_controls": 45,
+    "hospital_access": 42,
+    "library_services": 30,
+    "manufacturing_quality": 60,
+    "procurement_review": 45,
+    "robotics_operations": 48,
+    "university_registration": 45
+  }
+}

results/reports/closurebench_ask_act_replicate_manifest.json

@@ -0,0 +1,116 @@
+{
+  "dataset": "data/ask_act/full.jsonl",
+  "dataset_size": 960,
+  "repeats": 3,
+  "shards": 4,
+  "runs": [
+    {
+      "label": "mistral-small",
+      "model": "mistralai/mistral-small-2603",
+      "repeat": 1,
+      "scored": "results/scored/ask_act_replicates/closure_contract_v3_ask_act_mistral-small_repeat_1_scored.jsonl"
+    },
+    {
+      "label": "mistral-small",
+      "model": "mistralai/mistral-small-2603",
+      "repeat": 2,
+      "scored": "results/scored/ask_act_replicates/closure_contract_v3_ask_act_mistral-small_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "mistral-small",
+      "model": "mistralai/mistral-small-2603",
+      "repeat": 3,
+      "scored": "results/scored/ask_act_replicates/closure_contract_v3_ask_act_mistral-small_repeat_3_scored.jsonl"
+    },
+    {
+      "label": "deepseek-flash",
+      "model": "deepseek-v4-flash",
+      "repeat": 1,
+      "scored": "results/scored/ask_act_replicates/closure_contract_v3_ask_act_deepseek-flash_repeat_1_scored.jsonl"
+    },
+    {
+      "label": "deepseek-flash",
+      "model": "deepseek-v4-flash",
+      "repeat": 2,
+      "scored": "results/scored/ask_act_replicates/closure_contract_v3_ask_act_deepseek-flash_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "deepseek-flash",
+      "model": "deepseek-v4-flash",
+      "repeat": 3,
+      "scored": "results/scored/ask_act_replicates/closure_contract_v3_ask_act_deepseek-flash_repeat_3_scored.jsonl"
+    },
+    {
+      "label": "deepseek-pro",
+      "model": "deepseek-v4-pro",
+      "repeat": 1,
+      "scored": "results/scored/ask_act_replicates/closure_contract_v3_ask_act_deepseek-pro_repeat_1_scored.jsonl"
+    },
+    {
+      "label": "deepseek-pro",
+      "model": "deepseek-v4-pro",
+      "repeat": 2,
+      "scored": "results/scored/ask_act_replicates/closure_contract_v3_ask_act_deepseek-pro_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "deepseek-pro",
+      "model": "deepseek-v4-pro",
+      "repeat": 3,
+      "scored": "results/scored/ask_act_replicates/closure_contract_v3_ask_act_deepseek-pro_repeat_3_scored.jsonl"
+    },
+    {
+      "label": "mistral-medium",
+      "model": "mistralai/mistral-medium-3-5",
+      "repeat": 1,
+      "scored": "results/scored/ask_act_replicates/closure_contract_v3_ask_act_mistral-medium_repeat_1_scored.jsonl"
+    },
+    {
+      "label": "mistral-medium",
+      "model": "mistralai/mistral-medium-3-5",
+      "repeat": 2,
+      "scored": "results/scored/ask_act_replicates/closure_contract_v3_ask_act_mistral-medium_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "mistral-medium",
+      "model": "mistralai/mistral-medium-3-5",
+      "repeat": 3,
+      "scored": "results/scored/ask_act_replicates/closure_contract_v3_ask_act_mistral-medium_repeat_3_scored.jsonl"
+    },
+    {
+      "label": "llama-scout",
+      "model": "meta-llama/llama-4-scout",
+      "repeat": 1,
+      "scored": "results/scored/ask_act_replicates/closure_contract_v3_ask_act_llama-scout_repeat_1_scored.jsonl"
+    },
+    {
+      "label": "llama-scout",
+      "model": "meta-llama/llama-4-scout",
+      "repeat": 2,
+      "scored": "results/scored/ask_act_replicates/closure_contract_v3_ask_act_llama-scout_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "llama-scout",
+      "model": "meta-llama/llama-4-scout",
+      "repeat": 3,
+      "scored": "results/scored/ask_act_replicates/closure_contract_v3_ask_act_llama-scout_repeat_3_scored.jsonl"
+    },
+    {
+      "label": "llama-maverick",
+      "model": "meta-llama/llama-4-maverick",
+      "repeat": 1,
+      "scored": "results/scored/ask_act_replicates/closure_contract_v3_ask_act_llama-maverick_repeat_1_scored.jsonl"
+    },
+    {
+      "label": "llama-maverick",
+      "model": "meta-llama/llama-4-maverick",
+      "repeat": 2,
+      "scored": "results/scored/ask_act_replicates/closure_contract_v3_ask_act_llama-maverick_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "llama-maverick",
+      "model": "meta-llama/llama-4-maverick",
+      "repeat": 3,
+      "scored": "results/scored/ask_act_replicates/closure_contract_v3_ask_act_llama-maverick_repeat_3_scored.jsonl"
+    }
+  ]
+}

results/reports/closurebench_deepseek_v4_flash_report.json

@@ -0,0 +1,391 @@
+{
+  "n_scored": 960,
+  "n_dataset": 960,
+  "coverage": 100.0,
+  "api_errors": 0,
+  "json_valid": {
+    "n": 959,
+    "rate": 99.9
+  },
+  "answer_parse": {
+    "n": 960,
+    "rate": 100.0
+  },
+  "overall_accuracy": {
+    "correct": 872,
+    "n": 960,
+    "accuracy": 90.83
+  },
+  "by_semantics": {
+    "cwa": {
+      "n": 320,
+      "correct": 280,
+      "accuracy": 87.5
+    },
+    "lcwa": {
+      "n": 320,
+      "correct": 272,
+      "accuracy": 85.0
+    },
+    "owa": {
+      "n": 320,
+      "correct": 320,
+      "accuracy": 100.0
+    }
+  },
+  "by_subset": {
+    "control_entailed": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "control_explicit_negative": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "core_contrastive": {
+      "n": 576,
+      "correct": 488,
+      "accuracy": 84.72
+    }
+  },
+  "by_family": {
+    "closed_derived_missing_antecedent": {
+      "n": 96,
+      "correct": 33,
+      "accuracy": 34.38
+    },
+    "closed_missing_direct": {
+      "n": 96,
+      "correct": 86,
+      "accuracy": 89.58
+    },
+    "closed_missing_with_open_distractor": {
+      "n": 96,
+      "correct": 88,
+      "accuracy": 91.67
+    },
+    "entailed_closed_conclusion": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "entailed_open_conclusion": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "explicit_negative_closed": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "explicit_negative_open": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "open_derived_missing_antecedent": {
+      "n": 96,
+      "correct": 89,
+      "accuracy": 92.71
+    },
+    "open_missing_direct": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "open_missing_with_closed_distractor": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    }
+  },
+  "by_family_and_semantics": {
+    "closed_derived_missing_antecedent": {
+      "cwa": {
+        "n": 32,
+        "correct": 1,
+        "accuracy": 3.12
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 0,
+        "accuracy": 0.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "closed_missing_direct": {
+      "cwa": {
+        "n": 32,
+        "correct": 31,
+        "accuracy": 96.88
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 23,
+        "accuracy": 71.88
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "closed_missing_with_open_distractor": {
+      "cwa": {
+        "n": 32,
+        "correct": 31,
+        "accuracy": 96.88
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 25,
+        "accuracy": 78.12
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "entailed_closed_conclusion": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "entailed_open_conclusion": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "explicit_negative_closed": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "explicit_negative_open": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "open_derived_missing_antecedent": {
+      "cwa": {
+        "n": 32,
+        "correct": 25,
+        "accuracy": 78.12
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "open_missing_direct": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "open_missing_with_closed_distractor": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    }
+  },
+  "by_gold_answer": {
+    "false": {
+      "n": 480,
+      "correct": 392,
+      "accuracy": 81.67
+    },
+    "true": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "unknown": {
+      "n": 288,
+      "correct": 288,
+      "accuracy": 100.0
+    }
+  },
+  "by_gold_reason_type": {
+    "entailed_by_fact_or_rule": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "explicit_negative_fact": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "global_closed_world_underivable": {
+      "n": 192,
+      "correct": 152,
+      "accuracy": 79.17
+    },
+    "local_closed_world_underivable": {
+      "n": 96,
+      "correct": 48,
+      "accuracy": 50.0
+    },
+    "local_open_world_underivable": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "open_world_underivable": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    }
+  },
+  "semantic_switch_accuracy": {
+    "complete_base_groups": 320,
+    "correct_groups": 264,
+    "accuracy": 82.5
+  },
+  "core_contrastive_switch_accuracy": {
+    "complete_base_groups": 192,
+    "correct_groups": 136,
+    "accuracy": 70.83
+  },
+  "unknown_accuracy": {
+    "correct": 288,
+    "n": 288,
+    "accuracy": 100.0
+  },
+  "false_accuracy": {
+    "correct": 392,
+    "n": 480,
+    "accuracy": 81.67
+  },
+  "true_accuracy": {
+    "correct": 192,
+    "n": 192,
+    "accuracy": 100.0
+  },
+  "lcwa_open_scope_accuracy": {
+    "correct": 96,
+    "n": 96,
+    "accuracy": 100.0
+  },
+  "lcwa_closed_scope_accuracy": {
+    "correct": 48,
+    "n": 96,
+    "accuracy": 50.0
+  },
+  "semantic_bias_index": {
+    "owa_unknown_predicted_false_rate": 0.0,
+    "lcwa_open_predicted_false_rate": 0.0,
+    "lcwa_closed_predicted_unknown_rate": 50.0
+  },
+  "error_types": {
+    "correct": 872,
+    "open_world_bias": 40,
+    "missed_local_closure": 48
+  },
+  "baselines": {
+    "always_cwa": {
+      "correct": 672,
+      "accuracy": 70.0,
+      "n": 960
+    },
+    "always_owa": {
+      "correct": 672,
+      "accuracy": 70.0,
+      "n": 960
+    },
+    "ignore_lcwa": {
+      "correct": 864,
+      "accuracy": 90.0,
+      "n": 960
+    }
+  },
+  "models": [
+    "deepseek-v4-flash"
+  ]
+}

results/reports/closurebench_deepseek_v4_pro_report.json

@@ -0,0 +1,391 @@
+{
+  "n_scored": 960,
+  "n_dataset": 960,
+  "coverage": 100.0,
+  "api_errors": 0,
+  "json_valid": {
+    "n": 958,
+    "rate": 99.79
+  },
+  "answer_parse": {
+    "n": 960,
+    "rate": 100.0
+  },
+  "overall_accuracy": {
+    "correct": 898,
+    "n": 960,
+    "accuracy": 93.54
+  },
+  "by_semantics": {
+    "cwa": {
+      "n": 320,
+      "correct": 294,
+      "accuracy": 91.88
+    },
+    "lcwa": {
+      "n": 320,
+      "correct": 284,
+      "accuracy": 88.75
+    },
+    "owa": {
+      "n": 320,
+      "correct": 320,
+      "accuracy": 100.0
+    }
+  },
+  "by_subset": {
+    "control_entailed": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "control_explicit_negative": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "core_contrastive": {
+      "n": 576,
+      "correct": 514,
+      "accuracy": 89.24
+    }
+  },
+  "by_family": {
+    "closed_derived_missing_antecedent": {
+      "n": 96,
+      "correct": 47,
+      "accuracy": 48.96
+    },
+    "closed_missing_direct": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "closed_missing_with_open_distractor": {
+      "n": 96,
+      "correct": 91,
+      "accuracy": 94.79
+    },
+    "entailed_closed_conclusion": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "entailed_open_conclusion": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "explicit_negative_closed": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "explicit_negative_open": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "open_derived_missing_antecedent": {
+      "n": 96,
+      "correct": 88,
+      "accuracy": 91.67
+    },
+    "open_missing_direct": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "open_missing_with_closed_distractor": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    }
+  },
+  "by_family_and_semantics": {
+    "closed_derived_missing_antecedent": {
+      "cwa": {
+        "n": 32,
+        "correct": 15,
+        "accuracy": 46.88
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 0,
+        "accuracy": 0.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "closed_missing_direct": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "closed_missing_with_open_distractor": {
+      "cwa": {
+        "n": 32,
+        "correct": 31,
+        "accuracy": 96.88
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 28,
+        "accuracy": 87.5
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "entailed_closed_conclusion": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "entailed_open_conclusion": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "explicit_negative_closed": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "explicit_negative_open": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "open_derived_missing_antecedent": {
+      "cwa": {
+        "n": 32,
+        "correct": 24,
+        "accuracy": 75.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "open_missing_direct": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "open_missing_with_closed_distractor": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    }
+  },
+  "by_gold_answer": {
+    "false": {
+      "n": 480,
+      "correct": 418,
+      "accuracy": 87.08
+    },
+    "true": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "unknown": {
+      "n": 288,
+      "correct": 288,
+      "accuracy": 100.0
+    }
+  },
+  "by_gold_reason_type": {
+    "entailed_by_fact_or_rule": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "explicit_negative_fact": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "global_closed_world_underivable": {
+      "n": 192,
+      "correct": 166,
+      "accuracy": 86.46
+    },
+    "local_closed_world_underivable": {
+      "n": 96,
+      "correct": 60,
+      "accuracy": 62.5
+    },
+    "local_open_world_underivable": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "open_world_underivable": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    }
+  },
+  "semantic_switch_accuracy": {
+    "complete_base_groups": 320,
+    "correct_groups": 276,
+    "accuracy": 86.25
+  },
+  "core_contrastive_switch_accuracy": {
+    "complete_base_groups": 192,
+    "correct_groups": 148,
+    "accuracy": 77.08
+  },
+  "unknown_accuracy": {
+    "correct": 288,
+    "n": 288,
+    "accuracy": 100.0
+  },
+  "false_accuracy": {
+    "correct": 418,
+    "n": 480,
+    "accuracy": 87.08
+  },
+  "true_accuracy": {
+    "correct": 192,
+    "n": 192,
+    "accuracy": 100.0
+  },
+  "lcwa_open_scope_accuracy": {
+    "correct": 96,
+    "n": 96,
+    "accuracy": 100.0
+  },
+  "lcwa_closed_scope_accuracy": {
+    "correct": 60,
+    "n": 96,
+    "accuracy": 62.5
+  },
+  "semantic_bias_index": {
+    "owa_unknown_predicted_false_rate": 0.0,
+    "lcwa_open_predicted_false_rate": 0.0,
+    "lcwa_closed_predicted_unknown_rate": 37.5
+  },
+  "error_types": {
+    "correct": 898,
+    "open_world_bias": 26,
+    "missed_local_closure": 36
+  },
+  "baselines": {
+    "always_cwa": {
+      "correct": 672,
+      "accuracy": 70.0,
+      "n": 960
+    },
+    "always_owa": {
+      "correct": 672,
+      "accuracy": 70.0,
+      "n": 960
+    },
+    "ignore_lcwa": {
+      "correct": 864,
+      "accuracy": 90.0,
+      "n": 960
+    }
+  },
+  "models": [
+    "deepseek-v4-pro"
+  ]
+}

results/reports/closurebench_dynamic_dialogue_replicate_manifest.json

@@ -0,0 +1,116 @@
+{
+  "dataset": "data/dynamic_dialogue/full.jsonl",
+  "dataset_turns": 300,
+  "repeats": 3,
+  "shards": 4,
+  "runs": [
+    {
+      "label": "mistral-small",
+      "model": "mistralai/mistral-small-2603",
+      "repeat": 1,
+      "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_mistral-small_repeat_1_scored.jsonl"
+    },
+    {
+      "label": "mistral-small",
+      "model": "mistralai/mistral-small-2603",
+      "repeat": 2,
+      "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_mistral-small_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "mistral-small",
+      "model": "mistralai/mistral-small-2603",
+      "repeat": 3,
+      "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_mistral-small_repeat_3_scored.jsonl"
+    },
+    {
+      "label": "deepseek-flash",
+      "model": "deepseek-v4-flash",
+      "repeat": 1,
+      "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_deepseek-flash_repeat_1_scored.jsonl"
+    },
+    {
+      "label": "deepseek-flash",
+      "model": "deepseek-v4-flash",
+      "repeat": 2,
+      "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_deepseek-flash_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "deepseek-flash",
+      "model": "deepseek-v4-flash",
+      "repeat": 3,
+      "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_deepseek-flash_repeat_3_scored.jsonl"
+    },
+    {
+      "label": "deepseek-pro",
+      "model": "deepseek-v4-pro",
+      "repeat": 1,
+      "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_deepseek-pro_repeat_1_scored.jsonl"
+    },
+    {
+      "label": "deepseek-pro",
+      "model": "deepseek-v4-pro",
+      "repeat": 2,
+      "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_deepseek-pro_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "deepseek-pro",
+      "model": "deepseek-v4-pro",
+      "repeat": 3,
+      "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_deepseek-pro_repeat_3_scored.jsonl"
+    },
+    {
+      "label": "mistral-medium",
+      "model": "mistralai/mistral-medium-3-5",
+      "repeat": 1,
+      "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_mistral-medium_repeat_1_scored.jsonl"
+    },
+    {
+      "label": "mistral-medium",
+      "model": "mistralai/mistral-medium-3-5",
+      "repeat": 2,
+      "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_mistral-medium_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "mistral-medium",
+      "model": "mistralai/mistral-medium-3-5",
+      "repeat": 3,
+      "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_mistral-medium_repeat_3_scored.jsonl"
+    },
+    {
+      "label": "llama-scout",
+      "model": "meta-llama/llama-4-scout",
+      "repeat": 1,
+      "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_llama-scout_repeat_1_scored.jsonl"
+    },
+    {
+      "label": "llama-scout",
+      "model": "meta-llama/llama-4-scout",
+      "repeat": 2,
+      "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_llama-scout_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "llama-scout",
+      "model": "meta-llama/llama-4-scout",
+      "repeat": 3,
+      "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_llama-scout_repeat_3_scored.jsonl"
+    },
+    {
+      "label": "llama-maverick",
+      "model": "meta-llama/llama-4-maverick",
+      "repeat": 1,
+      "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_llama-maverick_repeat_1_scored.jsonl"
+    },
+    {
+      "label": "llama-maverick",
+      "model": "meta-llama/llama-4-maverick",
+      "repeat": 2,
+      "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_llama-maverick_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "llama-maverick",
+      "model": "meta-llama/llama-4-maverick",
+      "repeat": 3,
+      "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_llama-maverick_repeat_3_scored.jsonl"
+    }
+  ]
+}

results/reports/closurebench_dynamic_dialogue_summary.json

@@ -0,0 +1,3508 @@
+{
+  "dataset": "data/dynamic_dialogue/full.jsonl",
+  "n_dialogues": 100,
+  "n_turns": 300,
+  "manifest": "results/reports/closurebench_dynamic_dialogue_replicate_manifest.json",
+  "runs": {
+    "deepseek-flash": {
+      "model": "deepseek-v4-flash",
+      "n_repeats": 3,
+      "replicates": [
+        {
+          "n_dialogues": 100,
+          "n_turns": 300,
+          "n_scored": 300,
+          "coverage": 100.0,
+          "models": [
+            "deepseek-v4-flash"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "json_valid": {
+            "n": 295,
+            "rate": 98.33
+          },
+          "answer_accuracy": {
+            "correct": 290,
+            "n": 300,
+            "accuracy": 96.67
+          },
+          "full_turn_accuracy": {
+            "correct": 290,
+            "n": 300,
+            "accuracy": 96.67
+          },
+          "dialogue_all_answer_accuracy": {
+            "correct": 90,
+            "n": 100,
+            "accuracy": 90.0
+          },
+          "dialogue_all_turn_accuracy": {
+            "correct": 90,
+            "n": 100,
+            "accuracy": 90.0
+          },
+          "revision_accuracy": {
+            "correct": 150,
+            "n": 150,
+            "accuracy": 100.0
+          },
+          "persistence_accuracy": {
+            "correct": 40,
+            "n": 50,
+            "accuracy": 80.0
+          },
+          "completeness_addition_accuracy": {
+            "correct": 100,
+            "n": 100,
+            "accuracy": 100.0
+          },
+          "reopening_accuracy": {
+            "correct": 50,
+            "n": 50,
+            "accuracy": 100.0
+          },
+          "narrowing_persistence_accuracy": {
+            "correct": 40,
+            "n": 50,
+            "accuracy": 80.0
+          },
+          "changed_flag_accuracy": {
+            "correct": 190,
+            "n": 200,
+            "accuracy": 95.0
+          },
+          "applied_update_accuracy": {
+            "correct": 300,
+            "n": 300,
+            "accuracy": 100.0
+          },
+          "anchoring_error_rate": {
+            "errors": 0,
+            "n": 150,
+            "rate": 0.0
+          },
+          "closure_inertia_rate": {
+            "errors": 0,
+            "n": 50,
+            "rate": 0.0
+          },
+          "closure_retention_failure_rate": {
+            "errors": 10,
+            "n": 50,
+            "rate": 20.0
+          },
+          "by_turn": {
+            "1": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "2": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "3": {
+              "correct": 90,
+              "n": 100,
+              "accuracy": 90.0
+            }
+          },
+          "by_dialogue_type": {
+            "persist_after_narrowing": {
+              "correct": 140,
+              "n": 150,
+              "accuracy": 93.33
+            },
+            "reopen": {
+              "correct": 150,
+              "n": 150,
+              "accuracy": 100.0
+            }
+          },
+          "by_update_operation": {
+            "ADD_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "REPLACE_COMPLETE_PREDICATES": {
+              "correct": 90,
+              "n": 100,
+              "accuracy": 90.0
+            },
+            "SET_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            }
+          },
+          "error_types": {
+            "correct": 290,
+            "lost_relevant_local_closure": 10
+          },
+          "repeat": 1,
+          "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_deepseek-flash_repeat_1_scored.jsonl"
+        },
+        {
+          "n_dialogues": 100,
+          "n_turns": 300,
+          "n_scored": 300,
+          "coverage": 100.0,
+          "models": [
+            "deepseek-v4-flash"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "json_valid": {
+            "n": 299,
+            "rate": 99.67
+          },
+          "answer_accuracy": {
+            "correct": 291,
+            "n": 300,
+            "accuracy": 97.0
+          },
+          "full_turn_accuracy": {
+            "correct": 291,
+            "n": 300,
+            "accuracy": 97.0
+          },
+          "dialogue_all_answer_accuracy": {
+            "correct": 91,
+            "n": 100,
+            "accuracy": 91.0
+          },
+          "dialogue_all_turn_accuracy": {
+            "correct": 91,
+            "n": 100,
+            "accuracy": 91.0
+          },
+          "revision_accuracy": {
+            "correct": 150,
+            "n": 150,
+            "accuracy": 100.0
+          },
+          "persistence_accuracy": {
+            "correct": 41,
+            "n": 50,
+            "accuracy": 82.0
+          },
+          "completeness_addition_accuracy": {
+            "correct": 100,
+            "n": 100,
+            "accuracy": 100.0
+          },
+          "reopening_accuracy": {
+            "correct": 50,
+            "n": 50,
+            "accuracy": 100.0
+          },
+          "narrowing_persistence_accuracy": {
+            "correct": 41,
+            "n": 50,
+            "accuracy": 82.0
+          },
+          "changed_flag_accuracy": {
+            "correct": 191,
+            "n": 200,
+            "accuracy": 95.5
+          },
+          "applied_update_accuracy": {
+            "correct": 300,
+            "n": 300,
+            "accuracy": 100.0
+          },
+          "anchoring_error_rate": {
+            "errors": 0,
+            "n": 150,
+            "rate": 0.0
+          },
+          "closure_inertia_rate": {
+            "errors": 0,
+            "n": 50,
+            "rate": 0.0
+          },
+          "closure_retention_failure_rate": {
+            "errors": 9,
+            "n": 50,
+            "rate": 18.0
+          },
+          "by_turn": {
+            "1": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "2": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "3": {
+              "correct": 91,
+              "n": 100,
+              "accuracy": 91.0
+            }
+          },
+          "by_dialogue_type": {
+            "persist_after_narrowing": {
+              "correct": 141,
+              "n": 150,
+              "accuracy": 94.0
+            },
+            "reopen": {
+              "correct": 150,
+              "n": 150,
+              "accuracy": 100.0
+            }
+          },
+          "by_update_operation": {
+            "ADD_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "REPLACE_COMPLETE_PREDICATES": {
+              "correct": 91,
+              "n": 100,
+              "accuracy": 91.0
+            },
+            "SET_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            }
+          },
+          "error_types": {
+            "correct": 291,
+            "lost_relevant_local_closure": 9
+          },
+          "repeat": 2,
+          "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_deepseek-flash_repeat_2_scored.jsonl"
+        },
+        {
+          "n_dialogues": 100,
+          "n_turns": 300,
+          "n_scored": 300,
+          "coverage": 100.0,
+          "models": [
+            "deepseek-v4-flash"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "json_valid": {
+            "n": 299,
+            "rate": 99.67
+          },
+          "answer_accuracy": {
+            "correct": 292,
+            "n": 300,
+            "accuracy": 97.33
+          },
+          "full_turn_accuracy": {
+            "correct": 292,
+            "n": 300,
+            "accuracy": 97.33
+          },
+          "dialogue_all_answer_accuracy": {
+            "correct": 92,
+            "n": 100,
+            "accuracy": 92.0
+          },
+          "dialogue_all_turn_accuracy": {
+            "correct": 92,
+            "n": 100,
+            "accuracy": 92.0
+          },
+          "revision_accuracy": {
+            "correct": 150,
+            "n": 150,
+            "accuracy": 100.0
+          },
+          "persistence_accuracy": {
+            "correct": 42,
+            "n": 50,
+            "accuracy": 84.0
+          },
+          "completeness_addition_accuracy": {
+            "correct": 100,
+            "n": 100,
+            "accuracy": 100.0
+          },
+          "reopening_accuracy": {
+            "correct": 50,
+            "n": 50,
+            "accuracy": 100.0
+          },
+          "narrowing_persistence_accuracy": {
+            "correct": 42,
+            "n": 50,
+            "accuracy": 84.0
+          },
+          "changed_flag_accuracy": {
+            "correct": 192,
+            "n": 200,
+            "accuracy": 96.0
+          },
+          "applied_update_accuracy": {
+            "correct": 300,
+            "n": 300,
+            "accuracy": 100.0
+          },
+          "anchoring_error_rate": {
+            "errors": 0,
+            "n": 150,
+            "rate": 0.0
+          },
+          "closure_inertia_rate": {
+            "errors": 0,
+            "n": 50,
+            "rate": 0.0
+          },
+          "closure_retention_failure_rate": {
+            "errors": 8,
+            "n": 50,
+            "rate": 16.0
+          },
+          "by_turn": {
+            "1": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "2": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "3": {
+              "correct": 92,
+              "n": 100,
+              "accuracy": 92.0
+            }
+          },
+          "by_dialogue_type": {
+            "persist_after_narrowing": {
+              "correct": 142,
+              "n": 150,
+              "accuracy": 94.67
+            },
+            "reopen": {
+              "correct": 150,
+              "n": 150,
+              "accuracy": 100.0
+            }
+          },
+          "by_update_operation": {
+            "ADD_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "REPLACE_COMPLETE_PREDICATES": {
+              "correct": 92,
+              "n": 100,
+              "accuracy": 92.0
+            },
+            "SET_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            }
+          },
+          "error_types": {
+            "correct": 292,
+            "lost_relevant_local_closure": 8
+          },
+          "repeat": 3,
+          "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_deepseek-flash_repeat_3_scored.jsonl"
+        }
+      ],
+      "metrics": {
+        "dialogue_all_answer_accuracy": {
+          "mean": 91.0,
+          "sd": 1.0,
+          "ci95_half_width": 2.48,
+          "min": 90.0,
+          "max": 92.0,
+          "values": [
+            90.0,
+            91.0,
+            92.0
+          ]
+        },
+        "dialogue_all_turn_accuracy": {
+          "mean": 91.0,
+          "sd": 1.0,
+          "ci95_half_width": 2.48,
+          "min": 90.0,
+          "max": 92.0,
+          "values": [
+            90.0,
+            91.0,
+            92.0
+          ]
+        },
+        "answer_accuracy": {
+          "mean": 97.0,
+          "sd": 0.33,
+          "ci95_half_width": 0.82,
+          "min": 96.67,
+          "max": 97.33,
+          "values": [
+            96.67,
+            97.0,
+            97.33
+          ]
+        },
+        "revision_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "persistence_accuracy": {
+          "mean": 82.0,
+          "sd": 2.0,
+          "ci95_half_width": 4.97,
+          "min": 80.0,
+          "max": 84.0,
+          "values": [
+            80.0,
+            82.0,
+            84.0
+          ]
+        },
+        "completeness_addition_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "reopening_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "narrowing_persistence_accuracy": {
+          "mean": 82.0,
+          "sd": 2.0,
+          "ci95_half_width": 4.97,
+          "min": 80.0,
+          "max": 84.0,
+          "values": [
+            80.0,
+            82.0,
+            84.0
+          ]
+        },
+        "changed_flag_accuracy": {
+          "mean": 95.5,
+          "sd": 0.5,
+          "ci95_half_width": 1.24,
+          "min": 95.0,
+          "max": 96.0,
+          "values": [
+            95.0,
+            95.5,
+            96.0
+          ]
+        },
+        "applied_update_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "closure_inertia_rate": {
+          "mean": 0.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 0.0,
+          "max": 0.0,
+          "values": [
+            0.0,
+            0.0,
+            0.0
+          ]
+        },
+        "closure_retention_failure_rate": {
+          "mean": 18.0,
+          "sd": 2.0,
+          "ci95_half_width": 4.97,
+          "min": 16.0,
+          "max": 20.0,
+          "values": [
+            20.0,
+            18.0,
+            16.0
+          ]
+        },
+        "anchoring_error_rate": {
+          "mean": 0.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 0.0,
+          "max": 0.0,
+          "values": [
+            0.0,
+            0.0,
+            0.0
+          ]
+        }
+      },
+      "turn_decision_consistency": {
+        "n": 300,
+        "consistent": 292,
+        "rate": 97.33
+      },
+      "mean_parse_failures": 0.0,
+      "mean_api_errors": 0.0
+    },
+    "deepseek-pro": {
+      "model": "deepseek-v4-pro",
+      "n_repeats": 3,
+      "replicates": [
+        {
+          "n_dialogues": 100,
+          "n_turns": 300,
+          "n_scored": 300,
+          "coverage": 100.0,
+          "models": [
+            "deepseek-v4-pro"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "json_valid": {
+            "n": 300,
+            "rate": 100.0
+          },
+          "answer_accuracy": {
+            "correct": 297,
+            "n": 300,
+            "accuracy": 99.0
+          },
+          "full_turn_accuracy": {
+            "correct": 297,
+            "n": 300,
+            "accuracy": 99.0
+          },
+          "dialogue_all_answer_accuracy": {
+            "correct": 97,
+            "n": 100,
+            "accuracy": 97.0
+          },
+          "dialogue_all_turn_accuracy": {
+            "correct": 97,
+            "n": 100,
+            "accuracy": 97.0
+          },
+          "revision_accuracy": {
+            "correct": 150,
+            "n": 150,
+            "accuracy": 100.0
+          },
+          "persistence_accuracy": {
+            "correct": 47,
+            "n": 50,
+            "accuracy": 94.0
+          },
+          "completeness_addition_accuracy": {
+            "correct": 100,
+            "n": 100,
+            "accuracy": 100.0
+          },
+          "reopening_accuracy": {
+            "correct": 50,
+            "n": 50,
+            "accuracy": 100.0
+          },
+          "narrowing_persistence_accuracy": {
+            "correct": 47,
+            "n": 50,
+            "accuracy": 94.0
+          },
+          "changed_flag_accuracy": {
+            "correct": 197,
+            "n": 200,
+            "accuracy": 98.5
+          },
+          "applied_update_accuracy": {
+            "correct": 300,
+            "n": 300,
+            "accuracy": 100.0
+          },
+          "anchoring_error_rate": {
+            "errors": 0,
+            "n": 150,
+            "rate": 0.0
+          },
+          "closure_inertia_rate": {
+            "errors": 0,
+            "n": 50,
+            "rate": 0.0
+          },
+          "closure_retention_failure_rate": {
+            "errors": 3,
+            "n": 50,
+            "rate": 6.0
+          },
+          "by_turn": {
+            "1": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "2": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "3": {
+              "correct": 97,
+              "n": 100,
+              "accuracy": 97.0
+            }
+          },
+          "by_dialogue_type": {
+            "persist_after_narrowing": {
+              "correct": 147,
+              "n": 150,
+              "accuracy": 98.0
+            },
+            "reopen": {
+              "correct": 150,
+              "n": 150,
+              "accuracy": 100.0
+            }
+          },
+          "by_update_operation": {
+            "ADD_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "REPLACE_COMPLETE_PREDICATES": {
+              "correct": 97,
+              "n": 100,
+              "accuracy": 97.0
+            },
+            "SET_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            }
+          },
+          "error_types": {
+            "correct": 297,
+            "lost_relevant_local_closure": 3
+          },
+          "repeat": 1,
+          "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_deepseek-pro_repeat_1_scored.jsonl"
+        },
+        {
+          "n_dialogues": 100,
+          "n_turns": 300,
+          "n_scored": 300,
+          "coverage": 100.0,
+          "models": [
+            "deepseek-v4-pro"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "json_valid": {
+            "n": 300,
+            "rate": 100.0
+          },
+          "answer_accuracy": {
+            "correct": 299,
+            "n": 300,
+            "accuracy": 99.67
+          },
+          "full_turn_accuracy": {
+            "correct": 299,
+            "n": 300,
+            "accuracy": 99.67
+          },
+          "dialogue_all_answer_accuracy": {
+            "correct": 99,
+            "n": 100,
+            "accuracy": 99.0
+          },
+          "dialogue_all_turn_accuracy": {
+            "correct": 99,
+            "n": 100,
+            "accuracy": 99.0
+          },
+          "revision_accuracy": {
+            "correct": 150,
+            "n": 150,
+            "accuracy": 100.0
+          },
+          "persistence_accuracy": {
+            "correct": 49,
+            "n": 50,
+            "accuracy": 98.0
+          },
+          "completeness_addition_accuracy": {
+            "correct": 100,
+            "n": 100,
+            "accuracy": 100.0
+          },
+          "reopening_accuracy": {
+            "correct": 50,
+            "n": 50,
+            "accuracy": 100.0
+          },
+          "narrowing_persistence_accuracy": {
+            "correct": 49,
+            "n": 50,
+            "accuracy": 98.0
+          },
+          "changed_flag_accuracy": {
+            "correct": 199,
+            "n": 200,
+            "accuracy": 99.5
+          },
+          "applied_update_accuracy": {
+            "correct": 300,
+            "n": 300,
+            "accuracy": 100.0
+          },
+          "anchoring_error_rate": {
+            "errors": 0,
+            "n": 150,
+            "rate": 0.0
+          },
+          "closure_inertia_rate": {
+            "errors": 0,
+            "n": 50,
+            "rate": 0.0
+          },
+          "closure_retention_failure_rate": {
+            "errors": 1,
+            "n": 50,
+            "rate": 2.0
+          },
+          "by_turn": {
+            "1": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "2": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "3": {
+              "correct": 99,
+              "n": 100,
+              "accuracy": 99.0
+            }
+          },
+          "by_dialogue_type": {
+            "persist_after_narrowing": {
+              "correct": 149,
+              "n": 150,
+              "accuracy": 99.33
+            },
+            "reopen": {
+              "correct": 150,
+              "n": 150,
+              "accuracy": 100.0
+            }
+          },
+          "by_update_operation": {
+            "ADD_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "REPLACE_COMPLETE_PREDICATES": {
+              "correct": 99,
+              "n": 100,
+              "accuracy": 99.0
+            },
+            "SET_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            }
+          },
+          "error_types": {
+            "correct": 299,
+            "lost_relevant_local_closure": 1
+          },
+          "repeat": 2,
+          "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_deepseek-pro_repeat_2_scored.jsonl"
+        },
+        {
+          "n_dialogues": 100,
+          "n_turns": 300,
+          "n_scored": 300,
+          "coverage": 100.0,
+          "models": [
+            "deepseek-v4-pro"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "json_valid": {
+            "n": 300,
+            "rate": 100.0
+          },
+          "answer_accuracy": {
+            "correct": 298,
+            "n": 300,
+            "accuracy": 99.33
+          },
+          "full_turn_accuracy": {
+            "correct": 298,
+            "n": 300,
+            "accuracy": 99.33
+          },
+          "dialogue_all_answer_accuracy": {
+            "correct": 98,
+            "n": 100,
+            "accuracy": 98.0
+          },
+          "dialogue_all_turn_accuracy": {
+            "correct": 98,
+            "n": 100,
+            "accuracy": 98.0
+          },
+          "revision_accuracy": {
+            "correct": 150,
+            "n": 150,
+            "accuracy": 100.0
+          },
+          "persistence_accuracy": {
+            "correct": 48,
+            "n": 50,
+            "accuracy": 96.0
+          },
+          "completeness_addition_accuracy": {
+            "correct": 100,
+            "n": 100,
+            "accuracy": 100.0
+          },
+          "reopening_accuracy": {
+            "correct": 50,
+            "n": 50,
+            "accuracy": 100.0
+          },
+          "narrowing_persistence_accuracy": {
+            "correct": 48,
+            "n": 50,
+            "accuracy": 96.0
+          },
+          "changed_flag_accuracy": {
+            "correct": 198,
+            "n": 200,
+            "accuracy": 99.0
+          },
+          "applied_update_accuracy": {
+            "correct": 300,
+            "n": 300,
+            "accuracy": 100.0
+          },
+          "anchoring_error_rate": {
+            "errors": 0,
+            "n": 150,
+            "rate": 0.0
+          },
+          "closure_inertia_rate": {
+            "errors": 0,
+            "n": 50,
+            "rate": 0.0
+          },
+          "closure_retention_failure_rate": {
+            "errors": 2,
+            "n": 50,
+            "rate": 4.0
+          },
+          "by_turn": {
+            "1": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "2": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "3": {
+              "correct": 98,
+              "n": 100,
+              "accuracy": 98.0
+            }
+          },
+          "by_dialogue_type": {
+            "persist_after_narrowing": {
+              "correct": 148,
+              "n": 150,
+              "accuracy": 98.67
+            },
+            "reopen": {
+              "correct": 150,
+              "n": 150,
+              "accuracy": 100.0
+            }
+          },
+          "by_update_operation": {
+            "ADD_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "REPLACE_COMPLETE_PREDICATES": {
+              "correct": 98,
+              "n": 100,
+              "accuracy": 98.0
+            },
+            "SET_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            }
+          },
+          "error_types": {
+            "correct": 298,
+            "lost_relevant_local_closure": 2
+          },
+          "repeat": 3,
+          "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_deepseek-pro_repeat_3_scored.jsonl"
+        }
+      ],
+      "metrics": {
+        "dialogue_all_answer_accuracy": {
+          "mean": 98.0,
+          "sd": 1.0,
+          "ci95_half_width": 2.48,
+          "min": 97.0,
+          "max": 99.0,
+          "values": [
+            97.0,
+            99.0,
+            98.0
+          ]
+        },
+        "dialogue_all_turn_accuracy": {
+          "mean": 98.0,
+          "sd": 1.0,
+          "ci95_half_width": 2.48,
+          "min": 97.0,
+          "max": 99.0,
+          "values": [
+            97.0,
+            99.0,
+            98.0
+          ]
+        },
+        "answer_accuracy": {
+          "mean": 99.33,
+          "sd": 0.34,
+          "ci95_half_width": 0.83,
+          "min": 99.0,
+          "max": 99.67,
+          "values": [
+            99.0,
+            99.67,
+            99.33
+          ]
+        },
+        "revision_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "persistence_accuracy": {
+          "mean": 96.0,
+          "sd": 2.0,
+          "ci95_half_width": 4.97,
+          "min": 94.0,
+          "max": 98.0,
+          "values": [
+            94.0,
+            98.0,
+            96.0
+          ]
+        },
+        "completeness_addition_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "reopening_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "narrowing_persistence_accuracy": {
+          "mean": 96.0,
+          "sd": 2.0,
+          "ci95_half_width": 4.97,
+          "min": 94.0,
+          "max": 98.0,
+          "values": [
+            94.0,
+            98.0,
+            96.0
+          ]
+        },
+        "changed_flag_accuracy": {
+          "mean": 99.0,
+          "sd": 0.5,
+          "ci95_half_width": 1.24,
+          "min": 98.5,
+          "max": 99.5,
+          "values": [
+            98.5,
+            99.5,
+            99.0
+          ]
+        },
+        "applied_update_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "closure_inertia_rate": {
+          "mean": 0.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 0.0,
+          "max": 0.0,
+          "values": [
+            0.0,
+            0.0,
+            0.0
+          ]
+        },
+        "closure_retention_failure_rate": {
+          "mean": 4.0,
+          "sd": 2.0,
+          "ci95_half_width": 4.97,
+          "min": 2.0,
+          "max": 6.0,
+          "values": [
+            6.0,
+            2.0,
+            4.0
+          ]
+        },
+        "anchoring_error_rate": {
+          "mean": 0.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 0.0,
+          "max": 0.0,
+          "values": [
+            0.0,
+            0.0,
+            0.0
+          ]
+        }
+      },
+      "turn_decision_consistency": {
+        "n": 300,
+        "consistent": 295,
+        "rate": 98.33
+      },
+      "mean_parse_failures": 0.0,
+      "mean_api_errors": 0.0
+    },
+    "llama-maverick": {
+      "model": "meta-llama/llama-4-maverick",
+      "n_repeats": 3,
+      "replicates": [
+        {
+          "n_dialogues": 100,
+          "n_turns": 300,
+          "n_scored": 300,
+          "coverage": 100.0,
+          "models": [
+            "meta-llama/llama-4-maverick"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "json_valid": {
+            "n": 0,
+            "rate": 0.0
+          },
+          "answer_accuracy": {
+            "correct": 300,
+            "n": 300,
+            "accuracy": 100.0
+          },
+          "full_turn_accuracy": {
+            "correct": 300,
+            "n": 300,
+            "accuracy": 100.0
+          },
+          "dialogue_all_answer_accuracy": {
+            "correct": 100,
+            "n": 100,
+            "accuracy": 100.0
+          },
+          "dialogue_all_turn_accuracy": {
+            "correct": 100,
+            "n": 100,
+            "accuracy": 100.0
+          },
+          "revision_accuracy": {
+            "correct": 150,
+            "n": 150,
+            "accuracy": 100.0
+          },
+          "persistence_accuracy": {
+            "correct": 50,
+            "n": 50,
+            "accuracy": 100.0
+          },
+          "completeness_addition_accuracy": {
+            "correct": 100,
+            "n": 100,
+            "accuracy": 100.0
+          },
+          "reopening_accuracy": {
+            "correct": 50,
+            "n": 50,
+            "accuracy": 100.0
+          },
+          "narrowing_persistence_accuracy": {
+            "correct": 50,
+            "n": 50,
+            "accuracy": 100.0
+          },
+          "changed_flag_accuracy": {
+            "correct": 200,
+            "n": 200,
+            "accuracy": 100.0
+          },
+          "applied_update_accuracy": {
+            "correct": 300,
+            "n": 300,
+            "accuracy": 100.0
+          },
+          "anchoring_error_rate": {
+            "errors": 0,
+            "n": 150,
+            "rate": 0.0
+          },
+          "closure_inertia_rate": {
+            "errors": 0,
+            "n": 50,
+            "rate": 0.0
+          },
+          "closure_retention_failure_rate": {
+            "errors": 0,
+            "n": 50,
+            "rate": 0.0
+          },
+          "by_turn": {
+            "1": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "2": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "3": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            }
+          },
+          "by_dialogue_type": {
+            "persist_after_narrowing": {
+              "correct": 150,
+              "n": 150,
+              "accuracy": 100.0
+            },
+            "reopen": {
+              "correct": 150,
+              "n": 150,
+              "accuracy": 100.0
+            }
+          },
+          "by_update_operation": {
+            "ADD_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "REPLACE_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "SET_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            }
+          },
+          "error_types": {
+            "correct": 300
+          },
+          "repeat": 1,
+          "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_llama-maverick_repeat_1_scored.jsonl"
+        },
+        {
+          "n_dialogues": 100,
+          "n_turns": 300,
+          "n_scored": 300,
+          "coverage": 100.0,
+          "models": [
+            "meta-llama/llama-4-maverick"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "json_valid": {
+            "n": 0,
+            "rate": 0.0
+          },
+          "answer_accuracy": {
+            "correct": 300,
+            "n": 300,
+            "accuracy": 100.0
+          },
+          "full_turn_accuracy": {
+            "correct": 300,
+            "n": 300,
+            "accuracy": 100.0
+          },
+          "dialogue_all_answer_accuracy": {
+            "correct": 100,
+            "n": 100,
+            "accuracy": 100.0
+          },
+          "dialogue_all_turn_accuracy": {
+            "correct": 100,
+            "n": 100,
+            "accuracy": 100.0
+          },
+          "revision_accuracy": {
+            "correct": 150,
+            "n": 150,
+            "accuracy": 100.0
+          },
+          "persistence_accuracy": {
+            "correct": 50,
+            "n": 50,
+            "accuracy": 100.0
+          },
+          "completeness_addition_accuracy": {
+            "correct": 100,
+            "n": 100,
+            "accuracy": 100.0
+          },
+          "reopening_accuracy": {
+            "correct": 50,
+            "n": 50,
+            "accuracy": 100.0
+          },
+          "narrowing_persistence_accuracy": {
+            "correct": 50,
+            "n": 50,
+            "accuracy": 100.0
+          },
+          "changed_flag_accuracy": {
+            "correct": 200,
+            "n": 200,
+            "accuracy": 100.0
+          },
+          "applied_update_accuracy": {
+            "correct": 300,
+            "n": 300,
+            "accuracy": 100.0
+          },
+          "anchoring_error_rate": {
+            "errors": 0,
+            "n": 150,
+            "rate": 0.0
+          },
+          "closure_inertia_rate": {
+            "errors": 0,
+            "n": 50,
+            "rate": 0.0
+          },
+          "closure_retention_failure_rate": {
+            "errors": 0,
+            "n": 50,
+            "rate": 0.0
+          },
+          "by_turn": {
+            "1": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "2": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "3": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            }
+          },
+          "by_dialogue_type": {
+            "persist_after_narrowing": {
+              "correct": 150,
+              "n": 150,
+              "accuracy": 100.0
+            },
+            "reopen": {
+              "correct": 150,
+              "n": 150,
+              "accuracy": 100.0
+            }
+          },
+          "by_update_operation": {
+            "ADD_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "REPLACE_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "SET_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            }
+          },
+          "error_types": {
+            "correct": 300
+          },
+          "repeat": 2,
+          "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_llama-maverick_repeat_2_scored.jsonl"
+        },
+        {
+          "n_dialogues": 100,
+          "n_turns": 300,
+          "n_scored": 300,
+          "coverage": 100.0,
+          "models": [
+            "meta-llama/llama-4-maverick"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "json_valid": {
+            "n": 0,
+            "rate": 0.0
+          },
+          "answer_accuracy": {
+            "correct": 300,
+            "n": 300,
+            "accuracy": 100.0
+          },
+          "full_turn_accuracy": {
+            "correct": 297,
+            "n": 300,
+            "accuracy": 99.0
+          },
+          "dialogue_all_answer_accuracy": {
+            "correct": 100,
+            "n": 100,
+            "accuracy": 100.0
+          },
+          "dialogue_all_turn_accuracy": {
+            "correct": 97,
+            "n": 100,
+            "accuracy": 97.0
+          },
+          "revision_accuracy": {
+            "correct": 150,
+            "n": 150,
+            "accuracy": 100.0
+          },
+          "persistence_accuracy": {
+            "correct": 50,
+            "n": 50,
+            "accuracy": 100.0
+          },
+          "completeness_addition_accuracy": {
+            "correct": 100,
+            "n": 100,
+            "accuracy": 100.0
+          },
+          "reopening_accuracy": {
+            "correct": 50,
+            "n": 50,
+            "accuracy": 100.0
+          },
+          "narrowing_persistence_accuracy": {
+            "correct": 50,
+            "n": 50,
+            "accuracy": 100.0
+          },
+          "changed_flag_accuracy": {
+            "correct": 199,
+            "n": 200,
+            "accuracy": 99.5
+          },
+          "applied_update_accuracy": {
+            "correct": 297,
+            "n": 300,
+            "accuracy": 99.0
+          },
+          "anchoring_error_rate": {
+            "errors": 0,
+            "n": 150,
+            "rate": 0.0
+          },
+          "closure_inertia_rate": {
+            "errors": 0,
+            "n": 50,
+            "rate": 0.0
+          },
+          "closure_retention_failure_rate": {
+            "errors": 0,
+            "n": 50,
+            "rate": 0.0
+          },
+          "by_turn": {
+            "1": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "2": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "3": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            }
+          },
+          "by_dialogue_type": {
+            "persist_after_narrowing": {
+              "correct": 150,
+              "n": 150,
+              "accuracy": 100.0
+            },
+            "reopen": {
+              "correct": 150,
+              "n": 150,
+              "accuracy": 100.0
+            }
+          },
+          "by_update_operation": {
+            "ADD_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "REPLACE_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "SET_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            }
+          },
+          "error_types": {
+            "correct": 299,
+            "transition_flag_error": 1
+          },
+          "repeat": 3,
+          "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_llama-maverick_repeat_3_scored.jsonl"
+        }
+      ],
+      "metrics": {
+        "dialogue_all_answer_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "dialogue_all_turn_accuracy": {
+          "mean": 99.0,
+          "sd": 1.73,
+          "ci95_half_width": 4.3,
+          "min": 97.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            97.0
+          ]
+        },
+        "answer_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "revision_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "persistence_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "completeness_addition_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "reopening_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "narrowing_persistence_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "changed_flag_accuracy": {
+          "mean": 99.83,
+          "sd": 0.29,
+          "ci95_half_width": 0.72,
+          "min": 99.5,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            99.5
+          ]
+        },
+        "applied_update_accuracy": {
+          "mean": 99.67,
+          "sd": 0.58,
+          "ci95_half_width": 1.43,
+          "min": 99.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            99.0
+          ]
+        },
+        "closure_inertia_rate": {
+          "mean": 0.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 0.0,
+          "max": 0.0,
+          "values": [
+            0.0,
+            0.0,
+            0.0
+          ]
+        },
+        "closure_retention_failure_rate": {
+          "mean": 0.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 0.0,
+          "max": 0.0,
+          "values": [
+            0.0,
+            0.0,
+            0.0
+          ]
+        },
+        "anchoring_error_rate": {
+          "mean": 0.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 0.0,
+          "max": 0.0,
+          "values": [
+            0.0,
+            0.0,
+            0.0
+          ]
+        }
+      },
+      "turn_decision_consistency": {
+        "n": 300,
+        "consistent": 297,
+        "rate": 99.0
+      },
+      "mean_parse_failures": 0.0,
+      "mean_api_errors": 0.0
+    },
+    "llama-scout": {
+      "model": "meta-llama/llama-4-scout",
+      "n_repeats": 3,
+      "replicates": [
+        {
+          "n_dialogues": 100,
+          "n_turns": 300,
+          "n_scored": 300,
+          "coverage": 100.0,
+          "models": [
+            "meta-llama/llama-4-scout"
+          ],
+          "api_errors": 0,
+          "parse_failures": 1,
+          "json_valid": {
+            "n": 123,
+            "rate": 41.0
+          },
+          "answer_accuracy": {
+            "correct": 249,
+            "n": 300,
+            "accuracy": 83.0
+          },
+          "full_turn_accuracy": {
+            "correct": 244,
+            "n": 300,
+            "accuracy": 81.33
+          },
+          "dialogue_all_answer_accuracy": {
+            "correct": 50,
+            "n": 100,
+            "accuracy": 50.0
+          },
+          "dialogue_all_turn_accuracy": {
+            "correct": 50,
+            "n": 100,
+            "accuracy": 50.0
+          },
+          "revision_accuracy": {
+            "correct": 148,
+            "n": 150,
+            "accuracy": 98.67
+          },
+          "persistence_accuracy": {
+            "correct": 2,
+            "n": 50,
+            "accuracy": 4.0
+          },
+          "completeness_addition_accuracy": {
+            "correct": 98,
+            "n": 100,
+            "accuracy": 98.0
+          },
+          "reopening_accuracy": {
+            "correct": 50,
+            "n": 50,
+            "accuracy": 100.0
+          },
+          "narrowing_persistence_accuracy": {
+            "correct": 2,
+            "n": 50,
+            "accuracy": 4.0
+          },
+          "changed_flag_accuracy": {
+            "correct": 150,
+            "n": 200,
+            "accuracy": 75.0
+          },
+          "applied_update_accuracy": {
+            "correct": 297,
+            "n": 300,
+            "accuracy": 99.0
+          },
+          "anchoring_error_rate": {
+            "errors": 0,
+            "n": 150,
+            "rate": 0.0
+          },
+          "closure_inertia_rate": {
+            "errors": 0,
+            "n": 50,
+            "rate": 0.0
+          },
+          "closure_retention_failure_rate": {
+            "errors": 47,
+            "n": 50,
+            "rate": 94.0
+          },
+          "by_turn": {
+            "1": {
+              "correct": 99,
+              "n": 100,
+              "accuracy": 99.0
+            },
+            "2": {
+              "correct": 98,
+              "n": 100,
+              "accuracy": 98.0
+            },
+            "3": {
+              "correct": 52,
+              "n": 100,
+              "accuracy": 52.0
+            }
+          },
+          "by_dialogue_type": {
+            "persist_after_narrowing": {
+              "correct": 99,
+              "n": 150,
+              "accuracy": 66.0
+            },
+            "reopen": {
+              "correct": 150,
+              "n": 150,
+              "accuracy": 100.0
+            }
+          },
+          "by_update_operation": {
+            "ADD_COMPLETE_PREDICATES": {
+              "correct": 98,
+              "n": 100,
+              "accuracy": 98.0
+            },
+            "REPLACE_COMPLETE_PREDICATES": {
+              "correct": 52,
+              "n": 100,
+              "accuracy": 52.0
+            },
+            "SET_COMPLETE_PREDICATES": {
+              "correct": 99,
+              "n": 100,
+              "accuracy": 99.0
+            }
+          },
+          "error_types": {
+            "correct": 247,
+            "lost_relevant_local_closure": 47,
+            "other_answer_error": 2,
+            "transition_flag_error": 2,
+            "parse_failure": 1,
+            "missed_completeness_addition": 1
+          },
+          "repeat": 1,
+          "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_llama-scout_repeat_1_scored.jsonl"
+        },
+        {
+          "n_dialogues": 100,
+          "n_turns": 300,
+          "n_scored": 300,
+          "coverage": 100.0,
+          "models": [
+            "meta-llama/llama-4-scout"
+          ],
+          "api_errors": 0,
+          "parse_failures": 1,
+          "json_valid": {
+            "n": 108,
+            "rate": 36.0
+          },
+          "answer_accuracy": {
+            "correct": 250,
+            "n": 300,
+            "accuracy": 83.33
+          },
+          "full_turn_accuracy": {
+            "correct": 248,
+            "n": 300,
+            "accuracy": 82.67
+          },
+          "dialogue_all_answer_accuracy": {
+            "correct": 50,
+            "n": 100,
+            "accuracy": 50.0
+          },
+          "dialogue_all_turn_accuracy": {
+            "correct": 49,
+            "n": 100,
+            "accuracy": 49.0
+          },
+          "revision_accuracy": {
+            "correct": 149,
+            "n": 150,
+            "accuracy": 99.33
+          },
+          "persistence_accuracy": {
+            "correct": 1,
+            "n": 50,
+            "accuracy": 2.0
+          },
+          "completeness_addition_accuracy": {
+            "correct": 100,
+            "n": 100,
+            "accuracy": 100.0
+          },
+          "reopening_accuracy": {
+            "correct": 49,
+            "n": 50,
+            "accuracy": 98.0
+          },
+          "narrowing_persistence_accuracy": {
+            "correct": 1,
+            "n": 50,
+            "accuracy": 2.0
+          },
+          "changed_flag_accuracy": {
+            "correct": 150,
+            "n": 200,
+            "accuracy": 75.0
+          },
+          "applied_update_accuracy": {
+            "correct": 296,
+            "n": 300,
+            "accuracy": 98.67
+          },
+          "anchoring_error_rate": {
+            "errors": 0,
+            "n": 150,
+            "rate": 0.0
+          },
+          "closure_inertia_rate": {
+            "errors": 0,
+            "n": 50,
+            "rate": 0.0
+          },
+          "closure_retention_failure_rate": {
+            "errors": 49,
+            "n": 50,
+            "rate": 98.0
+          },
+          "by_turn": {
+            "1": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "2": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "3": {
+              "correct": 50,
+              "n": 100,
+              "accuracy": 50.0
+            }
+          },
+          "by_dialogue_type": {
+            "persist_after_narrowing": {
+              "correct": 101,
+              "n": 150,
+              "accuracy": 67.33
+            },
+            "reopen": {
+              "correct": 149,
+              "n": 150,
+              "accuracy": 99.33
+            }
+          },
+          "by_update_operation": {
+            "ADD_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "REPLACE_COMPLETE_PREDICATES": {
+              "correct": 50,
+              "n": 100,
+              "accuracy": 50.0
+            },
+            "SET_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            }
+          },
+          "error_types": {
+            "correct": 250,
+            "lost_relevant_local_closure": 49,
+            "parse_failure": 1
+          },
+          "repeat": 2,
+          "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_llama-scout_repeat_2_scored.jsonl"
+        },
+        {
+          "n_dialogues": 100,
+          "n_turns": 300,
+          "n_scored": 300,
+          "coverage": 100.0,
+          "models": [
+            "meta-llama/llama-4-scout"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "json_valid": {
+            "n": 130,
+            "rate": 43.33
+          },
+          "answer_accuracy": {
+            "correct": 250,
+            "n": 300,
+            "accuracy": 83.33
+          },
+          "full_turn_accuracy": {
+            "correct": 245,
+            "n": 300,
+            "accuracy": 81.67
+          },
+          "dialogue_all_answer_accuracy": {
+            "correct": 50,
+            "n": 100,
+            "accuracy": 50.0
+          },
+          "dialogue_all_turn_accuracy": {
+            "correct": 47,
+            "n": 100,
+            "accuracy": 47.0
+          },
+          "revision_accuracy": {
+            "correct": 150,
+            "n": 150,
+            "accuracy": 100.0
+          },
+          "persistence_accuracy": {
+            "correct": 0,
+            "n": 50,
+            "accuracy": 0.0
+          },
+          "completeness_addition_accuracy": {
+            "correct": 100,
+            "n": 100,
+            "accuracy": 100.0
+          },
+          "reopening_accuracy": {
+            "correct": 50,
+            "n": 50,
+            "accuracy": 100.0
+          },
+          "narrowing_persistence_accuracy": {
+            "correct": 0,
+            "n": 50,
+            "accuracy": 0.0
+          },
+          "changed_flag_accuracy": {
+            "correct": 148,
+            "n": 200,
+            "accuracy": 74.0
+          },
+          "applied_update_accuracy": {
+            "correct": 296,
+            "n": 300,
+            "accuracy": 98.67
+          },
+          "anchoring_error_rate": {
+            "errors": 0,
+            "n": 150,
+            "rate": 0.0
+          },
+          "closure_inertia_rate": {
+            "errors": 0,
+            "n": 50,
+            "rate": 0.0
+          },
+          "closure_retention_failure_rate": {
+            "errors": 50,
+            "n": 50,
+            "rate": 100.0
+          },
+          "by_turn": {
+            "1": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "2": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "3": {
+              "correct": 50,
+              "n": 100,
+              "accuracy": 50.0
+            }
+          },
+          "by_dialogue_type": {
+            "persist_after_narrowing": {
+              "correct": 100,
+              "n": 150,
+              "accuracy": 66.67
+            },
+            "reopen": {
+              "correct": 150,
+              "n": 150,
+              "accuracy": 100.0
+            }
+          },
+          "by_update_operation": {
+            "ADD_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "REPLACE_COMPLETE_PREDICATES": {
+              "correct": 50,
+              "n": 100,
+              "accuracy": 50.0
+            },
+            "SET_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            }
+          },
+          "error_types": {
+            "correct": 248,
+            "lost_relevant_local_closure": 50,
+            "transition_flag_error": 2
+          },
+          "repeat": 3,
+          "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_llama-scout_repeat_3_scored.jsonl"
+        }
+      ],
+      "metrics": {
+        "dialogue_all_answer_accuracy": {
+          "mean": 50.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 50.0,
+          "max": 50.0,
+          "values": [
+            50.0,
+            50.0,
+            50.0
+          ]
+        },
+        "dialogue_all_turn_accuracy": {
+          "mean": 48.67,
+          "sd": 1.53,
+          "ci95_half_width": 3.79,
+          "min": 47.0,
+          "max": 50.0,
+          "values": [
+            50.0,
+            49.0,
+            47.0
+          ]
+        },
+        "answer_accuracy": {
+          "mean": 83.22,
+          "sd": 0.19,
+          "ci95_half_width": 0.47,
+          "min": 83.0,
+          "max": 83.33,
+          "values": [
+            83.0,
+            83.33,
+            83.33
+          ]
+        },
+        "revision_accuracy": {
+          "mean": 99.33,
+          "sd": 0.67,
+          "ci95_half_width": 1.65,
+          "min": 98.67,
+          "max": 100.0,
+          "values": [
+            98.67,
+            99.33,
+            100.0
+          ]
+        },
+        "persistence_accuracy": {
+          "mean": 2.0,
+          "sd": 2.0,
+          "ci95_half_width": 4.97,
+          "min": 0.0,
+          "max": 4.0,
+          "values": [
+            4.0,
+            2.0,
+            0.0
+          ]
+        },
+        "completeness_addition_accuracy": {
+          "mean": 99.33,
+          "sd": 1.15,
+          "ci95_half_width": 2.87,
+          "min": 98.0,
+          "max": 100.0,
+          "values": [
+            98.0,
+            100.0,
+            100.0
+          ]
+        },
+        "reopening_accuracy": {
+          "mean": 99.33,
+          "sd": 1.15,
+          "ci95_half_width": 2.87,
+          "min": 98.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            98.0,
+            100.0
+          ]
+        },
+        "narrowing_persistence_accuracy": {
+          "mean": 2.0,
+          "sd": 2.0,
+          "ci95_half_width": 4.97,
+          "min": 0.0,
+          "max": 4.0,
+          "values": [
+            4.0,
+            2.0,
+            0.0
+          ]
+        },
+        "changed_flag_accuracy": {
+          "mean": 74.67,
+          "sd": 0.58,
+          "ci95_half_width": 1.43,
+          "min": 74.0,
+          "max": 75.0,
+          "values": [
+            75.0,
+            75.0,
+            74.0
+          ]
+        },
+        "applied_update_accuracy": {
+          "mean": 98.78,
+          "sd": 0.19,
+          "ci95_half_width": 0.47,
+          "min": 98.67,
+          "max": 99.0,
+          "values": [
+            99.0,
+            98.67,
+            98.67
+          ]
+        },
+        "closure_inertia_rate": {
+          "mean": 0.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 0.0,
+          "max": 0.0,
+          "values": [
+            0.0,
+            0.0,
+            0.0
+          ]
+        },
+        "closure_retention_failure_rate": {
+          "mean": 97.33,
+          "sd": 3.06,
+          "ci95_half_width": 7.59,
+          "min": 94.0,
+          "max": 100.0,
+          "values": [
+            94.0,
+            98.0,
+            100.0
+          ]
+        },
+        "anchoring_error_rate": {
+          "mean": 0.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 0.0,
+          "max": 0.0,
+          "values": [
+            0.0,
+            0.0,
+            0.0
+          ]
+        }
+      },
+      "turn_decision_consistency": {
+        "n": 300,
+        "consistent": 281,
+        "rate": 93.67
+      },
+      "mean_parse_failures": 0.67,
+      "mean_api_errors": 0.0
+    },
+    "mistral-medium": {
+      "model": "mistralai/mistral-medium-3-5",
+      "n_repeats": 3,
+      "replicates": [
+        {
+          "n_dialogues": 100,
+          "n_turns": 300,
+          "n_scored": 300,
+          "coverage": 100.0,
+          "models": [
+            "mistralai/mistral-medium-3-5"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "json_valid": {
+            "n": 300,
+            "rate": 100.0
+          },
+          "answer_accuracy": {
+            "correct": 284,
+            "n": 300,
+            "accuracy": 94.67
+          },
+          "full_turn_accuracy": {
+            "correct": 284,
+            "n": 300,
+            "accuracy": 94.67
+          },
+          "dialogue_all_answer_accuracy": {
+            "correct": 84,
+            "n": 100,
+            "accuracy": 84.0
+          },
+          "dialogue_all_turn_accuracy": {
+            "correct": 84,
+            "n": 100,
+            "accuracy": 84.0
+          },
+          "revision_accuracy": {
+            "correct": 150,
+            "n": 150,
+            "accuracy": 100.0
+          },
+          "persistence_accuracy": {
+            "correct": 34,
+            "n": 50,
+            "accuracy": 68.0
+          },
+          "completeness_addition_accuracy": {
+            "correct": 100,
+            "n": 100,
+            "accuracy": 100.0
+          },
+          "reopening_accuracy": {
+            "correct": 50,
+            "n": 50,
+            "accuracy": 100.0
+          },
+          "narrowing_persistence_accuracy": {
+            "correct": 34,
+            "n": 50,
+            "accuracy": 68.0
+          },
+          "changed_flag_accuracy": {
+            "correct": 184,
+            "n": 200,
+            "accuracy": 92.0
+          },
+          "applied_update_accuracy": {
+            "correct": 300,
+            "n": 300,
+            "accuracy": 100.0
+          },
+          "anchoring_error_rate": {
+            "errors": 0,
+            "n": 150,
+            "rate": 0.0
+          },
+          "closure_inertia_rate": {
+            "errors": 0,
+            "n": 50,
+            "rate": 0.0
+          },
+          "closure_retention_failure_rate": {
+            "errors": 16,
+            "n": 50,
+            "rate": 32.0
+          },
+          "by_turn": {
+            "1": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "2": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "3": {
+              "correct": 84,
+              "n": 100,
+              "accuracy": 84.0
+            }
+          },
+          "by_dialogue_type": {
+            "persist_after_narrowing": {
+              "correct": 134,
+              "n": 150,
+              "accuracy": 89.33
+            },
+            "reopen": {
+              "correct": 150,
+              "n": 150,
+              "accuracy": 100.0
+            }
+          },
+          "by_update_operation": {
+            "ADD_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "REPLACE_COMPLETE_PREDICATES": {
+              "correct": 84,
+              "n": 100,
+              "accuracy": 84.0
+            },
+            "SET_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            }
+          },
+          "error_types": {
+            "correct": 284,
+            "lost_relevant_local_closure": 16
+          },
+          "repeat": 1,
+          "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_mistral-medium_repeat_1_scored.jsonl"
+        },
+        {
+          "n_dialogues": 100,
+          "n_turns": 300,
+          "n_scored": 300,
+          "coverage": 100.0,
+          "models": [
+            "mistralai/mistral-medium-3-5"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "json_valid": {
+            "n": 300,
+            "rate": 100.0
+          },
+          "answer_accuracy": {
+            "correct": 283,
+            "n": 300,
+            "accuracy": 94.33
+          },
+          "full_turn_accuracy": {
+            "correct": 283,
+            "n": 300,
+            "accuracy": 94.33
+          },
+          "dialogue_all_answer_accuracy": {
+            "correct": 83,
+            "n": 100,
+            "accuracy": 83.0
+          },
+          "dialogue_all_turn_accuracy": {
+            "correct": 83,
+            "n": 100,
+            "accuracy": 83.0
+          },
+          "revision_accuracy": {
+            "correct": 150,
+            "n": 150,
+            "accuracy": 100.0
+          },
+          "persistence_accuracy": {
+            "correct": 33,
+            "n": 50,
+            "accuracy": 66.0
+          },
+          "completeness_addition_accuracy": {
+            "correct": 100,
+            "n": 100,
+            "accuracy": 100.0
+          },
+          "reopening_accuracy": {
+            "correct": 50,
+            "n": 50,
+            "accuracy": 100.0
+          },
+          "narrowing_persistence_accuracy": {
+            "correct": 33,
+            "n": 50,
+            "accuracy": 66.0
+          },
+          "changed_flag_accuracy": {
+            "correct": 183,
+            "n": 200,
+            "accuracy": 91.5
+          },
+          "applied_update_accuracy": {
+            "correct": 300,
+            "n": 300,
+            "accuracy": 100.0
+          },
+          "anchoring_error_rate": {
+            "errors": 0,
+            "n": 150,
+            "rate": 0.0
+          },
+          "closure_inertia_rate": {
+            "errors": 0,
+            "n": 50,
+            "rate": 0.0
+          },
+          "closure_retention_failure_rate": {
+            "errors": 17,
+            "n": 50,
+            "rate": 34.0
+          },
+          "by_turn": {
+            "1": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "2": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "3": {
+              "correct": 83,
+              "n": 100,
+              "accuracy": 83.0
+            }
+          },
+          "by_dialogue_type": {
+            "persist_after_narrowing": {
+              "correct": 133,
+              "n": 150,
+              "accuracy": 88.67
+            },
+            "reopen": {
+              "correct": 150,
+              "n": 150,
+              "accuracy": 100.0
+            }
+          },
+          "by_update_operation": {
+            "ADD_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "REPLACE_COMPLETE_PREDICATES": {
+              "correct": 83,
+              "n": 100,
+              "accuracy": 83.0
+            },
+            "SET_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            }
+          },
+          "error_types": {
+            "correct": 283,
+            "lost_relevant_local_closure": 17
+          },
+          "repeat": 2,
+          "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_mistral-medium_repeat_2_scored.jsonl"
+        },
+        {
+          "n_dialogues": 100,
+          "n_turns": 300,
+          "n_scored": 300,
+          "coverage": 100.0,
+          "models": [
+            "mistralai/mistral-medium-3-5"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "json_valid": {
+            "n": 300,
+            "rate": 100.0
+          },
+          "answer_accuracy": {
+            "correct": 283,
+            "n": 300,
+            "accuracy": 94.33
+          },
+          "full_turn_accuracy": {
+            "correct": 283,
+            "n": 300,
+            "accuracy": 94.33
+          },
+          "dialogue_all_answer_accuracy": {
+            "correct": 83,
+            "n": 100,
+            "accuracy": 83.0
+          },
+          "dialogue_all_turn_accuracy": {
+            "correct": 83,
+            "n": 100,
+            "accuracy": 83.0
+          },
+          "revision_accuracy": {
+            "correct": 150,
+            "n": 150,
+            "accuracy": 100.0
+          },
+          "persistence_accuracy": {
+            "correct": 33,
+            "n": 50,
+            "accuracy": 66.0
+          },
+          "completeness_addition_accuracy": {
+            "correct": 100,
+            "n": 100,
+            "accuracy": 100.0
+          },
+          "reopening_accuracy": {
+            "correct": 50,
+            "n": 50,
+            "accuracy": 100.0
+          },
+          "narrowing_persistence_accuracy": {
+            "correct": 33,
+            "n": 50,
+            "accuracy": 66.0
+          },
+          "changed_flag_accuracy": {
+            "correct": 183,
+            "n": 200,
+            "accuracy": 91.5
+          },
+          "applied_update_accuracy": {
+            "correct": 300,
+            "n": 300,
+            "accuracy": 100.0
+          },
+          "anchoring_error_rate": {
+            "errors": 0,
+            "n": 150,
+            "rate": 0.0
+          },
+          "closure_inertia_rate": {
+            "errors": 0,
+            "n": 50,
+            "rate": 0.0
+          },
+          "closure_retention_failure_rate": {
+            "errors": 17,
+            "n": 50,
+            "rate": 34.0
+          },
+          "by_turn": {
+            "1": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "2": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "3": {
+              "correct": 83,
+              "n": 100,
+              "accuracy": 83.0
+            }
+          },
+          "by_dialogue_type": {
+            "persist_after_narrowing": {
+              "correct": 133,
+              "n": 150,
+              "accuracy": 88.67
+            },
+            "reopen": {
+              "correct": 150,
+              "n": 150,
+              "accuracy": 100.0
+            }
+          },
+          "by_update_operation": {
+            "ADD_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "REPLACE_COMPLETE_PREDICATES": {
+              "correct": 83,
+              "n": 100,
+              "accuracy": 83.0
+            },
+            "SET_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            }
+          },
+          "error_types": {
+            "correct": 283,
+            "lost_relevant_local_closure": 17
+          },
+          "repeat": 3,
+          "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_mistral-medium_repeat_3_scored.jsonl"
+        }
+      ],
+      "metrics": {
+        "dialogue_all_answer_accuracy": {
+          "mean": 83.33,
+          "sd": 0.58,
+          "ci95_half_width": 1.43,
+          "min": 83.0,
+          "max": 84.0,
+          "values": [
+            84.0,
+            83.0,
+            83.0
+          ]
+        },
+        "dialogue_all_turn_accuracy": {
+          "mean": 83.33,
+          "sd": 0.58,
+          "ci95_half_width": 1.43,
+          "min": 83.0,
+          "max": 84.0,
+          "values": [
+            84.0,
+            83.0,
+            83.0
+          ]
+        },
+        "answer_accuracy": {
+          "mean": 94.44,
+          "sd": 0.2,
+          "ci95_half_width": 0.49,
+          "min": 94.33,
+          "max": 94.67,
+          "values": [
+            94.67,
+            94.33,
+            94.33
+          ]
+        },
+        "revision_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "persistence_accuracy": {
+          "mean": 66.67,
+          "sd": 1.15,
+          "ci95_half_width": 2.87,
+          "min": 66.0,
+          "max": 68.0,
+          "values": [
+            68.0,
+            66.0,
+            66.0
+          ]
+        },
+        "completeness_addition_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "reopening_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "narrowing_persistence_accuracy": {
+          "mean": 66.67,
+          "sd": 1.15,
+          "ci95_half_width": 2.87,
+          "min": 66.0,
+          "max": 68.0,
+          "values": [
+            68.0,
+            66.0,
+            66.0
+          ]
+        },
+        "changed_flag_accuracy": {
+          "mean": 91.67,
+          "sd": 0.29,
+          "ci95_half_width": 0.72,
+          "min": 91.5,
+          "max": 92.0,
+          "values": [
+            92.0,
+            91.5,
+            91.5
+          ]
+        },
+        "applied_update_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "closure_inertia_rate": {
+          "mean": 0.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 0.0,
+          "max": 0.0,
+          "values": [
+            0.0,
+            0.0,
+            0.0
+          ]
+        },
+        "closure_retention_failure_rate": {
+          "mean": 33.33,
+          "sd": 1.15,
+          "ci95_half_width": 2.87,
+          "min": 32.0,
+          "max": 34.0,
+          "values": [
+            32.0,
+            34.0,
+            34.0
+          ]
+        },
+        "anchoring_error_rate": {
+          "mean": 0.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 0.0,
+          "max": 0.0,
+          "values": [
+            0.0,
+            0.0,
+            0.0
+          ]
+        }
+      },
+      "turn_decision_consistency": {
+        "n": 300,
+        "consistent": 299,
+        "rate": 99.67
+      },
+      "mean_parse_failures": 0.0,
+      "mean_api_errors": 0.0
+    },
+    "mistral-small": {
+      "model": "mistralai/mistral-small-2603",
+      "n_repeats": 3,
+      "replicates": [
+        {
+          "n_dialogues": 100,
+          "n_turns": 300,
+          "n_scored": 300,
+          "coverage": 100.0,
+          "models": [
+            "mistralai/mistral-small-2603"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "json_valid": {
+            "n": 300,
+            "rate": 100.0
+          },
+          "answer_accuracy": {
+            "correct": 259,
+            "n": 300,
+            "accuracy": 86.33
+          },
+          "full_turn_accuracy": {
+            "correct": 259,
+            "n": 300,
+            "accuracy": 86.33
+          },
+          "dialogue_all_answer_accuracy": {
+            "correct": 59,
+            "n": 100,
+            "accuracy": 59.0
+          },
+          "dialogue_all_turn_accuracy": {
+            "correct": 59,
+            "n": 100,
+            "accuracy": 59.0
+          },
+          "revision_accuracy": {
+            "correct": 150,
+            "n": 150,
+            "accuracy": 100.0
+          },
+          "persistence_accuracy": {
+            "correct": 10,
+            "n": 50,
+            "accuracy": 20.0
+          },
+          "completeness_addition_accuracy": {
+            "correct": 100,
+            "n": 100,
+            "accuracy": 100.0
+          },
+          "reopening_accuracy": {
+            "correct": 50,
+            "n": 50,
+            "accuracy": 100.0
+          },
+          "narrowing_persistence_accuracy": {
+            "correct": 10,
+            "n": 50,
+            "accuracy": 20.0
+          },
+          "changed_flag_accuracy": {
+            "correct": 160,
+            "n": 200,
+            "accuracy": 80.0
+          },
+          "applied_update_accuracy": {
+            "correct": 300,
+            "n": 300,
+            "accuracy": 100.0
+          },
+          "anchoring_error_rate": {
+            "errors": 1,
+            "n": 150,
+            "rate": 0.67
+          },
+          "closure_inertia_rate": {
+            "errors": 0,
+            "n": 50,
+            "rate": 0.0
+          },
+          "closure_retention_failure_rate": {
+            "errors": 40,
+            "n": 50,
+            "rate": 80.0
+          },
+          "by_turn": {
+            "1": {
+              "correct": 99,
+              "n": 100,
+              "accuracy": 99.0
+            },
+            "2": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "3": {
+              "correct": 60,
+              "n": 100,
+              "accuracy": 60.0
+            }
+          },
+          "by_dialogue_type": {
+            "persist_after_narrowing": {
+              "correct": 110,
+              "n": 150,
+              "accuracy": 73.33
+            },
+            "reopen": {
+              "correct": 149,
+              "n": 150,
+              "accuracy": 99.33
+            }
+          },
+          "by_update_operation": {
+            "ADD_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "REPLACE_COMPLETE_PREDICATES": {
+              "correct": 60,
+              "n": 100,
+              "accuracy": 60.0
+            },
+            "SET_COMPLETE_PREDICATES": {
+              "correct": 99,
+              "n": 100,
+              "accuracy": 99.0
+            }
+          },
+          "error_types": {
+            "correct": 259,
+            "lost_relevant_local_closure": 40,
+            "other_answer_error": 1
+          },
+          "repeat": 1,
+          "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_mistral-small_repeat_1_scored.jsonl"
+        },
+        {
+          "n_dialogues": 100,
+          "n_turns": 300,
+          "n_scored": 300,
+          "coverage": 100.0,
+          "models": [
+            "mistralai/mistral-small-2603"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "json_valid": {
+            "n": 300,
+            "rate": 100.0
+          },
+          "answer_accuracy": {
+            "correct": 255,
+            "n": 300,
+            "accuracy": 85.0
+          },
+          "full_turn_accuracy": {
+            "correct": 255,
+            "n": 300,
+            "accuracy": 85.0
+          },
+          "dialogue_all_answer_accuracy": {
+            "correct": 55,
+            "n": 100,
+            "accuracy": 55.0
+          },
+          "dialogue_all_turn_accuracy": {
+            "correct": 55,
+            "n": 100,
+            "accuracy": 55.0
+          },
+          "revision_accuracy": {
+            "correct": 150,
+            "n": 150,
+            "accuracy": 100.0
+          },
+          "persistence_accuracy": {
+            "correct": 9,
+            "n": 50,
+            "accuracy": 18.0
+          },
+          "completeness_addition_accuracy": {
+            "correct": 100,
+            "n": 100,
+            "accuracy": 100.0
+          },
+          "reopening_accuracy": {
+            "correct": 50,
+            "n": 50,
+            "accuracy": 100.0
+          },
+          "narrowing_persistence_accuracy": {
+            "correct": 9,
+            "n": 50,
+            "accuracy": 18.0
+          },
+          "changed_flag_accuracy": {
+            "correct": 159,
+            "n": 200,
+            "accuracy": 79.5
+          },
+          "applied_update_accuracy": {
+            "correct": 300,
+            "n": 300,
+            "accuracy": 100.0
+          },
+          "anchoring_error_rate": {
+            "errors": 4,
+            "n": 150,
+            "rate": 2.67
+          },
+          "closure_inertia_rate": {
+            "errors": 0,
+            "n": 50,
+            "rate": 0.0
+          },
+          "closure_retention_failure_rate": {
+            "errors": 41,
+            "n": 50,
+            "rate": 82.0
+          },
+          "by_turn": {
+            "1": {
+              "correct": 96,
+              "n": 100,
+              "accuracy": 96.0
+            },
+            "2": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "3": {
+              "correct": 59,
+              "n": 100,
+              "accuracy": 59.0
+            }
+          },
+          "by_dialogue_type": {
+            "persist_after_narrowing": {
+              "correct": 109,
+              "n": 150,
+              "accuracy": 72.67
+            },
+            "reopen": {
+              "correct": 146,
+              "n": 150,
+              "accuracy": 97.33
+            }
+          },
+          "by_update_operation": {
+            "ADD_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "REPLACE_COMPLETE_PREDICATES": {
+              "correct": 59,
+              "n": 100,
+              "accuracy": 59.0
+            },
+            "SET_COMPLETE_PREDICATES": {
+              "correct": 96,
+              "n": 100,
+              "accuracy": 96.0
+            }
+          },
+          "error_types": {
+            "correct": 255,
+            "lost_relevant_local_closure": 41,
+            "other_answer_error": 4
+          },
+          "repeat": 2,
+          "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_mistral-small_repeat_2_scored.jsonl"
+        },
+        {
+          "n_dialogues": 100,
+          "n_turns": 300,
+          "n_scored": 300,
+          "coverage": 100.0,
+          "models": [
+            "mistralai/mistral-small-2603"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "json_valid": {
+            "n": 300,
+            "rate": 100.0
+          },
+          "answer_accuracy": {
+            "correct": 260,
+            "n": 300,
+            "accuracy": 86.67
+          },
+          "full_turn_accuracy": {
+            "correct": 259,
+            "n": 300,
+            "accuracy": 86.33
+          },
+          "dialogue_all_answer_accuracy": {
+            "correct": 60,
+            "n": 100,
+            "accuracy": 60.0
+          },
+          "dialogue_all_turn_accuracy": {
+            "correct": 60,
+            "n": 100,
+            "accuracy": 60.0
+          },
+          "revision_accuracy": {
+            "correct": 150,
+            "n": 150,
+            "accuracy": 100.0
+          },
+          "persistence_accuracy": {
+            "correct": 13,
+            "n": 50,
+            "accuracy": 26.0
+          },
+          "completeness_addition_accuracy": {
+            "correct": 100,
+            "n": 100,
+            "accuracy": 100.0
+          },
+          "reopening_accuracy": {
+            "correct": 50,
+            "n": 50,
+            "accuracy": 100.0
+          },
+          "narrowing_persistence_accuracy": {
+            "correct": 13,
+            "n": 50,
+            "accuracy": 26.0
+          },
+          "changed_flag_accuracy": {
+            "correct": 162,
+            "n": 200,
+            "accuracy": 81.0
+          },
+          "applied_update_accuracy": {
+            "correct": 300,
+            "n": 300,
+            "accuracy": 100.0
+          },
+          "anchoring_error_rate": {
+            "errors": 3,
+            "n": 150,
+            "rate": 2.0
+          },
+          "closure_inertia_rate": {
+            "errors": 0,
+            "n": 50,
+            "rate": 0.0
+          },
+          "closure_retention_failure_rate": {
+            "errors": 37,
+            "n": 50,
+            "rate": 74.0
+          },
+          "by_turn": {
+            "1": {
+              "correct": 97,
+              "n": 100,
+              "accuracy": 97.0
+            },
+            "2": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "3": {
+              "correct": 63,
+              "n": 100,
+              "accuracy": 63.0
+            }
+          },
+          "by_dialogue_type": {
+            "persist_after_narrowing": {
+              "correct": 113,
+              "n": 150,
+              "accuracy": 75.33
+            },
+            "reopen": {
+              "correct": 147,
+              "n": 150,
+              "accuracy": 98.0
+            }
+          },
+          "by_update_operation": {
+            "ADD_COMPLETE_PREDICATES": {
+              "correct": 100,
+              "n": 100,
+              "accuracy": 100.0
+            },
+            "REPLACE_COMPLETE_PREDICATES": {
+              "correct": 63,
+              "n": 100,
+              "accuracy": 63.0
+            },
+            "SET_COMPLETE_PREDICATES": {
+              "correct": 97,
+              "n": 100,
+              "accuracy": 97.0
+            }
+          },
+          "error_types": {
+            "correct": 259,
+            "lost_relevant_local_closure": 37,
+            "other_answer_error": 3,
+            "transition_flag_error": 1
+          },
+          "repeat": 3,
+          "scored": "results/scored/dynamic_dialogue_replicates/closure_contract_v3_dynamic_dialogue_mistral-small_repeat_3_scored.jsonl"
+        }
+      ],
+      "metrics": {
+        "dialogue_all_answer_accuracy": {
+          "mean": 58.0,
+          "sd": 2.65,
+          "ci95_half_width": 6.57,
+          "min": 55.0,
+          "max": 60.0,
+          "values": [
+            59.0,
+            55.0,
+            60.0
+          ]
+        },
+        "dialogue_all_turn_accuracy": {
+          "mean": 58.0,
+          "sd": 2.65,
+          "ci95_half_width": 6.57,
+          "min": 55.0,
+          "max": 60.0,
+          "values": [
+            59.0,
+            55.0,
+            60.0
+          ]
+        },
+        "answer_accuracy": {
+          "mean": 86.0,
+          "sd": 0.88,
+          "ci95_half_width": 2.19,
+          "min": 85.0,
+          "max": 86.67,
+          "values": [
+            86.33,
+            85.0,
+            86.67
+          ]
+        },
+        "revision_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "persistence_accuracy": {
+          "mean": 21.33,
+          "sd": 4.16,
+          "ci95_half_width": 10.34,
+          "min": 18.0,
+          "max": 26.0,
+          "values": [
+            20.0,
+            18.0,
+            26.0
+          ]
+        },
+        "completeness_addition_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "reopening_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "narrowing_persistence_accuracy": {
+          "mean": 21.33,
+          "sd": 4.16,
+          "ci95_half_width": 10.34,
+          "min": 18.0,
+          "max": 26.0,
+          "values": [
+            20.0,
+            18.0,
+            26.0
+          ]
+        },
+        "changed_flag_accuracy": {
+          "mean": 80.17,
+          "sd": 0.76,
+          "ci95_half_width": 1.9,
+          "min": 79.5,
+          "max": 81.0,
+          "values": [
+            80.0,
+            79.5,
+            81.0
+          ]
+        },
+        "applied_update_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "closure_inertia_rate": {
+          "mean": 0.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 0.0,
+          "max": 0.0,
+          "values": [
+            0.0,
+            0.0,
+            0.0
+          ]
+        },
+        "closure_retention_failure_rate": {
+          "mean": 78.67,
+          "sd": 4.16,
+          "ci95_half_width": 10.34,
+          "min": 74.0,
+          "max": 82.0,
+          "values": [
+            80.0,
+            82.0,
+            74.0
+          ]
+        },
+        "anchoring_error_rate": {
+          "mean": 1.78,
+          "sd": 1.02,
+          "ci95_half_width": 2.53,
+          "min": 0.67,
+          "max": 2.67,
+          "values": [
+            0.67,
+            2.67,
+            2.0
+          ]
+        }
+      },
+      "turn_decision_consistency": {
+        "n": 300,
+        "consistent": 277,
+        "rate": 92.33
+      },
+      "mean_parse_failures": 0.0,
+      "mean_api_errors": 0.0
+    }
+  }
+}

results/reports/closurebench_llama_4_maverick_report.json

@@ -0,0 +1,390 @@
+{
+  "n_scored": 960,
+  "n_dataset": 960,
+  "coverage": 100.0,
+  "api_errors": 0,
+  "json_valid": {
+    "n": 46,
+    "rate": 4.79
+  },
+  "answer_parse": {
+    "n": 960,
+    "rate": 100.0
+  },
+  "overall_accuracy": {
+    "correct": 959,
+    "n": 960,
+    "accuracy": 99.9
+  },
+  "by_semantics": {
+    "cwa": {
+      "n": 320,
+      "correct": 319,
+      "accuracy": 99.69
+    },
+    "lcwa": {
+      "n": 320,
+      "correct": 320,
+      "accuracy": 100.0
+    },
+    "owa": {
+      "n": 320,
+      "correct": 320,
+      "accuracy": 100.0
+    }
+  },
+  "by_subset": {
+    "control_entailed": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "control_explicit_negative": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "core_contrastive": {
+      "n": 576,
+      "correct": 575,
+      "accuracy": 99.83
+    }
+  },
+  "by_family": {
+    "closed_derived_missing_antecedent": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "closed_missing_direct": {
+      "n": 96,
+      "correct": 95,
+      "accuracy": 98.96
+    },
+    "closed_missing_with_open_distractor": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "entailed_closed_conclusion": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "entailed_open_conclusion": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "explicit_negative_closed": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "explicit_negative_open": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "open_derived_missing_antecedent": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "open_missing_direct": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "open_missing_with_closed_distractor": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    }
+  },
+  "by_family_and_semantics": {
+    "closed_derived_missing_antecedent": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "closed_missing_direct": {
+      "cwa": {
+        "n": 32,
+        "correct": 31,
+        "accuracy": 96.88
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "closed_missing_with_open_distractor": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "entailed_closed_conclusion": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "entailed_open_conclusion": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "explicit_negative_closed": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "explicit_negative_open": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "open_derived_missing_antecedent": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "open_missing_direct": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "open_missing_with_closed_distractor": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    }
+  },
+  "by_gold_answer": {
+    "false": {
+      "n": 480,
+      "correct": 479,
+      "accuracy": 99.79
+    },
+    "true": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "unknown": {
+      "n": 288,
+      "correct": 288,
+      "accuracy": 100.0
+    }
+  },
+  "by_gold_reason_type": {
+    "entailed_by_fact_or_rule": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "explicit_negative_fact": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "global_closed_world_underivable": {
+      "n": 192,
+      "correct": 191,
+      "accuracy": 99.48
+    },
+    "local_closed_world_underivable": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "local_open_world_underivable": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "open_world_underivable": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    }
+  },
+  "semantic_switch_accuracy": {
+    "complete_base_groups": 320,
+    "correct_groups": 319,
+    "accuracy": 99.69
+  },
+  "core_contrastive_switch_accuracy": {
+    "complete_base_groups": 192,
+    "correct_groups": 191,
+    "accuracy": 99.48
+  },
+  "unknown_accuracy": {
+    "correct": 288,
+    "n": 288,
+    "accuracy": 100.0
+  },
+  "false_accuracy": {
+    "correct": 479,
+    "n": 480,
+    "accuracy": 99.79
+  },
+  "true_accuracy": {
+    "correct": 192,
+    "n": 192,
+    "accuracy": 100.0
+  },
+  "lcwa_open_scope_accuracy": {
+    "correct": 96,
+    "n": 96,
+    "accuracy": 100.0
+  },
+  "lcwa_closed_scope_accuracy": {
+    "correct": 96,
+    "n": 96,
+    "accuracy": 100.0
+  },
+  "semantic_bias_index": {
+    "owa_unknown_predicted_false_rate": 0.0,
+    "lcwa_open_predicted_false_rate": 0.0,
+    "lcwa_closed_predicted_unknown_rate": 0.0
+  },
+  "error_types": {
+    "correct": 959,
+    "open_world_bias": 1
+  },
+  "baselines": {
+    "always_cwa": {
+      "correct": 672,
+      "accuracy": 70.0,
+      "n": 960
+    },
+    "always_owa": {
+      "correct": 672,
+      "accuracy": 70.0,
+      "n": 960
+    },
+    "ignore_lcwa": {
+      "correct": 864,
+      "accuracy": 90.0,
+      "n": 960
+    }
+  },
+  "models": [
+    "meta-llama/llama-4-maverick"
+  ]
+}

results/reports/closurebench_llama_4_scout_report.json

@@ -0,0 +1,394 @@
+{
+  "n_scored": 960,
+  "n_dataset": 960,
+  "coverage": 100.0,
+  "api_errors": 0,
+  "json_valid": {
+    "n": 882,
+    "rate": 91.88
+  },
+  "answer_parse": {
+    "n": 960,
+    "rate": 100.0
+  },
+  "overall_accuracy": {
+    "correct": 683,
+    "n": 960,
+    "accuracy": 71.15
+  },
+  "by_semantics": {
+    "cwa": {
+      "n": 320,
+      "correct": 221,
+      "accuracy": 69.06
+    },
+    "lcwa": {
+      "n": 320,
+      "correct": 185,
+      "accuracy": 57.81
+    },
+    "owa": {
+      "n": 320,
+      "correct": 277,
+      "accuracy": 86.56
+    }
+  },
+  "by_subset": {
+    "control_entailed": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "control_explicit_negative": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "core_contrastive": {
+      "n": 576,
+      "correct": 299,
+      "accuracy": 51.91
+    }
+  },
+  "by_family": {
+    "closed_derived_missing_antecedent": {
+      "n": 96,
+      "correct": 43,
+      "accuracy": 44.79
+    },
+    "closed_missing_direct": {
+      "n": 96,
+      "correct": 33,
+      "accuracy": 34.38
+    },
+    "closed_missing_with_open_distractor": {
+      "n": 96,
+      "correct": 36,
+      "accuracy": 37.5
+    },
+    "entailed_closed_conclusion": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "entailed_open_conclusion": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "explicit_negative_closed": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "explicit_negative_open": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "open_derived_missing_antecedent": {
+      "n": 96,
+      "correct": 37,
+      "accuracy": 38.54
+    },
+    "open_missing_direct": {
+      "n": 96,
+      "correct": 65,
+      "accuracy": 67.71
+    },
+    "open_missing_with_closed_distractor": {
+      "n": 96,
+      "correct": 85,
+      "accuracy": 88.54
+    }
+  },
+  "by_family_and_semantics": {
+    "closed_derived_missing_antecedent": {
+      "cwa": {
+        "n": 32,
+        "correct": 10,
+        "accuracy": 31.25
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 1,
+        "accuracy": 3.12
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "closed_missing_direct": {
+      "cwa": {
+        "n": 32,
+        "correct": 0,
+        "accuracy": 0.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 1,
+        "accuracy": 3.12
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "closed_missing_with_open_distractor": {
+      "cwa": {
+        "n": 32,
+        "correct": 3,
+        "accuracy": 9.38
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 1,
+        "accuracy": 3.12
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "entailed_closed_conclusion": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "entailed_open_conclusion": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "explicit_negative_closed": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "explicit_negative_open": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "open_derived_missing_antecedent": {
+      "cwa": {
+        "n": 32,
+        "correct": 31,
+        "accuracy": 96.88
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 2,
+        "accuracy": 6.25
+      },
+      "owa": {
+        "n": 32,
+        "correct": 4,
+        "accuracy": 12.5
+      }
+    },
+    "open_missing_direct": {
+      "cwa": {
+        "n": 32,
+        "correct": 28,
+        "accuracy": 87.5
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 20,
+        "accuracy": 62.5
+      },
+      "owa": {
+        "n": 32,
+        "correct": 17,
+        "accuracy": 53.12
+      }
+    },
+    "open_missing_with_closed_distractor": {
+      "cwa": {
+        "n": 32,
+        "correct": 21,
+        "accuracy": 65.62
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    }
+  },
+  "by_gold_answer": {
+    "false": {
+      "n": 480,
+      "correct": 288,
+      "accuracy": 60.0
+    },
+    "true": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "unknown": {
+      "n": 288,
+      "correct": 203,
+      "accuracy": 70.49
+    }
+  },
+  "by_gold_reason_type": {
+    "entailed_by_fact_or_rule": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "explicit_negative_fact": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "global_closed_world_underivable": {
+      "n": 192,
+      "correct": 93,
+      "accuracy": 48.44
+    },
+    "local_closed_world_underivable": {
+      "n": 96,
+      "correct": 3,
+      "accuracy": 3.12
+    },
+    "local_open_world_underivable": {
+      "n": 96,
+      "correct": 54,
+      "accuracy": 56.25
+    },
+    "open_world_underivable": {
+      "n": 192,
+      "correct": 149,
+      "accuracy": 77.6
+    }
+  },
+  "semantic_switch_accuracy": {
+    "complete_base_groups": 320,
+    "correct_groups": 163,
+    "accuracy": 50.94
+  },
+  "core_contrastive_switch_accuracy": {
+    "complete_base_groups": 192,
+    "correct_groups": 35,
+    "accuracy": 18.23
+  },
+  "unknown_accuracy": {
+    "correct": 203,
+    "n": 288,
+    "accuracy": 70.49
+  },
+  "false_accuracy": {
+    "correct": 288,
+    "n": 480,
+    "accuracy": 60.0
+  },
+  "true_accuracy": {
+    "correct": 192,
+    "n": 192,
+    "accuracy": 100.0
+  },
+  "lcwa_open_scope_accuracy": {
+    "correct": 54,
+    "n": 96,
+    "accuracy": 56.25
+  },
+  "lcwa_closed_scope_accuracy": {
+    "correct": 3,
+    "n": 96,
+    "accuracy": 3.12
+  },
+  "semantic_bias_index": {
+    "owa_unknown_predicted_false_rate": 22.4,
+    "lcwa_open_predicted_false_rate": 43.75,
+    "lcwa_closed_predicted_unknown_rate": 93.75
+  },
+  "error_types": {
+    "correct": 683,
+    "open_world_bias": 99,
+    "missed_local_closure": 90,
+    "closure_scope_error_false_for_open_predicate": 42,
+    "closed_world_bias": 43,
+    "hallucinated_true": 3
+  },
+  "baselines": {
+    "always_cwa": {
+      "correct": 672,
+      "accuracy": 70.0,
+      "n": 960
+    },
+    "always_owa": {
+      "correct": 672,
+      "accuracy": 70.0,
+      "n": 960
+    },
+    "ignore_lcwa": {
+      "correct": 864,
+      "accuracy": 90.0,
+      "n": 960
+    }
+  },
+  "models": [
+    "meta-llama/llama-4-scout"
+  ]
+}

results/reports/closurebench_mistral_medium_3_5_report.json

@@ -0,0 +1,391 @@
+{
+  "n_scored": 960,
+  "n_dataset": 960,
+  "coverage": 100.0,
+  "api_errors": 0,
+  "json_valid": {
+    "n": 958,
+    "rate": 99.79
+  },
+  "answer_parse": {
+    "n": 960,
+    "rate": 100.0
+  },
+  "overall_accuracy": {
+    "correct": 922,
+    "n": 960,
+    "accuracy": 96.04
+  },
+  "by_semantics": {
+    "cwa": {
+      "n": 320,
+      "correct": 318,
+      "accuracy": 99.38
+    },
+    "lcwa": {
+      "n": 320,
+      "correct": 284,
+      "accuracy": 88.75
+    },
+    "owa": {
+      "n": 320,
+      "correct": 320,
+      "accuracy": 100.0
+    }
+  },
+  "by_subset": {
+    "control_entailed": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "control_explicit_negative": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "core_contrastive": {
+      "n": 576,
+      "correct": 538,
+      "accuracy": 93.4
+    }
+  },
+  "by_family": {
+    "closed_derived_missing_antecedent": {
+      "n": 96,
+      "correct": 68,
+      "accuracy": 70.83
+    },
+    "closed_missing_direct": {
+      "n": 96,
+      "correct": 91,
+      "accuracy": 94.79
+    },
+    "closed_missing_with_open_distractor": {
+      "n": 96,
+      "correct": 91,
+      "accuracy": 94.79
+    },
+    "entailed_closed_conclusion": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "entailed_open_conclusion": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "explicit_negative_closed": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "explicit_negative_open": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "open_derived_missing_antecedent": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "open_missing_direct": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "open_missing_with_closed_distractor": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    }
+  },
+  "by_family_and_semantics": {
+    "closed_derived_missing_antecedent": {
+      "cwa": {
+        "n": 32,
+        "correct": 30,
+        "accuracy": 93.75
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 6,
+        "accuracy": 18.75
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "closed_missing_direct": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 27,
+        "accuracy": 84.38
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "closed_missing_with_open_distractor": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 27,
+        "accuracy": 84.38
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "entailed_closed_conclusion": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "entailed_open_conclusion": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "explicit_negative_closed": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "explicit_negative_open": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "open_derived_missing_antecedent": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "open_missing_direct": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "open_missing_with_closed_distractor": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    }
+  },
+  "by_gold_answer": {
+    "false": {
+      "n": 480,
+      "correct": 442,
+      "accuracy": 92.08
+    },
+    "true": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "unknown": {
+      "n": 288,
+      "correct": 288,
+      "accuracy": 100.0
+    }
+  },
+  "by_gold_reason_type": {
+    "entailed_by_fact_or_rule": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "explicit_negative_fact": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "global_closed_world_underivable": {
+      "n": 192,
+      "correct": 190,
+      "accuracy": 98.96
+    },
+    "local_closed_world_underivable": {
+      "n": 96,
+      "correct": 60,
+      "accuracy": 62.5
+    },
+    "local_open_world_underivable": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "open_world_underivable": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    }
+  },
+  "semantic_switch_accuracy": {
+    "complete_base_groups": 320,
+    "correct_groups": 284,
+    "accuracy": 88.75
+  },
+  "core_contrastive_switch_accuracy": {
+    "complete_base_groups": 192,
+    "correct_groups": 156,
+    "accuracy": 81.25
+  },
+  "unknown_accuracy": {
+    "correct": 288,
+    "n": 288,
+    "accuracy": 100.0
+  },
+  "false_accuracy": {
+    "correct": 442,
+    "n": 480,
+    "accuracy": 92.08
+  },
+  "true_accuracy": {
+    "correct": 192,
+    "n": 192,
+    "accuracy": 100.0
+  },
+  "lcwa_open_scope_accuracy": {
+    "correct": 96,
+    "n": 96,
+    "accuracy": 100.0
+  },
+  "lcwa_closed_scope_accuracy": {
+    "correct": 60,
+    "n": 96,
+    "accuracy": 62.5
+  },
+  "semantic_bias_index": {
+    "owa_unknown_predicted_false_rate": 0.0,
+    "lcwa_open_predicted_false_rate": 0.0,
+    "lcwa_closed_predicted_unknown_rate": 37.5
+  },
+  "error_types": {
+    "correct": 922,
+    "missed_local_closure": 36,
+    "open_world_bias": 2
+  },
+  "baselines": {
+    "always_cwa": {
+      "correct": 672,
+      "accuracy": 70.0,
+      "n": 960
+    },
+    "always_owa": {
+      "correct": 672,
+      "accuracy": 70.0,
+      "n": 960
+    },
+    "ignore_lcwa": {
+      "correct": 864,
+      "accuracy": 90.0,
+      "n": 960
+    }
+  },
+  "models": [
+    "mistralai/mistral-medium-3-5"
+  ]
+}

results/reports/closurebench_mistral_small_2603_report.json

@@ -0,0 +1,393 @@
+{
+  "n_scored": 960,
+  "n_dataset": 960,
+  "coverage": 100.0,
+  "api_errors": 0,
+  "json_valid": {
+    "n": 960,
+    "rate": 100.0
+  },
+  "answer_parse": {
+    "n": 960,
+    "rate": 100.0
+  },
+  "overall_accuracy": {
+    "correct": 774,
+    "n": 960,
+    "accuracy": 80.62
+  },
+  "by_semantics": {
+    "cwa": {
+      "n": 320,
+      "correct": 294,
+      "accuracy": 91.88
+    },
+    "lcwa": {
+      "n": 320,
+      "correct": 185,
+      "accuracy": 57.81
+    },
+    "owa": {
+      "n": 320,
+      "correct": 295,
+      "accuracy": 92.19
+    }
+  },
+  "by_subset": {
+    "control_entailed": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "control_explicit_negative": {
+      "n": 192,
+      "correct": 190,
+      "accuracy": 98.96
+    },
+    "core_contrastive": {
+      "n": 576,
+      "correct": 392,
+      "accuracy": 68.06
+    }
+  },
+  "by_family": {
+    "closed_derived_missing_antecedent": {
+      "n": 96,
+      "correct": 72,
+      "accuracy": 75.0
+    },
+    "closed_missing_direct": {
+      "n": 96,
+      "correct": 64,
+      "accuracy": 66.67
+    },
+    "closed_missing_with_open_distractor": {
+      "n": 96,
+      "correct": 59,
+      "accuracy": 61.46
+    },
+    "entailed_closed_conclusion": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "entailed_open_conclusion": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "explicit_negative_closed": {
+      "n": 96,
+      "correct": 96,
+      "accuracy": 100.0
+    },
+    "explicit_negative_open": {
+      "n": 96,
+      "correct": 94,
+      "accuracy": 97.92
+    },
+    "open_derived_missing_antecedent": {
+      "n": 96,
+      "correct": 43,
+      "accuracy": 44.79
+    },
+    "open_missing_direct": {
+      "n": 96,
+      "correct": 82,
+      "accuracy": 85.42
+    },
+    "open_missing_with_closed_distractor": {
+      "n": 96,
+      "correct": 72,
+      "accuracy": 75.0
+    }
+  },
+  "by_family_and_semantics": {
+    "closed_derived_missing_antecedent": {
+      "cwa": {
+        "n": 32,
+        "correct": 31,
+        "accuracy": 96.88
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 11,
+        "accuracy": 34.38
+      },
+      "owa": {
+        "n": 32,
+        "correct": 30,
+        "accuracy": 93.75
+      }
+    },
+    "closed_missing_direct": {
+      "cwa": {
+        "n": 32,
+        "correct": 21,
+        "accuracy": 65.62
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 11,
+        "accuracy": 34.38
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "closed_missing_with_open_distractor": {
+      "cwa": {
+        "n": 32,
+        "correct": 21,
+        "accuracy": 65.62
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 6,
+        "accuracy": 18.75
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "entailed_closed_conclusion": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "entailed_open_conclusion": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "explicit_negative_closed": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "explicit_negative_open": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 30,
+        "accuracy": 93.75
+      }
+    },
+    "open_derived_missing_antecedent": {
+      "cwa": {
+        "n": 32,
+        "correct": 31,
+        "accuracy": 96.88
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 1,
+        "accuracy": 3.12
+      },
+      "owa": {
+        "n": 32,
+        "correct": 11,
+        "accuracy": 34.38
+      }
+    },
+    "open_missing_direct": {
+      "cwa": {
+        "n": 32,
+        "correct": 30,
+        "accuracy": 93.75
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 20,
+        "accuracy": 62.5
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    },
+    "open_missing_with_closed_distractor": {
+      "cwa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      },
+      "lcwa": {
+        "n": 32,
+        "correct": 8,
+        "accuracy": 25.0
+      },
+      "owa": {
+        "n": 32,
+        "correct": 32,
+        "accuracy": 100.0
+      }
+    }
+  },
+  "by_gold_answer": {
+    "false": {
+      "n": 480,
+      "correct": 384,
+      "accuracy": 80.0
+    },
+    "true": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "unknown": {
+      "n": 288,
+      "correct": 198,
+      "accuracy": 68.75
+    }
+  },
+  "by_gold_reason_type": {
+    "entailed_by_fact_or_rule": {
+      "n": 192,
+      "correct": 192,
+      "accuracy": 100.0
+    },
+    "explicit_negative_fact": {
+      "n": 192,
+      "correct": 190,
+      "accuracy": 98.96
+    },
+    "global_closed_world_underivable": {
+      "n": 192,
+      "correct": 166,
+      "accuracy": 86.46
+    },
+    "local_closed_world_underivable": {
+      "n": 96,
+      "correct": 28,
+      "accuracy": 29.17
+    },
+    "local_open_world_underivable": {
+      "n": 96,
+      "correct": 29,
+      "accuracy": 30.21
+    },
+    "open_world_underivable": {
+      "n": 192,
+      "correct": 169,
+      "accuracy": 88.02
+    }
+  },
+  "semantic_switch_accuracy": {
+    "complete_base_groups": 320,
+    "correct_groups": 179,
+    "accuracy": 55.94
+  },
+  "core_contrastive_switch_accuracy": {
+    "complete_base_groups": 192,
+    "correct_groups": 53,
+    "accuracy": 27.6
+  },
+  "unknown_accuracy": {
+    "correct": 198,
+    "n": 288,
+    "accuracy": 68.75
+  },
+  "false_accuracy": {
+    "correct": 384,
+    "n": 480,
+    "accuracy": 80.0
+  },
+  "true_accuracy": {
+    "correct": 192,
+    "n": 192,
+    "accuracy": 100.0
+  },
+  "lcwa_open_scope_accuracy": {
+    "correct": 29,
+    "n": 96,
+    "accuracy": 30.21
+  },
+  "lcwa_closed_scope_accuracy": {
+    "correct": 28,
+    "n": 96,
+    "accuracy": 29.17
+  },
+  "semantic_bias_index": {
+    "owa_unknown_predicted_false_rate": 11.98,
+    "lcwa_open_predicted_false_rate": 69.79,
+    "lcwa_closed_predicted_unknown_rate": 70.83
+  },
+  "error_types": {
+    "correct": 774,
+    "open_world_bias": 28,
+    "closure_scope_error_false_for_open_predicate": 67,
+    "closed_world_bias": 23,
+    "missed_local_closure": 68
+  },
+  "baselines": {
+    "always_cwa": {
+      "correct": 672,
+      "accuracy": 70.0,
+      "n": 960
+    },
+    "always_owa": {
+      "correct": 672,
+      "accuracy": 70.0,
+      "n": 960
+    },
+    "ignore_lcwa": {
+      "correct": 864,
+      "accuracy": 90.0,
+      "n": 960
+    }
+  },
+  "models": [
+    "mistralai/mistral-small-2603"
+  ]
+}

results/reports/closurebench_model_comparison.json

@@ -0,0 +1,336 @@
+{
+  "dataset": "data/base/full.jsonl",
+  "n_dataset": 960,
+  "runs": {
+    "mistral-small": {
+      "n_scored": 960,
+      "n_dataset": 960,
+      "coverage": 100.0,
+      "models": [
+        "mistralai/mistral-small-2603"
+      ],
+      "api_errors": 0,
+      "parse_failures": 0,
+      "semantic_switch_accuracy": {
+        "correct": 179,
+        "n": 320,
+        "accuracy": 55.94
+      },
+      "core_contrastive_switch_accuracy": {
+        "correct": 53,
+        "n": 192,
+        "accuracy": 27.6
+      },
+      "lcwa_closed_scope_accuracy": {
+        "correct": 28,
+        "n": 96,
+        "accuracy": 29.17
+      },
+      "lcwa_open_scope_accuracy": {
+        "correct": 29,
+        "n": 96,
+        "accuracy": 30.21
+      },
+      "cwa_accuracy": {
+        "correct": 294,
+        "n": 320,
+        "accuracy": 91.88
+      },
+      "owa_accuracy": {
+        "correct": 295,
+        "n": 320,
+        "accuracy": 92.19
+      },
+      "false_accuracy": {
+        "correct": 384,
+        "n": 480,
+        "accuracy": 80.0
+      },
+      "unknown_accuracy": {
+        "correct": 198,
+        "n": 288,
+        "accuracy": 68.75
+      },
+      "overall_accuracy": {
+        "correct": 774,
+        "n": 960,
+        "accuracy": 80.62
+      }
+    },
+    "deepseek-flash": {
+      "n_scored": 960,
+      "n_dataset": 960,
+      "coverage": 100.0,
+      "models": [
+        "deepseek-v4-flash"
+      ],
+      "api_errors": 0,
+      "parse_failures": 0,
+      "semantic_switch_accuracy": {
+        "correct": 264,
+        "n": 320,
+        "accuracy": 82.5
+      },
+      "core_contrastive_switch_accuracy": {
+        "correct": 136,
+        "n": 192,
+        "accuracy": 70.83
+      },
+      "lcwa_closed_scope_accuracy": {
+        "correct": 48,
+        "n": 96,
+        "accuracy": 50.0
+      },
+      "lcwa_open_scope_accuracy": {
+        "correct": 96,
+        "n": 96,
+        "accuracy": 100.0
+      },
+      "cwa_accuracy": {
+        "correct": 280,
+        "n": 320,
+        "accuracy": 87.5
+      },
+      "owa_accuracy": {
+        "correct": 320,
+        "n": 320,
+        "accuracy": 100.0
+      },
+      "false_accuracy": {
+        "correct": 392,
+        "n": 480,
+        "accuracy": 81.67
+      },
+      "unknown_accuracy": {
+        "correct": 288,
+        "n": 288,
+        "accuracy": 100.0
+      },
+      "overall_accuracy": {
+        "correct": 872,
+        "n": 960,
+        "accuracy": 90.83
+      }
+    },
+    "deepseek-pro": {
+      "n_scored": 960,
+      "n_dataset": 960,
+      "coverage": 100.0,
+      "models": [
+        "deepseek-v4-pro"
+      ],
+      "api_errors": 0,
+      "parse_failures": 0,
+      "semantic_switch_accuracy": {
+        "correct": 276,
+        "n": 320,
+        "accuracy": 86.25
+      },
+      "core_contrastive_switch_accuracy": {
+        "correct": 148,
+        "n": 192,
+        "accuracy": 77.08
+      },
+      "lcwa_closed_scope_accuracy": {
+        "correct": 60,
+        "n": 96,
+        "accuracy": 62.5
+      },
+      "lcwa_open_scope_accuracy": {
+        "correct": 96,
+        "n": 96,
+        "accuracy": 100.0
+      },
+      "cwa_accuracy": {
+        "correct": 294,
+        "n": 320,
+        "accuracy": 91.88
+      },
+      "owa_accuracy": {
+        "correct": 320,
+        "n": 320,
+        "accuracy": 100.0
+      },
+      "false_accuracy": {
+        "correct": 418,
+        "n": 480,
+        "accuracy": 87.08
+      },
+      "unknown_accuracy": {
+        "correct": 288,
+        "n": 288,
+        "accuracy": 100.0
+      },
+      "overall_accuracy": {
+        "correct": 898,
+        "n": 960,
+        "accuracy": 93.54
+      }
+    },
+    "mistral-medium": {
+      "n_scored": 960,
+      "n_dataset": 960,
+      "coverage": 100.0,
+      "models": [
+        "mistralai/mistral-medium-3-5"
+      ],
+      "api_errors": 0,
+      "parse_failures": 0,
+      "semantic_switch_accuracy": {
+        "correct": 284,
+        "n": 320,
+        "accuracy": 88.75
+      },
+      "core_contrastive_switch_accuracy": {
+        "correct": 156,
+        "n": 192,
+        "accuracy": 81.25
+      },
+      "lcwa_closed_scope_accuracy": {
+        "correct": 60,
+        "n": 96,
+        "accuracy": 62.5
+      },
+      "lcwa_open_scope_accuracy": {
+        "correct": 96,
+        "n": 96,
+        "accuracy": 100.0
+      },
+      "cwa_accuracy": {
+        "correct": 318,
+        "n": 320,
+        "accuracy": 99.38
+      },
+      "owa_accuracy": {
+        "correct": 320,
+        "n": 320,
+        "accuracy": 100.0
+      },
+      "false_accuracy": {
+        "correct": 442,
+        "n": 480,
+        "accuracy": 92.08
+      },
+      "unknown_accuracy": {
+        "correct": 288,
+        "n": 288,
+        "accuracy": 100.0
+      },
+      "overall_accuracy": {
+        "correct": 922,
+        "n": 960,
+        "accuracy": 96.04
+      }
+    },
+    "llama-scout": {
+      "n_scored": 960,
+      "n_dataset": 960,
+      "coverage": 100.0,
+      "models": [
+        "meta-llama/llama-4-scout"
+      ],
+      "api_errors": 0,
+      "parse_failures": 0,
+      "semantic_switch_accuracy": {
+        "correct": 163,
+        "n": 320,
+        "accuracy": 50.94
+      },
+      "core_contrastive_switch_accuracy": {
+        "correct": 35,
+        "n": 192,
+        "accuracy": 18.23
+      },
+      "lcwa_closed_scope_accuracy": {
+        "correct": 3,
+        "n": 96,
+        "accuracy": 3.12
+      },
+      "lcwa_open_scope_accuracy": {
+        "correct": 54,
+        "n": 96,
+        "accuracy": 56.25
+      },
+      "cwa_accuracy": {
+        "correct": 221,
+        "n": 320,
+        "accuracy": 69.06
+      },
+      "owa_accuracy": {
+        "correct": 277,
+        "n": 320,
+        "accuracy": 86.56
+      },
+      "false_accuracy": {
+        "correct": 288,
+        "n": 480,
+        "accuracy": 60.0
+      },
+      "unknown_accuracy": {
+        "correct": 203,
+        "n": 288,
+        "accuracy": 70.49
+      },
+      "overall_accuracy": {
+        "correct": 683,
+        "n": 960,
+        "accuracy": 71.15
+      }
+    },
+    "llama-maverick": {
+      "n_scored": 960,
+      "n_dataset": 960,
+      "coverage": 100.0,
+      "models": [
+        "meta-llama/llama-4-maverick"
+      ],
+      "api_errors": 0,
+      "parse_failures": 0,
+      "semantic_switch_accuracy": {
+        "correct": 319,
+        "n": 320,
+        "accuracy": 99.69
+      },
+      "core_contrastive_switch_accuracy": {
+        "correct": 191,
+        "n": 192,
+        "accuracy": 99.48
+      },
+      "lcwa_closed_scope_accuracy": {
+        "correct": 96,
+        "n": 96,
+        "accuracy": 100.0
+      },
+      "lcwa_open_scope_accuracy": {
+        "correct": 96,
+        "n": 96,
+        "accuracy": 100.0
+      },
+      "cwa_accuracy": {
+        "correct": 319,
+        "n": 320,
+        "accuracy": 99.69
+      },
+      "owa_accuracy": {
+        "correct": 320,
+        "n": 320,
+        "accuracy": 100.0
+      },
+      "false_accuracy": {
+        "correct": 479,
+        "n": 480,
+        "accuracy": 99.79
+      },
+      "unknown_accuracy": {
+        "correct": 288,
+        "n": 288,
+        "accuracy": 100.0
+      },
+      "overall_accuracy": {
+        "correct": 959,
+        "n": 960,
+        "accuracy": 99.9
+      }
+    }
+  }
+}

results/reports/closurebench_multi_agent_manifest.json

@@ -0,0 +1,37 @@
+{
+  "dataset": "data/multi_agent/full.jsonl",
+  "dataset_size": 360,
+  "shards": 4,
+  "runs": [
+    {
+      "label": "mistral-small",
+      "model": "mistralai/mistral-small-2603",
+      "scored": "results/scored/multi_agent/closure_contract_v3_multi_agent_mistral-small_scored.jsonl"
+    },
+    {
+      "label": "deepseek-flash",
+      "model": "deepseek-v4-flash",
+      "scored": "results/scored/multi_agent/closure_contract_v3_multi_agent_deepseek-flash_scored.jsonl"
+    },
+    {
+      "label": "deepseek-pro",
+      "model": "deepseek-v4-pro",
+      "scored": "results/scored/multi_agent/closure_contract_v3_multi_agent_deepseek-pro_scored.jsonl"
+    },
+    {
+      "label": "mistral-medium",
+      "model": "mistralai/mistral-medium-3-5",
+      "scored": "results/scored/multi_agent/closure_contract_v3_multi_agent_mistral-medium_scored.jsonl"
+    },
+    {
+      "label": "llama-scout",
+      "model": "meta-llama/llama-4-scout",
+      "scored": "results/scored/multi_agent/closure_contract_v3_multi_agent_llama-scout_scored.jsonl"
+    },
+    {
+      "label": "llama-maverick",
+      "model": "meta-llama/llama-4-maverick",
+      "scored": "results/scored/multi_agent/closure_contract_v3_multi_agent_llama-maverick_scored.jsonl"
+    }
+  ]
+}

results/reports/closurebench_multi_agent_replicate_manifest.json

@@ -0,0 +1,135 @@
+{
+  "dataset": "data/multi_agent/full.jsonl",
+  "dataset_size": 360,
+  "repeats": 3,
+  "shards": 4,
+  "temperature": 0.0,
+  "runs": [
+    {
+      "label": "mistral-small",
+      "model": "mistralai/mistral-small-2603",
+      "scored": "results/scored/multi_agent/closure_contract_v3_multi_agent_mistral-small_scored.jsonl",
+      "repeat": 1,
+      "temperature": 0.0
+    },
+    {
+      "label": "mistral-small",
+      "model": "mistralai/mistral-small-2603",
+      "repeat": 2,
+      "temperature": 0.0,
+      "scored": "results/scored/multi_agent/closurebench_multi_agent_mistral-small_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "mistral-small",
+      "model": "mistralai/mistral-small-2603",
+      "repeat": 3,
+      "temperature": 0.0,
+      "scored": "results/scored/multi_agent/closurebench_multi_agent_mistral-small_repeat_3_scored.jsonl"
+    },
+    {
+      "label": "deepseek-flash",
+      "model": "deepseek-v4-flash",
+      "scored": "results/scored/multi_agent/closure_contract_v3_multi_agent_deepseek-flash_scored.jsonl",
+      "repeat": 1,
+      "temperature": 0.0
+    },
+    {
+      "label": "deepseek-flash",
+      "model": "deepseek-v4-flash",
+      "repeat": 2,
+      "temperature": 0.0,
+      "scored": "results/scored/multi_agent/closurebench_multi_agent_deepseek-flash_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "deepseek-flash",
+      "model": "deepseek-v4-flash",
+      "repeat": 3,
+      "temperature": 0.0,
+      "scored": "results/scored/multi_agent/closurebench_multi_agent_deepseek-flash_repeat_3_scored.jsonl"
+    },
+    {
+      "label": "deepseek-pro",
+      "model": "deepseek-v4-pro",
+      "scored": "results/scored/multi_agent/closure_contract_v3_multi_agent_deepseek-pro_scored.jsonl",
+      "repeat": 1,
+      "temperature": 0.0
+    },
+    {
+      "label": "deepseek-pro",
+      "model": "deepseek-v4-pro",
+      "repeat": 2,
+      "temperature": 0.0,
+      "scored": "results/scored/multi_agent/closurebench_multi_agent_deepseek-pro_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "deepseek-pro",
+      "model": "deepseek-v4-pro",
+      "repeat": 3,
+      "temperature": 0.0,
+      "scored": "results/scored/multi_agent/closurebench_multi_agent_deepseek-pro_repeat_3_scored.jsonl"
+    },
+    {
+      "label": "mistral-medium",
+      "model": "mistralai/mistral-medium-3-5",
+      "scored": "results/scored/multi_agent/closure_contract_v3_multi_agent_mistral-medium_scored.jsonl",
+      "repeat": 1,
+      "temperature": 0.0
+    },
+    {
+      "label": "mistral-medium",
+      "model": "mistralai/mistral-medium-3-5",
+      "repeat": 2,
+      "temperature": 0.0,
+      "scored": "results/scored/multi_agent/closurebench_multi_agent_mistral-medium_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "mistral-medium",
+      "model": "mistralai/mistral-medium-3-5",
+      "repeat": 3,
+      "temperature": 0.0,
+      "scored": "results/scored/multi_agent/closurebench_multi_agent_mistral-medium_repeat_3_scored.jsonl"
+    },
+    {
+      "label": "llama-scout",
+      "model": "meta-llama/llama-4-scout",
+      "scored": "results/scored/multi_agent/closure_contract_v3_multi_agent_llama-scout_scored.jsonl",
+      "repeat": 1,
+      "temperature": 0.0
+    },
+    {
+      "label": "llama-scout",
+      "model": "meta-llama/llama-4-scout",
+      "repeat": 2,
+      "temperature": 0.0,
+      "scored": "results/scored/multi_agent/closurebench_multi_agent_llama-scout_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "llama-scout",
+      "model": "meta-llama/llama-4-scout",
+      "repeat": 3,
+      "temperature": 0.0,
+      "scored": "results/scored/multi_agent/closurebench_multi_agent_llama-scout_repeat_3_scored.jsonl"
+    },
+    {
+      "label": "llama-maverick",
+      "model": "meta-llama/llama-4-maverick",
+      "scored": "results/scored/multi_agent/closure_contract_v3_multi_agent_llama-maverick_scored.jsonl",
+      "repeat": 1,
+      "temperature": 0.0
+    },
+    {
+      "label": "llama-maverick",
+      "model": "meta-llama/llama-4-maverick",
+      "repeat": 2,
+      "temperature": 0.0,
+      "scored": "results/scored/multi_agent/closurebench_multi_agent_llama-maverick_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "llama-maverick",
+      "model": "meta-llama/llama-4-maverick",
+      "repeat": 3,
+      "temperature": 0.0,
+      "scored": "results/scored/multi_agent/closurebench_multi_agent_llama-maverick_repeat_3_scored.jsonl"
+    }
+  ]
+}

results/reports/closurebench_multi_agent_summary.json

@@ -0,0 +1,3762 @@
+{
+  "dataset": "data/multi_agent/full.jsonl",
+  "n_dataset": 360,
+  "runs": [
+    {
+      "label": "mistral-small",
+      "model": "mistralai/mistral-small-2603",
+      "repeat": 1,
+      "temperature": 0.0,
+      "report": {
+        "n_dataset": 360,
+        "n_scored": 360,
+        "coverage": 100.0,
+        "models": [
+          "mistralai/mistral-small-2603"
+        ],
+        "api_errors": 0,
+        "parse_failures": 0,
+        "json_valid": {
+          "n": 360,
+          "rate": 100.0
+        },
+        "truth_accuracy": {
+          "correct": 291,
+          "n": 360,
+          "accuracy": 80.83
+        },
+        "source_accuracy": {
+          "correct": 230,
+          "n": 360,
+          "accuracy": 63.89
+        },
+        "closure_accuracy": {
+          "correct": 311,
+          "n": 360,
+          "accuracy": 86.39
+        },
+        "full_accuracy": {
+          "correct": 195,
+          "n": 360,
+          "accuracy": 54.17
+        },
+        "semantic_switch_full_accuracy": {
+          "correct": 17,
+          "n": 120,
+          "accuracy": 14.17
+        },
+        "core_switch_full_accuracy": {
+          "correct": 11,
+          "n": 72,
+          "accuracy": 15.28
+        },
+        "closed_absence_accuracy": {
+          "correct": 31,
+          "n": 106,
+          "accuracy": 29.25
+        },
+        "open_absence_accuracy": {
+          "correct": 93,
+          "n": 110,
+          "accuracy": 84.55
+        },
+        "closure_leakage_rate": {
+          "count": 12,
+          "n": 110,
+          "rate": 10.91
+        },
+        "closure_loss_rate": {
+          "count": 68,
+          "n": 106,
+          "rate": 64.15
+        },
+        "agent_a_source_accuracy": {
+          "correct": 71,
+          "n": 166,
+          "accuracy": 42.77
+        },
+        "agent_b_source_accuracy": {
+          "correct": 144,
+          "n": 179,
+          "accuracy": 80.45
+        },
+        "by_semantics": {
+          "cwa": {
+            "n": 120,
+            "truth": 56.67,
+            "source": 35.0,
+            "closure": 75.83,
+            "full": 19.17
+          },
+          "lcwa": {
+            "n": 120,
+            "truth": 85.83,
+            "source": 75.83,
+            "closure": 88.33,
+            "full": 66.67
+          },
+          "owa": {
+            "n": 120,
+            "truth": 100.0,
+            "source": 80.83,
+            "closure": 95.0,
+            "full": 76.67
+          }
+        },
+        "by_closure_handling": {
+          "closed_absence": {
+            "n": 106,
+            "truth": 34.91,
+            "source": 30.19,
+            "closure": 86.79,
+            "full": 29.25
+          },
+          "entailed_or_explicit": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 30.56,
+            "closure": 95.83,
+            "full": 26.39
+          },
+          "explicit_negative": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 72.22,
+            "full": 72.22
+          },
+          "open_absence": {
+            "n": 110,
+            "truth": 100.0,
+            "source": 94.55,
+            "closure": 89.09,
+            "full": 84.55
+          }
+        },
+        "by_source_used": {
+          "agent_a": {
+            "correct": 71,
+            "n": 166,
+            "accuracy": 42.77
+          },
+          "agent_b": {
+            "correct": 144,
+            "n": 179,
+            "accuracy": 80.45
+          },
+          "both": {
+            "correct": 15,
+            "n": 15,
+            "accuracy": 100.0
+          }
+        },
+        "by_family": {
+          "closed_derived_missing_antecedent": {
+            "correct": 8,
+            "n": 21,
+            "accuracy": 38.1
+          },
+          "closed_missing_direct": {
+            "correct": 25,
+            "n": 42,
+            "accuracy": 59.52
+          },
+          "closed_missing_with_open_distractor": {
+            "correct": 18,
+            "n": 39,
+            "accuracy": 46.15
+          },
+          "entailed_closed_conclusion": {
+            "correct": 18,
+            "n": 45,
+            "accuracy": 40.0
+          },
+          "entailed_open_conclusion": {
+            "correct": 1,
+            "n": 27,
+            "accuracy": 3.7
+          },
+          "explicit_negative_closed": {
+            "correct": 27,
+            "n": 36,
+            "accuracy": 75.0
+          },
+          "explicit_negative_open": {
+            "correct": 25,
+            "n": 36,
+            "accuracy": 69.44
+          },
+          "open_derived_missing_antecedent": {
+            "correct": 12,
+            "n": 27,
+            "accuracy": 44.44
+          },
+          "open_missing_direct": {
+            "correct": 30,
+            "n": 36,
+            "accuracy": 83.33
+          },
+          "open_missing_with_closed_distractor": {
+            "correct": 31,
+            "n": 51,
+            "accuracy": 60.78
+          }
+        },
+        "error_types": {
+          "closure_loss_closed_treated_open": 68,
+          "correct": 195,
+          "truth_correct_source_wrong": 60,
+          "truth_correct_closure_wrong": 24,
+          "closure_leakage_open_treated_closed": 12,
+          "truth_wrong": 1
+        }
+      }
+    },
+    {
+      "label": "mistral-small",
+      "model": "mistralai/mistral-small-2603",
+      "repeat": 2,
+      "temperature": 0.0,
+      "report": {
+        "n_dataset": 360,
+        "n_scored": 360,
+        "coverage": 100.0,
+        "models": [
+          "mistralai/mistral-small-2603"
+        ],
+        "api_errors": 0,
+        "parse_failures": 0,
+        "json_valid": {
+          "n": 360,
+          "rate": 100.0
+        },
+        "truth_accuracy": {
+          "correct": 265,
+          "n": 360,
+          "accuracy": 73.61
+        },
+        "source_accuracy": {
+          "correct": 254,
+          "n": 360,
+          "accuracy": 70.56
+        },
+        "closure_accuracy": {
+          "correct": 292,
+          "n": 360,
+          "accuracy": 81.11
+        },
+        "full_accuracy": {
+          "correct": 241,
+          "n": 360,
+          "accuracy": 66.94
+        },
+        "semantic_switch_full_accuracy": {
+          "correct": 40,
+          "n": 120,
+          "accuracy": 33.33
+        },
+        "core_switch_full_accuracy": {
+          "correct": 10,
+          "n": 72,
+          "accuracy": 13.89
+        },
+        "closed_absence_accuracy": {
+          "correct": 15,
+          "n": 106,
+          "accuracy": 14.15
+        },
+        "open_absence_accuracy": {
+          "correct": 102,
+          "n": 110,
+          "accuracy": 92.73
+        },
+        "closure_leakage_rate": {
+          "count": 6,
+          "n": 110,
+          "rate": 5.45
+        },
+        "closure_loss_rate": {
+          "count": 89,
+          "n": 106,
+          "rate": 83.96
+        },
+        "agent_a_source_accuracy": {
+          "correct": 74,
+          "n": 166,
+          "accuracy": 44.58
+        },
+        "agent_b_source_accuracy": {
+          "correct": 165,
+          "n": 179,
+          "accuracy": 92.18
+        },
+        "by_semantics": {
+          "cwa": {
+            "n": 120,
+            "truth": 50.83,
+            "source": 50.0,
+            "closure": 60.0,
+            "full": 45.83
+          },
+          "lcwa": {
+            "n": 120,
+            "truth": 72.5,
+            "source": 70.0,
+            "closure": 88.33,
+            "full": 66.67
+          },
+          "owa": {
+            "n": 120,
+            "truth": 97.5,
+            "source": 91.67,
+            "closure": 95.0,
+            "full": 88.33
+          }
+        },
+        "by_closure_handling": {
+          "closed_absence": {
+            "n": 106,
+            "truth": 14.15,
+            "source": 14.15,
+            "closure": 50.0,
+            "full": 14.15
+          },
+          "entailed_or_explicit": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 83.33,
+            "closure": 95.83,
+            "full": 79.17
+          },
+          "explicit_negative": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 93.06,
+            "full": 93.06
+          },
+          "open_absence": {
+            "n": 110,
+            "truth": 96.36,
+            "source": 97.27,
+            "closure": 93.64,
+            "full": 92.73
+          }
+        },
+        "by_source_used": {
+          "agent_a": {
+            "correct": 74,
+            "n": 166,
+            "accuracy": 44.58
+          },
+          "agent_b": {
+            "correct": 165,
+            "n": 179,
+            "accuracy": 92.18
+          },
+          "both": {
+            "correct": 15,
+            "n": 15,
+            "accuracy": 100.0
+          }
+        },
+        "by_family": {
+          "closed_derived_missing_antecedent": {
+            "correct": 7,
+            "n": 21,
+            "accuracy": 33.33
+          },
+          "closed_missing_direct": {
+            "correct": 16,
+            "n": 42,
+            "accuracy": 38.1
+          },
+          "closed_missing_with_open_distractor": {
+            "correct": 12,
+            "n": 39,
+            "accuracy": 30.77
+          },
+          "entailed_closed_conclusion": {
+            "correct": 36,
+            "n": 45,
+            "accuracy": 80.0
+          },
+          "entailed_open_conclusion": {
+            "correct": 21,
+            "n": 27,
+            "accuracy": 77.78
+          },
+          "explicit_negative_closed": {
+            "correct": 33,
+            "n": 36,
+            "accuracy": 91.67
+          },
+          "explicit_negative_open": {
+            "correct": 34,
+            "n": 36,
+            "accuracy": 94.44
+          },
+          "open_derived_missing_antecedent": {
+            "correct": 20,
+            "n": 27,
+            "accuracy": 74.07
+          },
+          "open_missing_direct": {
+            "correct": 28,
+            "n": 36,
+            "accuracy": 77.78
+          },
+          "open_missing_with_closed_distractor": {
+            "correct": 34,
+            "n": 51,
+            "accuracy": 66.67
+          }
+        },
+        "error_types": {
+          "closure_loss_closed_treated_open": 89,
+          "correct": 241,
+          "closure_leakage_open_treated_closed": 6,
+          "truth_correct_source_wrong": 13,
+          "truth_wrong": 3,
+          "truth_correct_closure_wrong": 8
+        }
+      }
+    },
+    {
+      "label": "mistral-small",
+      "model": "mistralai/mistral-small-2603",
+      "repeat": 3,
+      "temperature": 0.0,
+      "report": {
+        "n_dataset": 360,
+        "n_scored": 360,
+        "coverage": 100.0,
+        "models": [
+          "mistralai/mistral-small-2603"
+        ],
+        "api_errors": 0,
+        "parse_failures": 0,
+        "json_valid": {
+          "n": 360,
+          "rate": 100.0
+        },
+        "truth_accuracy": {
+          "correct": 267,
+          "n": 360,
+          "accuracy": 74.17
+        },
+        "source_accuracy": {
+          "correct": 253,
+          "n": 360,
+          "accuracy": 70.28
+        },
+        "closure_accuracy": {
+          "correct": 290,
+          "n": 360,
+          "accuracy": 80.56
+        },
+        "full_accuracy": {
+          "correct": 231,
+          "n": 360,
+          "accuracy": 64.17
+        },
+        "semantic_switch_full_accuracy": {
+          "correct": 31,
+          "n": 120,
+          "accuracy": 25.83
+        },
+        "core_switch_full_accuracy": {
+          "correct": 8,
+          "n": 72,
+          "accuracy": 11.11
+        },
+        "closed_absence_accuracy": {
+          "correct": 17,
+          "n": 106,
+          "accuracy": 16.04
+        },
+        "open_absence_accuracy": {
+          "correct": 98,
+          "n": 110,
+          "accuracy": 89.09
+        },
+        "closure_leakage_rate": {
+          "count": 10,
+          "n": 110,
+          "rate": 9.09
+        },
+        "closure_loss_rate": {
+          "count": 88,
+          "n": 106,
+          "rate": 83.02
+        },
+        "agent_a_source_accuracy": {
+          "correct": 75,
+          "n": 166,
+          "accuracy": 45.18
+        },
+        "agent_b_source_accuracy": {
+          "correct": 163,
+          "n": 179,
+          "accuracy": 91.06
+        },
+        "by_semantics": {
+          "cwa": {
+            "n": 120,
+            "truth": 50.83,
+            "source": 49.17,
+            "closure": 60.83,
+            "full": 40.83
+          },
+          "lcwa": {
+            "n": 120,
+            "truth": 75.0,
+            "source": 70.83,
+            "closure": 90.0,
+            "full": 69.17
+          },
+          "owa": {
+            "n": 120,
+            "truth": 96.67,
+            "source": 90.83,
+            "closure": 90.83,
+            "full": 82.5
+          }
+        },
+        "by_closure_handling": {
+          "closed_absence": {
+            "n": 106,
+            "truth": 16.04,
+            "source": 16.04,
+            "closure": 55.66,
+            "full": 16.04
+          },
+          "entailed_or_explicit": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 77.78,
+            "closure": 97.22,
+            "full": 75.0
+          },
+          "explicit_negative": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 86.11,
+            "full": 86.11
+          },
+          "open_absence": {
+            "n": 110,
+            "truth": 96.36,
+            "source": 98.18,
+            "closure": 90.0,
+            "full": 89.09
+          }
+        },
+        "by_source_used": {
+          "agent_a": {
+            "correct": 75,
+            "n": 166,
+            "accuracy": 45.18
+          },
+          "agent_b": {
+            "correct": 163,
+            "n": 179,
+            "accuracy": 91.06
+          },
+          "both": {
+            "correct": 15,
+            "n": 15,
+            "accuracy": 100.0
+          }
+        },
+        "by_family": {
+          "closed_derived_missing_antecedent": {
+            "correct": 8,
+            "n": 21,
+            "accuracy": 38.1
+          },
+          "closed_missing_direct": {
+            "correct": 17,
+            "n": 42,
+            "accuracy": 40.48
+          },
+          "closed_missing_with_open_distractor": {
+            "correct": 11,
+            "n": 39,
+            "accuracy": 28.21
+          },
+          "entailed_closed_conclusion": {
+            "correct": 35,
+            "n": 45,
+            "accuracy": 77.78
+          },
+          "entailed_open_conclusion": {
+            "correct": 19,
+            "n": 27,
+            "accuracy": 70.37
+          },
+          "explicit_negative_closed": {
+            "correct": 29,
+            "n": 36,
+            "accuracy": 80.56
+          },
+          "explicit_negative_open": {
+            "correct": 33,
+            "n": 36,
+            "accuracy": 91.67
+          },
+          "open_derived_missing_antecedent": {
+            "correct": 21,
+            "n": 27,
+            "accuracy": 77.78
+          },
+          "open_missing_direct": {
+            "correct": 28,
+            "n": 36,
+            "accuracy": 77.78
+          },
+          "open_missing_with_closed_distractor": {
+            "correct": 30,
+            "n": 51,
+            "accuracy": 58.82
+          }
+        },
+        "error_types": {
+          "closure_loss_closed_treated_open": 88,
+          "correct": 231,
+          "closure_leakage_open_treated_closed": 10,
+          "truth_correct_source_wrong": 17,
+          "truth_correct_closure_wrong": 12,
+          "truth_wrong": 2
+        }
+      }
+    },
+    {
+      "label": "deepseek-flash",
+      "model": "deepseek-v4-flash",
+      "repeat": 1,
+      "temperature": 0.0,
+      "report": {
+        "n_dataset": 360,
+        "n_scored": 360,
+        "coverage": 100.0,
+        "models": [
+          "deepseek-v4-flash"
+        ],
+        "api_errors": 0,
+        "parse_failures": 0,
+        "json_valid": {
+          "n": 102,
+          "rate": 28.33
+        },
+        "truth_accuracy": {
+          "correct": 326,
+          "n": 360,
+          "accuracy": 90.56
+        },
+        "source_accuracy": {
+          "correct": 353,
+          "n": 360,
+          "accuracy": 98.06
+        },
+        "closure_accuracy": {
+          "correct": 342,
+          "n": 360,
+          "accuracy": 95.0
+        },
+        "full_accuracy": {
+          "correct": 326,
+          "n": 360,
+          "accuracy": 90.56
+        },
+        "semantic_switch_full_accuracy": {
+          "correct": 96,
+          "n": 120,
+          "accuracy": 80.0
+        },
+        "core_switch_full_accuracy": {
+          "correct": 48,
+          "n": 72,
+          "accuracy": 66.67
+        },
+        "closed_absence_accuracy": {
+          "correct": 101,
+          "n": 106,
+          "accuracy": 95.28
+        },
+        "open_absence_accuracy": {
+          "correct": 81,
+          "n": 110,
+          "accuracy": 73.64
+        },
+        "closure_leakage_rate": {
+          "count": 29,
+          "n": 110,
+          "rate": 26.36
+        },
+        "closure_loss_rate": {
+          "count": 5,
+          "n": 106,
+          "rate": 4.72
+        },
+        "agent_a_source_accuracy": {
+          "correct": 161,
+          "n": 166,
+          "accuracy": 96.99
+        },
+        "agent_b_source_accuracy": {
+          "correct": 177,
+          "n": 179,
+          "accuracy": 98.88
+        },
+        "by_semantics": {
+          "cwa": {
+            "n": 120,
+            "truth": 96.67,
+            "source": 96.67,
+            "closure": 96.67,
+            "full": 96.67
+          },
+          "lcwa": {
+            "n": 120,
+            "truth": 86.67,
+            "source": 99.17,
+            "closure": 92.5,
+            "full": 86.67
+          },
+          "owa": {
+            "n": 120,
+            "truth": 88.33,
+            "source": 98.33,
+            "closure": 95.83,
+            "full": 88.33
+          }
+        },
+        "by_closure_handling": {
+          "closed_absence": {
+            "n": 106,
+            "truth": 95.28,
+            "source": 95.28,
+            "closure": 95.28,
+            "full": 95.28
+          },
+          "entailed_or_explicit": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 100.0,
+            "full": 100.0
+          },
+          "explicit_negative": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 100.0,
+            "full": 100.0
+          },
+          "open_absence": {
+            "n": 110,
+            "truth": 73.64,
+            "source": 98.18,
+            "closure": 88.18,
+            "full": 73.64
+          }
+        },
+        "by_source_used": {
+          "agent_a": {
+            "correct": 161,
+            "n": 166,
+            "accuracy": 96.99
+          },
+          "agent_b": {
+            "correct": 177,
+            "n": 179,
+            "accuracy": 98.88
+          },
+          "both": {
+            "correct": 15,
+            "n": 15,
+            "accuracy": 100.0
+          }
+        },
+        "by_family": {
+          "closed_derived_missing_antecedent": {
+            "correct": 17,
+            "n": 21,
+            "accuracy": 80.95
+          },
+          "closed_missing_direct": {
+            "correct": 42,
+            "n": 42,
+            "accuracy": 100.0
+          },
+          "closed_missing_with_open_distractor": {
+            "correct": 38,
+            "n": 39,
+            "accuracy": 97.44
+          },
+          "entailed_closed_conclusion": {
+            "correct": 45,
+            "n": 45,
+            "accuracy": 100.0
+          },
+          "entailed_open_conclusion": {
+            "correct": 27,
+            "n": 27,
+            "accuracy": 100.0
+          },
+          "explicit_negative_closed": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "explicit_negative_open": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "open_derived_missing_antecedent": {
+            "correct": 12,
+            "n": 27,
+            "accuracy": 44.44
+          },
+          "open_missing_direct": {
+            "correct": 24,
+            "n": 36,
+            "accuracy": 66.67
+          },
+          "open_missing_with_closed_distractor": {
+            "correct": 49,
+            "n": 51,
+            "accuracy": 96.08
+          }
+        },
+        "error_types": {
+          "correct": 326,
+          "closure_leakage_open_treated_closed": 29,
+          "closure_loss_closed_treated_open": 5
+        }
+      }
+    },
+    {
+      "label": "deepseek-flash",
+      "model": "deepseek-v4-flash",
+      "repeat": 2,
+      "temperature": 0.0,
+      "report": {
+        "n_dataset": 360,
+        "n_scored": 360,
+        "coverage": 100.0,
+        "models": [
+          "deepseek-v4-flash"
+        ],
+        "api_errors": 0,
+        "parse_failures": 17,
+        "json_valid": {
+          "n": 154,
+          "rate": 42.78
+        },
+        "truth_accuracy": {
+          "correct": 321,
+          "n": 360,
+          "accuracy": 89.17
+        },
+        "source_accuracy": {
+          "correct": 352,
+          "n": 360,
+          "accuracy": 97.78
+        },
+        "closure_accuracy": {
+          "correct": 330,
+          "n": 360,
+          "accuracy": 91.67
+        },
+        "full_accuracy": {
+          "correct": 297,
+          "n": 360,
+          "accuracy": 82.5
+        },
+        "semantic_switch_full_accuracy": {
+          "correct": 73,
+          "n": 120,
+          "accuracy": 60.83
+        },
+        "core_switch_full_accuracy": {
+          "correct": 45,
+          "n": 72,
+          "accuracy": 62.5
+        },
+        "closed_absence_accuracy": {
+          "correct": 101,
+          "n": 106,
+          "accuracy": 95.28
+        },
+        "open_absence_accuracy": {
+          "correct": 76,
+          "n": 110,
+          "accuracy": 69.09
+        },
+        "closure_leakage_rate": {
+          "count": 34,
+          "n": 110,
+          "rate": 30.91
+        },
+        "closure_loss_rate": {
+          "count": 5,
+          "n": 106,
+          "rate": 4.72
+        },
+        "agent_a_source_accuracy": {
+          "correct": 161,
+          "n": 166,
+          "accuracy": 96.99
+        },
+        "agent_b_source_accuracy": {
+          "correct": 176,
+          "n": 179,
+          "accuracy": 98.32
+        },
+        "by_semantics": {
+          "cwa": {
+            "n": 120,
+            "truth": 95.83,
+            "source": 95.83,
+            "closure": 95.83,
+            "full": 95.83
+          },
+          "lcwa": {
+            "n": 120,
+            "truth": 88.33,
+            "source": 99.17,
+            "closure": 91.67,
+            "full": 80.83
+          },
+          "owa": {
+            "n": 120,
+            "truth": 83.33,
+            "source": 98.33,
+            "closure": 87.5,
+            "full": 70.83
+          }
+        },
+        "by_closure_handling": {
+          "closed_absence": {
+            "n": 106,
+            "truth": 95.28,
+            "source": 95.28,
+            "closure": 95.28,
+            "full": 95.28
+          },
+          "entailed_or_explicit": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 98.61,
+            "closure": 68.06,
+            "full": 66.67
+          },
+          "explicit_negative": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 100.0,
+            "full": 100.0
+          },
+          "open_absence": {
+            "n": 110,
+            "truth": 69.09,
+            "source": 98.18,
+            "closure": 98.18,
+            "full": 69.09
+          }
+        },
+        "by_source_used": {
+          "agent_a": {
+            "correct": 161,
+            "n": 166,
+            "accuracy": 96.99
+          },
+          "agent_b": {
+            "correct": 176,
+            "n": 179,
+            "accuracy": 98.32
+          },
+          "both": {
+            "correct": 15,
+            "n": 15,
+            "accuracy": 100.0
+          }
+        },
+        "by_family": {
+          "closed_derived_missing_antecedent": {
+            "correct": 16,
+            "n": 21,
+            "accuracy": 76.19
+          },
+          "closed_missing_direct": {
+            "correct": 42,
+            "n": 42,
+            "accuracy": 100.0
+          },
+          "closed_missing_with_open_distractor": {
+            "correct": 38,
+            "n": 39,
+            "accuracy": 97.44
+          },
+          "entailed_closed_conclusion": {
+            "correct": 34,
+            "n": 45,
+            "accuracy": 75.56
+          },
+          "entailed_open_conclusion": {
+            "correct": 14,
+            "n": 27,
+            "accuracy": 51.85
+          },
+          "explicit_negative_closed": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "explicit_negative_open": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "open_derived_missing_antecedent": {
+            "correct": 9,
+            "n": 27,
+            "accuracy": 33.33
+          },
+          "open_missing_direct": {
+            "correct": 25,
+            "n": 36,
+            "accuracy": 69.44
+          },
+          "open_missing_with_closed_distractor": {
+            "correct": 47,
+            "n": 51,
+            "accuracy": 92.16
+          }
+        },
+        "error_types": {
+          "correct": 297,
+          "closure_leakage_open_treated_closed": 34,
+          "truth_correct_closure_wrong": 23,
+          "closure_loss_closed_treated_open": 5,
+          "truth_correct_source_wrong": 1
+        }
+      }
+    },
+    {
+      "label": "deepseek-flash",
+      "model": "deepseek-v4-flash",
+      "repeat": 3,
+      "temperature": 0.0,
+      "report": {
+        "n_dataset": 360,
+        "n_scored": 360,
+        "coverage": 100.0,
+        "models": [
+          "deepseek-v4-flash"
+        ],
+        "api_errors": 0,
+        "parse_failures": 15,
+        "json_valid": {
+          "n": 151,
+          "rate": 41.94
+        },
+        "truth_accuracy": {
+          "correct": 320,
+          "n": 360,
+          "accuracy": 88.89
+        },
+        "source_accuracy": {
+          "correct": 351,
+          "n": 360,
+          "accuracy": 97.5
+        },
+        "closure_accuracy": {
+          "correct": 332,
+          "n": 360,
+          "accuracy": 92.22
+        },
+        "full_accuracy": {
+          "correct": 297,
+          "n": 360,
+          "accuracy": 82.5
+        },
+        "semantic_switch_full_accuracy": {
+          "correct": 73,
+          "n": 120,
+          "accuracy": 60.83
+        },
+        "core_switch_full_accuracy": {
+          "correct": 45,
+          "n": 72,
+          "accuracy": 62.5
+        },
+        "closed_absence_accuracy": {
+          "correct": 101,
+          "n": 106,
+          "accuracy": 95.28
+        },
+        "open_absence_accuracy": {
+          "correct": 75,
+          "n": 110,
+          "accuracy": 68.18
+        },
+        "closure_leakage_rate": {
+          "count": 35,
+          "n": 110,
+          "rate": 31.82
+        },
+        "closure_loss_rate": {
+          "count": 5,
+          "n": 106,
+          "rate": 4.72
+        },
+        "agent_a_source_accuracy": {
+          "correct": 161,
+          "n": 166,
+          "accuracy": 96.99
+        },
+        "agent_b_source_accuracy": {
+          "correct": 175,
+          "n": 179,
+          "accuracy": 97.77
+        },
+        "by_semantics": {
+          "cwa": {
+            "n": 120,
+            "truth": 95.83,
+            "source": 95.83,
+            "closure": 95.83,
+            "full": 95.83
+          },
+          "lcwa": {
+            "n": 120,
+            "truth": 87.5,
+            "source": 99.17,
+            "closure": 91.67,
+            "full": 80.0
+          },
+          "owa": {
+            "n": 120,
+            "truth": 83.33,
+            "source": 97.5,
+            "closure": 89.17,
+            "full": 71.67
+          }
+        },
+        "by_closure_handling": {
+          "closed_absence": {
+            "n": 106,
+            "truth": 95.28,
+            "source": 95.28,
+            "closure": 95.28,
+            "full": 95.28
+          },
+          "entailed_or_explicit": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 97.22,
+            "closure": 70.83,
+            "full": 68.06
+          },
+          "explicit_negative": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 100.0,
+            "full": 100.0
+          },
+          "open_absence": {
+            "n": 110,
+            "truth": 68.18,
+            "source": 98.18,
+            "closure": 98.18,
+            "full": 68.18
+          }
+        },
+        "by_source_used": {
+          "agent_a": {
+            "correct": 161,
+            "n": 166,
+            "accuracy": 96.99
+          },
+          "agent_b": {
+            "correct": 175,
+            "n": 179,
+            "accuracy": 97.77
+          },
+          "both": {
+            "correct": 15,
+            "n": 15,
+            "accuracy": 100.0
+          }
+        },
+        "by_family": {
+          "closed_derived_missing_antecedent": {
+            "correct": 16,
+            "n": 21,
+            "accuracy": 76.19
+          },
+          "closed_missing_direct": {
+            "correct": 42,
+            "n": 42,
+            "accuracy": 100.0
+          },
+          "closed_missing_with_open_distractor": {
+            "correct": 38,
+            "n": 39,
+            "accuracy": 97.44
+          },
+          "entailed_closed_conclusion": {
+            "correct": 34,
+            "n": 45,
+            "accuracy": 75.56
+          },
+          "entailed_open_conclusion": {
+            "correct": 15,
+            "n": 27,
+            "accuracy": 55.56
+          },
+          "explicit_negative_closed": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "explicit_negative_open": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "open_derived_missing_antecedent": {
+            "correct": 9,
+            "n": 27,
+            "accuracy": 33.33
+          },
+          "open_missing_direct": {
+            "correct": 24,
+            "n": 36,
+            "accuracy": 66.67
+          },
+          "open_missing_with_closed_distractor": {
+            "correct": 47,
+            "n": 51,
+            "accuracy": 92.16
+          }
+        },
+        "error_types": {
+          "correct": 297,
+          "closure_leakage_open_treated_closed": 35,
+          "truth_correct_closure_wrong": 21,
+          "closure_loss_closed_treated_open": 5,
+          "truth_correct_source_wrong": 2
+        }
+      }
+    },
+    {
+      "label": "deepseek-pro",
+      "model": "deepseek-v4-pro",
+      "repeat": 1,
+      "temperature": 0.0,
+      "report": {
+        "n_dataset": 360,
+        "n_scored": 360,
+        "coverage": 100.0,
+        "models": [
+          "deepseek-v4-pro"
+        ],
+        "api_errors": 0,
+        "parse_failures": 0,
+        "json_valid": {
+          "n": 359,
+          "rate": 99.72
+        },
+        "truth_accuracy": {
+          "correct": 343,
+          "n": 360,
+          "accuracy": 95.28
+        },
+        "source_accuracy": {
+          "correct": 338,
+          "n": 360,
+          "accuracy": 93.89
+        },
+        "closure_accuracy": {
+          "correct": 339,
+          "n": 360,
+          "accuracy": 94.17
+        },
+        "full_accuracy": {
+          "correct": 327,
+          "n": 360,
+          "accuracy": 90.83
+        },
+        "semantic_switch_full_accuracy": {
+          "correct": 94,
+          "n": 120,
+          "accuracy": 78.33
+        },
+        "core_switch_full_accuracy": {
+          "correct": 55,
+          "n": 72,
+          "accuracy": 76.39
+        },
+        "closed_absence_accuracy": {
+          "correct": 89,
+          "n": 106,
+          "accuracy": 83.96
+        },
+        "open_absence_accuracy": {
+          "correct": 103,
+          "n": 110,
+          "accuracy": 93.64
+        },
+        "closure_leakage_rate": {
+          "count": 0,
+          "n": 110,
+          "rate": 0.0
+        },
+        "closure_loss_rate": {
+          "count": 17,
+          "n": 106,
+          "rate": 16.04
+        },
+        "agent_a_source_accuracy": {
+          "correct": 153,
+          "n": 166,
+          "accuracy": 92.17
+        },
+        "agent_b_source_accuracy": {
+          "correct": 170,
+          "n": 179,
+          "accuracy": 94.97
+        },
+        "by_semantics": {
+          "cwa": {
+            "n": 120,
+            "truth": 87.5,
+            "source": 90.83,
+            "closure": 90.0,
+            "full": 87.5
+          },
+          "lcwa": {
+            "n": 120,
+            "truth": 98.33,
+            "source": 95.0,
+            "closure": 97.5,
+            "full": 94.17
+          },
+          "owa": {
+            "n": 120,
+            "truth": 100.0,
+            "source": 95.83,
+            "closure": 95.0,
+            "full": 90.83
+          }
+        },
+        "by_closure_handling": {
+          "closed_absence": {
+            "n": 106,
+            "truth": 83.96,
+            "source": 87.74,
+            "closure": 86.79,
+            "full": 83.96
+          },
+          "entailed_or_explicit": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 97.22,
+            "closure": 90.28,
+            "full": 87.5
+          },
+          "explicit_negative": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 100.0,
+            "full": 100.0
+          },
+          "open_absence": {
+            "n": 110,
+            "truth": 100.0,
+            "source": 93.64,
+            "closure": 100.0,
+            "full": 93.64
+          }
+        },
+        "by_source_used": {
+          "agent_a": {
+            "correct": 153,
+            "n": 166,
+            "accuracy": 92.17
+          },
+          "agent_b": {
+            "correct": 170,
+            "n": 179,
+            "accuracy": 94.97
+          },
+          "both": {
+            "correct": 15,
+            "n": 15,
+            "accuracy": 100.0
+          }
+        },
+        "by_family": {
+          "closed_derived_missing_antecedent": {
+            "correct": 15,
+            "n": 21,
+            "accuracy": 71.43
+          },
+          "closed_missing_direct": {
+            "correct": 38,
+            "n": 42,
+            "accuracy": 90.48
+          },
+          "closed_missing_with_open_distractor": {
+            "correct": 38,
+            "n": 39,
+            "accuracy": 97.44
+          },
+          "entailed_closed_conclusion": {
+            "correct": 38,
+            "n": 45,
+            "accuracy": 84.44
+          },
+          "entailed_open_conclusion": {
+            "correct": 25,
+            "n": 27,
+            "accuracy": 92.59
+          },
+          "explicit_negative_closed": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "explicit_negative_open": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "open_derived_missing_antecedent": {
+            "correct": 14,
+            "n": 27,
+            "accuracy": 51.85
+          },
+          "open_missing_direct": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "open_missing_with_closed_distractor": {
+            "correct": 51,
+            "n": 51,
+            "accuracy": 100.0
+          }
+        },
+        "error_types": {
+          "correct": 327,
+          "truth_correct_source_wrong": 9,
+          "closure_loss_closed_treated_open": 17,
+          "truth_correct_closure_wrong": 7
+        }
+      }
+    },
+    {
+      "label": "deepseek-pro",
+      "model": "deepseek-v4-pro",
+      "repeat": 2,
+      "temperature": 0.0,
+      "report": {
+        "n_dataset": 360,
+        "n_scored": 360,
+        "coverage": 100.0,
+        "models": [
+          "deepseek-v4-pro"
+        ],
+        "api_errors": 0,
+        "parse_failures": 0,
+        "json_valid": {
+          "n": 360,
+          "rate": 100.0
+        },
+        "truth_accuracy": {
+          "correct": 345,
+          "n": 360,
+          "accuracy": 95.83
+        },
+        "source_accuracy": {
+          "correct": 352,
+          "n": 360,
+          "accuracy": 97.78
+        },
+        "closure_accuracy": {
+          "correct": 351,
+          "n": 360,
+          "accuracy": 97.5
+        },
+        "full_accuracy": {
+          "correct": 344,
+          "n": 360,
+          "accuracy": 95.56
+        },
+        "semantic_switch_full_accuracy": {
+          "correct": 108,
+          "n": 120,
+          "accuracy": 90.0
+        },
+        "core_switch_full_accuracy": {
+          "correct": 61,
+          "n": 72,
+          "accuracy": 84.72
+        },
+        "closed_absence_accuracy": {
+          "correct": 91,
+          "n": 106,
+          "accuracy": 85.85
+        },
+        "open_absence_accuracy": {
+          "correct": 110,
+          "n": 110,
+          "accuracy": 100.0
+        },
+        "closure_leakage_rate": {
+          "count": 0,
+          "n": 110,
+          "rate": 0.0
+        },
+        "closure_loss_rate": {
+          "count": 15,
+          "n": 106,
+          "rate": 14.15
+        },
+        "agent_a_source_accuracy": {
+          "correct": 158,
+          "n": 166,
+          "accuracy": 95.18
+        },
+        "agent_b_source_accuracy": {
+          "correct": 179,
+          "n": 179,
+          "accuracy": 100.0
+        },
+        "by_semantics": {
+          "cwa": {
+            "n": 120,
+            "truth": 90.83,
+            "source": 96.67,
+            "closure": 96.67,
+            "full": 90.83
+          },
+          "lcwa": {
+            "n": 120,
+            "truth": 96.67,
+            "source": 96.67,
+            "closure": 95.83,
+            "full": 95.83
+          },
+          "owa": {
+            "n": 120,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 100.0,
+            "full": 100.0
+          }
+        },
+        "by_closure_handling": {
+          "closed_absence": {
+            "n": 106,
+            "truth": 85.85,
+            "source": 92.45,
+            "closure": 92.45,
+            "full": 85.85
+          },
+          "entailed_or_explicit": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 98.61,
+            "full": 98.61
+          },
+          "explicit_negative": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 100.0,
+            "full": 100.0
+          },
+          "open_absence": {
+            "n": 110,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 100.0,
+            "full": 100.0
+          }
+        },
+        "by_source_used": {
+          "agent_a": {
+            "correct": 158,
+            "n": 166,
+            "accuracy": 95.18
+          },
+          "agent_b": {
+            "correct": 179,
+            "n": 179,
+            "accuracy": 100.0
+          },
+          "both": {
+            "correct": 15,
+            "n": 15,
+            "accuracy": 100.0
+          }
+        },
+        "by_family": {
+          "closed_derived_missing_antecedent": {
+            "correct": 11,
+            "n": 21,
+            "accuracy": 52.38
+          },
+          "closed_missing_direct": {
+            "correct": 41,
+            "n": 42,
+            "accuracy": 97.62
+          },
+          "closed_missing_with_open_distractor": {
+            "correct": 39,
+            "n": 39,
+            "accuracy": 100.0
+          },
+          "entailed_closed_conclusion": {
+            "correct": 45,
+            "n": 45,
+            "accuracy": 100.0
+          },
+          "entailed_open_conclusion": {
+            "correct": 26,
+            "n": 27,
+            "accuracy": 96.3
+          },
+          "explicit_negative_closed": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "explicit_negative_open": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "open_derived_missing_antecedent": {
+            "correct": 23,
+            "n": 27,
+            "accuracy": 85.19
+          },
+          "open_missing_direct": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "open_missing_with_closed_distractor": {
+            "correct": 51,
+            "n": 51,
+            "accuracy": 100.0
+          }
+        },
+        "error_types": {
+          "correct": 344,
+          "closure_loss_closed_treated_open": 15,
+          "truth_correct_closure_wrong": 1
+        }
+      }
+    },
+    {
+      "label": "deepseek-pro",
+      "model": "deepseek-v4-pro",
+      "repeat": 3,
+      "temperature": 0.0,
+      "report": {
+        "n_dataset": 360,
+        "n_scored": 360,
+        "coverage": 100.0,
+        "models": [
+          "deepseek-v4-pro"
+        ],
+        "api_errors": 0,
+        "parse_failures": 0,
+        "json_valid": {
+          "n": 360,
+          "rate": 100.0
+        },
+        "truth_accuracy": {
+          "correct": 346,
+          "n": 360,
+          "accuracy": 96.11
+        },
+        "source_accuracy": {
+          "correct": 351,
+          "n": 360,
+          "accuracy": 97.5
+        },
+        "closure_accuracy": {
+          "correct": 346,
+          "n": 360,
+          "accuracy": 96.11
+        },
+        "full_accuracy": {
+          "correct": 339,
+          "n": 360,
+          "accuracy": 94.17
+        },
+        "semantic_switch_full_accuracy": {
+          "correct": 103,
+          "n": 120,
+          "accuracy": 85.83
+        },
+        "core_switch_full_accuracy": {
+          "correct": 61,
+          "n": 72,
+          "accuracy": 84.72
+        },
+        "closed_absence_accuracy": {
+          "correct": 92,
+          "n": 106,
+          "accuracy": 86.79
+        },
+        "open_absence_accuracy": {
+          "correct": 110,
+          "n": 110,
+          "accuracy": 100.0
+        },
+        "closure_leakage_rate": {
+          "count": 0,
+          "n": 110,
+          "rate": 0.0
+        },
+        "closure_loss_rate": {
+          "count": 14,
+          "n": 106,
+          "rate": 13.21
+        },
+        "agent_a_source_accuracy": {
+          "correct": 158,
+          "n": 166,
+          "accuracy": 95.18
+        },
+        "agent_b_source_accuracy": {
+          "correct": 178,
+          "n": 179,
+          "accuracy": 99.44
+        },
+        "by_semantics": {
+          "cwa": {
+            "n": 120,
+            "truth": 91.67,
+            "source": 96.67,
+            "closure": 96.67,
+            "full": 91.67
+          },
+          "lcwa": {
+            "n": 120,
+            "truth": 96.67,
+            "source": 96.67,
+            "closure": 95.83,
+            "full": 95.83
+          },
+          "owa": {
+            "n": 120,
+            "truth": 100.0,
+            "source": 99.17,
+            "closure": 95.83,
+            "full": 95.0
+          }
+        },
+        "by_closure_handling": {
+          "closed_absence": {
+            "n": 106,
+            "truth": 86.79,
+            "source": 92.45,
+            "closure": 92.45,
+            "full": 86.79
+          },
+          "entailed_or_explicit": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 98.61,
+            "closure": 91.67,
+            "full": 90.28
+          },
+          "explicit_negative": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 100.0,
+            "full": 100.0
+          },
+          "open_absence": {
+            "n": 110,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 100.0,
+            "full": 100.0
+          }
+        },
+        "by_source_used": {
+          "agent_a": {
+            "correct": 158,
+            "n": 166,
+            "accuracy": 95.18
+          },
+          "agent_b": {
+            "correct": 178,
+            "n": 179,
+            "accuracy": 99.44
+          },
+          "both": {
+            "correct": 15,
+            "n": 15,
+            "accuracy": 100.0
+          }
+        },
+        "by_family": {
+          "closed_derived_missing_antecedent": {
+            "correct": 11,
+            "n": 21,
+            "accuracy": 52.38
+          },
+          "closed_missing_direct": {
+            "correct": 42,
+            "n": 42,
+            "accuracy": 100.0
+          },
+          "closed_missing_with_open_distractor": {
+            "correct": 39,
+            "n": 39,
+            "accuracy": 100.0
+          },
+          "entailed_closed_conclusion": {
+            "correct": 40,
+            "n": 45,
+            "accuracy": 88.89
+          },
+          "entailed_open_conclusion": {
+            "correct": 25,
+            "n": 27,
+            "accuracy": 92.59
+          },
+          "explicit_negative_closed": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "explicit_negative_open": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "open_derived_missing_antecedent": {
+            "correct": 23,
+            "n": 27,
+            "accuracy": 85.19
+          },
+          "open_missing_direct": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "open_missing_with_closed_distractor": {
+            "correct": 51,
+            "n": 51,
+            "accuracy": 100.0
+          }
+        },
+        "error_types": {
+          "correct": 339,
+          "closure_loss_closed_treated_open": 14,
+          "truth_correct_closure_wrong": 6,
+          "truth_correct_source_wrong": 1
+        }
+      }
+    },
+    {
+      "label": "mistral-medium",
+      "model": "mistralai/mistral-medium-3-5",
+      "repeat": 1,
+      "temperature": 0.0,
+      "report": {
+        "n_dataset": 360,
+        "n_scored": 360,
+        "coverage": 100.0,
+        "models": [
+          "mistralai/mistral-medium-3-5"
+        ],
+        "api_errors": 0,
+        "parse_failures": 0,
+        "json_valid": {
+          "n": 360,
+          "rate": 100.0
+        },
+        "truth_accuracy": {
+          "correct": 329,
+          "n": 360,
+          "accuracy": 91.39
+        },
+        "source_accuracy": {
+          "correct": 333,
+          "n": 360,
+          "accuracy": 92.5
+        },
+        "closure_accuracy": {
+          "correct": 285,
+          "n": 360,
+          "accuracy": 79.17
+        },
+        "full_accuracy": {
+          "correct": 262,
+          "n": 360,
+          "accuracy": 72.78
+        },
+        "semantic_switch_full_accuracy": {
+          "correct": 48,
+          "n": 120,
+          "accuracy": 40.0
+        },
+        "core_switch_full_accuracy": {
+          "correct": 48,
+          "n": 72,
+          "accuracy": 66.67
+        },
+        "closed_absence_accuracy": {
+          "correct": 75,
+          "n": 106,
+          "accuracy": 70.75
+        },
+        "open_absence_accuracy": {
+          "correct": 110,
+          "n": 110,
+          "accuracy": 100.0
+        },
+        "closure_leakage_rate": {
+          "count": 0,
+          "n": 110,
+          "rate": 0.0
+        },
+        "closure_loss_rate": {
+          "count": 31,
+          "n": 106,
+          "rate": 29.25
+        },
+        "agent_a_source_accuracy": {
+          "correct": 141,
+          "n": 166,
+          "accuracy": 84.94
+        },
+        "agent_b_source_accuracy": {
+          "correct": 177,
+          "n": 179,
+          "accuracy": 98.88
+        },
+        "by_semantics": {
+          "cwa": {
+            "n": 120,
+            "truth": 80.0,
+            "source": 85.0,
+            "closure": 76.67,
+            "full": 61.67
+          },
+          "lcwa": {
+            "n": 120,
+            "truth": 94.17,
+            "source": 94.17,
+            "closure": 80.0,
+            "full": 76.67
+          },
+          "owa": {
+            "n": 120,
+            "truth": 100.0,
+            "source": 98.33,
+            "closure": 80.83,
+            "full": 80.0
+          }
+        },
+        "by_closure_handling": {
+          "closed_absence": {
+            "n": 106,
+            "truth": 70.75,
+            "source": 76.42,
+            "closure": 91.51,
+            "full": 70.75
+          },
+          "entailed_or_explicit": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 97.22,
+            "closure": 55.56,
+            "full": 54.17
+          },
+          "explicit_negative": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 52.78,
+            "full": 52.78
+          },
+          "open_absence": {
+            "n": 110,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 100.0,
+            "full": 100.0
+          }
+        },
+        "by_source_used": {
+          "agent_a": {
+            "correct": 141,
+            "n": 166,
+            "accuracy": 84.94
+          },
+          "agent_b": {
+            "correct": 177,
+            "n": 179,
+            "accuracy": 98.88
+          },
+          "both": {
+            "correct": 15,
+            "n": 15,
+            "accuracy": 100.0
+          }
+        },
+        "by_family": {
+          "closed_derived_missing_antecedent": {
+            "correct": 7,
+            "n": 21,
+            "accuracy": 33.33
+          },
+          "closed_missing_direct": {
+            "correct": 37,
+            "n": 42,
+            "accuracy": 88.1
+          },
+          "closed_missing_with_open_distractor": {
+            "correct": 36,
+            "n": 39,
+            "accuracy": 92.31
+          },
+          "entailed_closed_conclusion": {
+            "correct": 30,
+            "n": 45,
+            "accuracy": 66.67
+          },
+          "entailed_open_conclusion": {
+            "correct": 9,
+            "n": 27,
+            "accuracy": 33.33
+          },
+          "explicit_negative_closed": {
+            "correct": 14,
+            "n": 36,
+            "accuracy": 38.89
+          },
+          "explicit_negative_open": {
+            "correct": 24,
+            "n": 36,
+            "accuracy": 66.67
+          },
+          "open_derived_missing_antecedent": {
+            "correct": 20,
+            "n": 27,
+            "accuracy": 74.07
+          },
+          "open_missing_direct": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "open_missing_with_closed_distractor": {
+            "correct": 49,
+            "n": 51,
+            "accuracy": 96.08
+          }
+        },
+        "error_types": {
+          "correct": 262,
+          "closure_loss_closed_treated_open": 31,
+          "truth_correct_closure_wrong": 65,
+          "truth_correct_source_wrong": 2
+        }
+      }
+    },
+    {
+      "label": "mistral-medium",
+      "model": "mistralai/mistral-medium-3-5",
+      "repeat": 2,
+      "temperature": 0.0,
+      "report": {
+        "n_dataset": 360,
+        "n_scored": 360,
+        "coverage": 100.0,
+        "models": [
+          "mistralai/mistral-medium-3-5"
+        ],
+        "api_errors": 0,
+        "parse_failures": 0,
+        "json_valid": {
+          "n": 360,
+          "rate": 100.0
+        },
+        "truth_accuracy": {
+          "correct": 330,
+          "n": 360,
+          "accuracy": 91.67
+        },
+        "source_accuracy": {
+          "correct": 336,
+          "n": 360,
+          "accuracy": 93.33
+        },
+        "closure_accuracy": {
+          "correct": 315,
+          "n": 360,
+          "accuracy": 87.5
+        },
+        "full_accuracy": {
+          "correct": 297,
+          "n": 360,
+          "accuracy": 82.5
+        },
+        "semantic_switch_full_accuracy": {
+          "correct": 73,
+          "n": 120,
+          "accuracy": 60.83
+        },
+        "core_switch_full_accuracy": {
+          "correct": 49,
+          "n": 72,
+          "accuracy": 68.06
+        },
+        "closed_absence_accuracy": {
+          "correct": 76,
+          "n": 106,
+          "accuracy": 71.7
+        },
+        "open_absence_accuracy": {
+          "correct": 110,
+          "n": 110,
+          "accuracy": 100.0
+        },
+        "closure_leakage_rate": {
+          "count": 0,
+          "n": 110,
+          "rate": 0.0
+        },
+        "closure_loss_rate": {
+          "count": 30,
+          "n": 106,
+          "rate": 28.3
+        },
+        "agent_a_source_accuracy": {
+          "correct": 143,
+          "n": 166,
+          "accuracy": 86.14
+        },
+        "agent_b_source_accuracy": {
+          "correct": 178,
+          "n": 179,
+          "accuracy": 99.44
+        },
+        "by_semantics": {
+          "cwa": {
+            "n": 120,
+            "truth": 80.83,
+            "source": 86.67,
+            "closure": 95.83,
+            "full": 80.83
+          },
+          "lcwa": {
+            "n": 120,
+            "truth": 94.17,
+            "source": 94.17,
+            "closure": 86.67,
+            "full": 86.67
+          },
+          "owa": {
+            "n": 120,
+            "truth": 100.0,
+            "source": 99.17,
+            "closure": 80.0,
+            "full": 80.0
+          }
+        },
+        "by_closure_handling": {
+          "closed_absence": {
+            "n": 106,
+            "truth": 71.7,
+            "source": 78.3,
+            "closure": 88.68,
+            "full": 71.7
+          },
+          "entailed_or_explicit": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 98.61,
+            "closure": 54.17,
+            "full": 54.17
+          },
+          "explicit_negative": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 100.0,
+            "full": 100.0
+          },
+          "open_absence": {
+            "n": 110,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 100.0,
+            "full": 100.0
+          }
+        },
+        "by_source_used": {
+          "agent_a": {
+            "correct": 143,
+            "n": 166,
+            "accuracy": 86.14
+          },
+          "agent_b": {
+            "correct": 178,
+            "n": 179,
+            "accuracy": 99.44
+          },
+          "both": {
+            "correct": 15,
+            "n": 15,
+            "accuracy": 100.0
+          }
+        },
+        "by_family": {
+          "closed_derived_missing_antecedent": {
+            "correct": 7,
+            "n": 21,
+            "accuracy": 33.33
+          },
+          "closed_missing_direct": {
+            "correct": 35,
+            "n": 42,
+            "accuracy": 83.33
+          },
+          "closed_missing_with_open_distractor": {
+            "correct": 35,
+            "n": 39,
+            "accuracy": 89.74
+          },
+          "entailed_closed_conclusion": {
+            "correct": 30,
+            "n": 45,
+            "accuracy": 66.67
+          },
+          "entailed_open_conclusion": {
+            "correct": 9,
+            "n": 27,
+            "accuracy": 33.33
+          },
+          "explicit_negative_closed": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "explicit_negative_open": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "open_derived_missing_antecedent": {
+            "correct": 22,
+            "n": 27,
+            "accuracy": 81.48
+          },
+          "open_missing_direct": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "open_missing_with_closed_distractor": {
+            "correct": 51,
+            "n": 51,
+            "accuracy": 100.0
+          }
+        },
+        "error_types": {
+          "correct": 297,
+          "closure_loss_closed_treated_open": 30,
+          "truth_correct_closure_wrong": 32,
+          "truth_correct_source_wrong": 1
+        }
+      }
+    },
+    {
+      "label": "mistral-medium",
+      "model": "mistralai/mistral-medium-3-5",
+      "repeat": 3,
+      "temperature": 0.0,
+      "report": {
+        "n_dataset": 360,
+        "n_scored": 360,
+        "coverage": 100.0,
+        "models": [
+          "mistralai/mistral-medium-3-5"
+        ],
+        "api_errors": 0,
+        "parse_failures": 0,
+        "json_valid": {
+          "n": 360,
+          "rate": 100.0
+        },
+        "truth_accuracy": {
+          "correct": 330,
+          "n": 360,
+          "accuracy": 91.67
+        },
+        "source_accuracy": {
+          "correct": 337,
+          "n": 360,
+          "accuracy": 93.61
+        },
+        "closure_accuracy": {
+          "correct": 316,
+          "n": 360,
+          "accuracy": 87.78
+        },
+        "full_accuracy": {
+          "correct": 297,
+          "n": 360,
+          "accuracy": 82.5
+        },
+        "semantic_switch_full_accuracy": {
+          "correct": 73,
+          "n": 120,
+          "accuracy": 60.83
+        },
+        "core_switch_full_accuracy": {
+          "correct": 49,
+          "n": 72,
+          "accuracy": 68.06
+        },
+        "closed_absence_accuracy": {
+          "correct": 76,
+          "n": 106,
+          "accuracy": 71.7
+        },
+        "open_absence_accuracy": {
+          "correct": 110,
+          "n": 110,
+          "accuracy": 100.0
+        },
+        "closure_leakage_rate": {
+          "count": 0,
+          "n": 110,
+          "rate": 0.0
+        },
+        "closure_loss_rate": {
+          "count": 30,
+          "n": 106,
+          "rate": 28.3
+        },
+        "agent_a_source_accuracy": {
+          "correct": 143,
+          "n": 166,
+          "accuracy": 86.14
+        },
+        "agent_b_source_accuracy": {
+          "correct": 179,
+          "n": 179,
+          "accuracy": 100.0
+        },
+        "by_semantics": {
+          "cwa": {
+            "n": 120,
+            "truth": 80.83,
+            "source": 86.67,
+            "closure": 95.83,
+            "full": 80.83
+          },
+          "lcwa": {
+            "n": 120,
+            "truth": 94.17,
+            "source": 94.17,
+            "closure": 87.5,
+            "full": 86.67
+          },
+          "owa": {
+            "n": 120,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 80.0,
+            "full": 80.0
+          }
+        },
+        "by_closure_handling": {
+          "closed_absence": {
+            "n": 106,
+            "truth": 71.7,
+            "source": 78.3,
+            "closure": 89.62,
+            "full": 71.7
+          },
+          "entailed_or_explicit": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 54.17,
+            "full": 54.17
+          },
+          "explicit_negative": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 100.0,
+            "full": 100.0
+          },
+          "open_absence": {
+            "n": 110,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 100.0,
+            "full": 100.0
+          }
+        },
+        "by_source_used": {
+          "agent_a": {
+            "correct": 143,
+            "n": 166,
+            "accuracy": 86.14
+          },
+          "agent_b": {
+            "correct": 179,
+            "n": 179,
+            "accuracy": 100.0
+          },
+          "both": {
+            "correct": 15,
+            "n": 15,
+            "accuracy": 100.0
+          }
+        },
+        "by_family": {
+          "closed_derived_missing_antecedent": {
+            "correct": 7,
+            "n": 21,
+            "accuracy": 33.33
+          },
+          "closed_missing_direct": {
+            "correct": 37,
+            "n": 42,
+            "accuracy": 88.1
+          },
+          "closed_missing_with_open_distractor": {
+            "correct": 34,
+            "n": 39,
+            "accuracy": 87.18
+          },
+          "entailed_closed_conclusion": {
+            "correct": 30,
+            "n": 45,
+            "accuracy": 66.67
+          },
+          "entailed_open_conclusion": {
+            "correct": 9,
+            "n": 27,
+            "accuracy": 33.33
+          },
+          "explicit_negative_closed": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "explicit_negative_open": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "open_derived_missing_antecedent": {
+            "correct": 21,
+            "n": 27,
+            "accuracy": 77.78
+          },
+          "open_missing_direct": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "open_missing_with_closed_distractor": {
+            "correct": 51,
+            "n": 51,
+            "accuracy": 100.0
+          }
+        },
+        "error_types": {
+          "correct": 297,
+          "closure_loss_closed_treated_open": 30,
+          "truth_correct_closure_wrong": 33
+        }
+      }
+    },
+    {
+      "label": "llama-scout",
+      "model": "meta-llama/llama-4-scout",
+      "repeat": 1,
+      "temperature": 0.0,
+      "report": {
+        "n_dataset": 360,
+        "n_scored": 360,
+        "coverage": 100.0,
+        "models": [
+          "meta-llama/llama-4-scout"
+        ],
+        "api_errors": 0,
+        "parse_failures": 11,
+        "json_valid": {
+          "n": 104,
+          "rate": 28.89
+        },
+        "truth_accuracy": {
+          "correct": 278,
+          "n": 360,
+          "accuracy": 77.22
+        },
+        "source_accuracy": {
+          "correct": 274,
+          "n": 360,
+          "accuracy": 76.11
+        },
+        "closure_accuracy": {
+          "correct": 267,
+          "n": 360,
+          "accuracy": 74.17
+        },
+        "full_accuracy": {
+          "correct": 248,
+          "n": 360,
+          "accuracy": 68.89
+        },
+        "semantic_switch_full_accuracy": {
+          "correct": 42,
+          "n": 120,
+          "accuracy": 35.0
+        },
+        "core_switch_full_accuracy": {
+          "correct": 10,
+          "n": 72,
+          "accuracy": 13.89
+        },
+        "closed_absence_accuracy": {
+          "correct": 18,
+          "n": 106,
+          "accuracy": 16.98
+        },
+        "open_absence_accuracy": {
+          "correct": 106,
+          "n": 110,
+          "accuracy": 96.36
+        },
+        "closure_leakage_rate": {
+          "count": 1,
+          "n": 110,
+          "rate": 0.91
+        },
+        "closure_loss_rate": {
+          "count": 76,
+          "n": 106,
+          "rate": 71.7
+        },
+        "agent_a_source_accuracy": {
+          "correct": 92,
+          "n": 166,
+          "accuracy": 55.42
+        },
+        "agent_b_source_accuracy": {
+          "correct": 167,
+          "n": 179,
+          "accuracy": 93.3
+        },
+        "by_semantics": {
+          "cwa": {
+            "n": 120,
+            "truth": 58.33,
+            "source": 65.83,
+            "closure": 54.17,
+            "full": 50.0
+          },
+          "lcwa": {
+            "n": 120,
+            "truth": 74.17,
+            "source": 66.67,
+            "closure": 70.0,
+            "full": 62.5
+          },
+          "owa": {
+            "n": 120,
+            "truth": 99.17,
+            "source": 95.83,
+            "closure": 98.33,
+            "full": 94.17
+          }
+        },
+        "by_closure_handling": {
+          "closed_absence": {
+            "n": 106,
+            "truth": 27.36,
+            "source": 33.02,
+            "closure": 23.58,
+            "full": 16.98
+          },
+          "entailed_or_explicit": {
+            "n": 72,
+            "truth": 95.83,
+            "source": 87.5,
+            "closure": 95.83,
+            "full": 84.72
+          },
+          "explicit_negative": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 94.44,
+            "closure": 90.28,
+            "full": 87.5
+          },
+          "open_absence": {
+            "n": 110,
+            "truth": 98.18,
+            "source": 98.18,
+            "closure": 98.18,
+            "full": 96.36
+          }
+        },
+        "by_source_used": {
+          "agent_a": {
+            "correct": 92,
+            "n": 166,
+            "accuracy": 55.42
+          },
+          "agent_b": {
+            "correct": 167,
+            "n": 179,
+            "accuracy": 93.3
+          },
+          "both": {
+            "correct": 15,
+            "n": 15,
+            "accuracy": 100.0
+          }
+        },
+        "by_family": {
+          "closed_derived_missing_antecedent": {
+            "correct": 8,
+            "n": 21,
+            "accuracy": 38.1
+          },
+          "closed_missing_direct": {
+            "correct": 17,
+            "n": 42,
+            "accuracy": 40.48
+          },
+          "closed_missing_with_open_distractor": {
+            "correct": 16,
+            "n": 39,
+            "accuracy": 41.03
+          },
+          "entailed_closed_conclusion": {
+            "correct": 39,
+            "n": 45,
+            "accuracy": 86.67
+          },
+          "entailed_open_conclusion": {
+            "correct": 22,
+            "n": 27,
+            "accuracy": 81.48
+          },
+          "explicit_negative_closed": {
+            "correct": 31,
+            "n": 36,
+            "accuracy": 86.11
+          },
+          "explicit_negative_open": {
+            "correct": 32,
+            "n": 36,
+            "accuracy": 88.89
+          },
+          "open_derived_missing_antecedent": {
+            "correct": 22,
+            "n": 27,
+            "accuracy": 81.48
+          },
+          "open_missing_direct": {
+            "correct": 27,
+            "n": 36,
+            "accuracy": 75.0
+          },
+          "open_missing_with_closed_distractor": {
+            "correct": 34,
+            "n": 51,
+            "accuracy": 66.67
+          }
+        },
+        "error_types": {
+          "closure_loss_closed_treated_open": 76,
+          "truth_correct_source_wrong": 21,
+          "correct": 248,
+          "truth_correct_closure_wrong": 9,
+          "closure_leakage_open_treated_closed": 1,
+          "truth_wrong": 2,
+          "parse_failure": 3
+        }
+      }
+    },
+    {
+      "label": "llama-scout",
+      "model": "meta-llama/llama-4-scout",
+      "repeat": 2,
+      "temperature": 0.0,
+      "report": {
+        "n_dataset": 360,
+        "n_scored": 360,
+        "coverage": 100.0,
+        "models": [
+          "meta-llama/llama-4-scout"
+        ],
+        "api_errors": 0,
+        "parse_failures": 6,
+        "json_valid": {
+          "n": 118,
+          "rate": 32.78
+        },
+        "truth_accuracy": {
+          "correct": 277,
+          "n": 360,
+          "accuracy": 76.94
+        },
+        "source_accuracy": {
+          "correct": 281,
+          "n": 360,
+          "accuracy": 78.06
+        },
+        "closure_accuracy": {
+          "correct": 270,
+          "n": 360,
+          "accuracy": 75.0
+        },
+        "full_accuracy": {
+          "correct": 255,
+          "n": 360,
+          "accuracy": 70.83
+        },
+        "semantic_switch_full_accuracy": {
+          "correct": 42,
+          "n": 120,
+          "accuracy": 35.0
+        },
+        "core_switch_full_accuracy": {
+          "correct": 11,
+          "n": 72,
+          "accuracy": 15.28
+        },
+        "closed_absence_accuracy": {
+          "correct": 24,
+          "n": 106,
+          "accuracy": 22.64
+        },
+        "open_absence_accuracy": {
+          "correct": 104,
+          "n": 110,
+          "accuracy": 94.55
+        },
+        "closure_leakage_rate": {
+          "count": 5,
+          "n": 110,
+          "rate": 4.55
+        },
+        "closure_loss_rate": {
+          "count": 74,
+          "n": 106,
+          "rate": 69.81
+        },
+        "agent_a_source_accuracy": {
+          "correct": 89,
+          "n": 166,
+          "accuracy": 53.61
+        },
+        "agent_b_source_accuracy": {
+          "correct": 177,
+          "n": 179,
+          "accuracy": 98.88
+        },
+        "by_semantics": {
+          "cwa": {
+            "n": 120,
+            "truth": 57.5,
+            "source": 59.17,
+            "closure": 55.83,
+            "full": 48.33
+          },
+          "lcwa": {
+            "n": 120,
+            "truth": 76.67,
+            "source": 75.83,
+            "closure": 75.0,
+            "full": 71.67
+          },
+          "owa": {
+            "n": 120,
+            "truth": 96.67,
+            "source": 99.17,
+            "closure": 94.17,
+            "full": 92.5
+          }
+        },
+        "by_closure_handling": {
+          "closed_absence": {
+            "n": 106,
+            "truth": 28.3,
+            "source": 33.02,
+            "closure": 28.3,
+            "full": 22.64
+          },
+          "entailed_or_explicit": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 90.28,
+            "closure": 93.06,
+            "full": 83.33
+          },
+          "explicit_negative": {
+            "n": 72,
+            "truth": 98.61,
+            "source": 98.61,
+            "closure": 94.44,
+            "full": 93.06
+          },
+          "open_absence": {
+            "n": 110,
+            "truth": 94.55,
+            "source": 100.0,
+            "closure": 95.45,
+            "full": 94.55
+          }
+        },
+        "by_source_used": {
+          "agent_a": {
+            "correct": 89,
+            "n": 166,
+            "accuracy": 53.61
+          },
+          "agent_b": {
+            "correct": 177,
+            "n": 179,
+            "accuracy": 98.88
+          },
+          "both": {
+            "correct": 15,
+            "n": 15,
+            "accuracy": 100.0
+          }
+        },
+        "by_family": {
+          "closed_derived_missing_antecedent": {
+            "correct": 7,
+            "n": 21,
+            "accuracy": 33.33
+          },
+          "closed_missing_direct": {
+            "correct": 20,
+            "n": 42,
+            "accuracy": 47.62
+          },
+          "closed_missing_with_open_distractor": {
+            "correct": 18,
+            "n": 39,
+            "accuracy": 46.15
+          },
+          "entailed_closed_conclusion": {
+            "correct": 38,
+            "n": 45,
+            "accuracy": 84.44
+          },
+          "entailed_open_conclusion": {
+            "correct": 22,
+            "n": 27,
+            "accuracy": 81.48
+          },
+          "explicit_negative_closed": {
+            "correct": 33,
+            "n": 36,
+            "accuracy": 91.67
+          },
+          "explicit_negative_open": {
+            "correct": 34,
+            "n": 36,
+            "accuracy": 94.44
+          },
+          "open_derived_missing_antecedent": {
+            "correct": 21,
+            "n": 27,
+            "accuracy": 77.78
+          },
+          "open_missing_direct": {
+            "correct": 26,
+            "n": 36,
+            "accuracy": 72.22
+          },
+          "open_missing_with_closed_distractor": {
+            "correct": 36,
+            "n": 51,
+            "accuracy": 70.59
+          }
+        },
+        "error_types": {
+          "closure_loss_closed_treated_open": 74,
+          "correct": 255,
+          "closure_leakage_open_treated_closed": 5,
+          "truth_correct_source_wrong": 12,
+          "truth_correct_closure_wrong": 10,
+          "truth_wrong": 4
+        }
+      }
+    },
+    {
+      "label": "llama-scout",
+      "model": "meta-llama/llama-4-scout",
+      "repeat": 3,
+      "temperature": 0.0,
+      "report": {
+        "n_dataset": 360,
+        "n_scored": 360,
+        "coverage": 100.0,
+        "models": [
+          "meta-llama/llama-4-scout"
+        ],
+        "api_errors": 0,
+        "parse_failures": 10,
+        "json_valid": {
+          "n": 121,
+          "rate": 33.61
+        },
+        "truth_accuracy": {
+          "correct": 275,
+          "n": 360,
+          "accuracy": 76.39
+        },
+        "source_accuracy": {
+          "correct": 286,
+          "n": 360,
+          "accuracy": 79.44
+        },
+        "closure_accuracy": {
+          "correct": 266,
+          "n": 360,
+          "accuracy": 73.89
+        },
+        "full_accuracy": {
+          "correct": 260,
+          "n": 360,
+          "accuracy": 72.22
+        },
+        "semantic_switch_full_accuracy": {
+          "correct": 48,
+          "n": 120,
+          "accuracy": 40.0
+        },
+        "core_switch_full_accuracy": {
+          "correct": 11,
+          "n": 72,
+          "accuracy": 15.28
+        },
+        "closed_absence_accuracy": {
+          "correct": 24,
+          "n": 106,
+          "accuracy": 22.64
+        },
+        "open_absence_accuracy": {
+          "correct": 105,
+          "n": 110,
+          "accuracy": 95.45
+        },
+        "closure_leakage_rate": {
+          "count": 3,
+          "n": 110,
+          "rate": 2.73
+        },
+        "closure_loss_rate": {
+          "count": 76,
+          "n": 106,
+          "rate": 71.7
+        },
+        "agent_a_source_accuracy": {
+          "correct": 97,
+          "n": 166,
+          "accuracy": 58.43
+        },
+        "agent_b_source_accuracy": {
+          "correct": 175,
+          "n": 179,
+          "accuracy": 97.77
+        },
+        "by_semantics": {
+          "cwa": {
+            "n": 120,
+            "truth": 53.33,
+            "source": 63.33,
+            "closure": 52.5,
+            "full": 50.83
+          },
+          "lcwa": {
+            "n": 120,
+            "truth": 77.5,
+            "source": 76.67,
+            "closure": 74.17,
+            "full": 72.5
+          },
+          "owa": {
+            "n": 120,
+            "truth": 98.33,
+            "source": 98.33,
+            "closure": 95.0,
+            "full": 93.33
+          }
+        },
+        "by_closure_handling": {
+          "closed_absence": {
+            "n": 106,
+            "truth": 24.53,
+            "source": 36.79,
+            "closure": 22.64,
+            "full": 22.64
+          },
+          "entailed_or_explicit": {
+            "n": 72,
+            "truth": 98.61,
+            "source": 91.67,
+            "closure": 91.67,
+            "full": 84.72
+          },
+          "explicit_negative": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 97.22,
+            "full": 97.22
+          },
+          "open_absence": {
+            "n": 110,
+            "truth": 96.36,
+            "source": 99.09,
+            "closure": 96.36,
+            "full": 95.45
+          }
+        },
+        "by_source_used": {
+          "agent_a": {
+            "correct": 97,
+            "n": 166,
+            "accuracy": 58.43
+          },
+          "agent_b": {
+            "correct": 175,
+            "n": 179,
+            "accuracy": 97.77
+          },
+          "both": {
+            "correct": 14,
+            "n": 15,
+            "accuracy": 93.33
+          }
+        },
+        "by_family": {
+          "closed_derived_missing_antecedent": {
+            "correct": 7,
+            "n": 21,
+            "accuracy": 33.33
+          },
+          "closed_missing_direct": {
+            "correct": 19,
+            "n": 42,
+            "accuracy": 45.24
+          },
+          "closed_missing_with_open_distractor": {
+            "correct": 19,
+            "n": 39,
+            "accuracy": 48.72
+          },
+          "entailed_closed_conclusion": {
+            "correct": 41,
+            "n": 45,
+            "accuracy": 91.11
+          },
+          "entailed_open_conclusion": {
+            "correct": 20,
+            "n": 27,
+            "accuracy": 74.07
+          },
+          "explicit_negative_closed": {
+            "correct": 34,
+            "n": 36,
+            "accuracy": 94.44
+          },
+          "explicit_negative_open": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "open_derived_missing_antecedent": {
+            "correct": 18,
+            "n": 27,
+            "accuracy": 66.67
+          },
+          "open_missing_direct": {
+            "correct": 28,
+            "n": 36,
+            "accuracy": 77.78
+          },
+          "open_missing_with_closed_distractor": {
+            "correct": 38,
+            "n": 51,
+            "accuracy": 74.51
+          }
+        },
+        "error_types": {
+          "closure_loss_closed_treated_open": 76,
+          "correct": 260,
+          "parse_failure": 3,
+          "truth_correct_closure_wrong": 9,
+          "closure_leakage_open_treated_closed": 3,
+          "truth_correct_source_wrong": 6,
+          "truth_wrong": 3
+        }
+      }
+    },
+    {
+      "label": "llama-maverick",
+      "model": "meta-llama/llama-4-maverick",
+      "repeat": 1,
+      "temperature": 0.0,
+      "report": {
+        "n_dataset": 360,
+        "n_scored": 360,
+        "coverage": 100.0,
+        "models": [
+          "meta-llama/llama-4-maverick"
+        ],
+        "api_errors": 0,
+        "parse_failures": 0,
+        "json_valid": {
+          "n": 0,
+          "rate": 0.0
+        },
+        "truth_accuracy": {
+          "correct": 360,
+          "n": 360,
+          "accuracy": 100.0
+        },
+        "source_accuracy": {
+          "correct": 351,
+          "n": 360,
+          "accuracy": 97.5
+        },
+        "closure_accuracy": {
+          "correct": 351,
+          "n": 360,
+          "accuracy": 97.5
+        },
+        "full_accuracy": {
+          "correct": 343,
+          "n": 360,
+          "accuracy": 95.28
+        },
+        "semantic_switch_full_accuracy": {
+          "correct": 106,
+          "n": 120,
+          "accuracy": 88.33
+        },
+        "core_switch_full_accuracy": {
+          "correct": 60,
+          "n": 72,
+          "accuracy": 83.33
+        },
+        "closed_absence_accuracy": {
+          "correct": 93,
+          "n": 106,
+          "accuracy": 87.74
+        },
+        "open_absence_accuracy": {
+          "correct": 108,
+          "n": 110,
+          "accuracy": 98.18
+        },
+        "closure_leakage_rate": {
+          "count": 0,
+          "n": 110,
+          "rate": 0.0
+        },
+        "closure_loss_rate": {
+          "count": 0,
+          "n": 106,
+          "rate": 0.0
+        },
+        "agent_a_source_accuracy": {
+          "correct": 159,
+          "n": 166,
+          "accuracy": 95.78
+        },
+        "agent_b_source_accuracy": {
+          "correct": 177,
+          "n": 179,
+          "accuracy": 98.88
+        },
+        "by_semantics": {
+          "cwa": {
+            "n": 120,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 95.0,
+            "full": 95.0
+          },
+          "lcwa": {
+            "n": 120,
+            "truth": 100.0,
+            "source": 93.33,
+            "closure": 98.33,
+            "full": 92.5
+          },
+          "owa": {
+            "n": 120,
+            "truth": 100.0,
+            "source": 99.17,
+            "closure": 99.17,
+            "full": 98.33
+          }
+        },
+        "by_closure_handling": {
+          "closed_absence": {
+            "n": 106,
+            "truth": 100.0,
+            "source": 93.4,
+            "closure": 93.4,
+            "full": 87.74
+          },
+          "entailed_or_explicit": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 97.22,
+            "full": 97.22
+          },
+          "explicit_negative": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 100.0,
+            "full": 100.0
+          },
+          "open_absence": {
+            "n": 110,
+            "truth": 100.0,
+            "source": 98.18,
+            "closure": 100.0,
+            "full": 98.18
+          }
+        },
+        "by_source_used": {
+          "agent_a": {
+            "correct": 159,
+            "n": 166,
+            "accuracy": 95.78
+          },
+          "agent_b": {
+            "correct": 177,
+            "n": 179,
+            "accuracy": 98.88
+          },
+          "both": {
+            "correct": 15,
+            "n": 15,
+            "accuracy": 100.0
+          }
+        },
+        "by_family": {
+          "closed_derived_missing_antecedent": {
+            "correct": 12,
+            "n": 21,
+            "accuracy": 57.14
+          },
+          "closed_missing_direct": {
+            "correct": 40,
+            "n": 42,
+            "accuracy": 95.24
+          },
+          "closed_missing_with_open_distractor": {
+            "correct": 38,
+            "n": 39,
+            "accuracy": 97.44
+          },
+          "entailed_closed_conclusion": {
+            "correct": 43,
+            "n": 45,
+            "accuracy": 95.56
+          },
+          "entailed_open_conclusion": {
+            "correct": 27,
+            "n": 27,
+            "accuracy": 100.0
+          },
+          "explicit_negative_closed": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "explicit_negative_open": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "open_derived_missing_antecedent": {
+            "correct": 24,
+            "n": 27,
+            "accuracy": 88.89
+          },
+          "open_missing_direct": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "open_missing_with_closed_distractor": {
+            "correct": 51,
+            "n": 51,
+            "accuracy": 100.0
+          }
+        },
+        "error_types": {
+          "correct": 343,
+          "truth_correct_closure_wrong": 8,
+          "truth_correct_source_wrong": 9
+        }
+      }
+    },
+    {
+      "label": "llama-maverick",
+      "model": "meta-llama/llama-4-maverick",
+      "repeat": 2,
+      "temperature": 0.0,
+      "report": {
+        "n_dataset": 360,
+        "n_scored": 360,
+        "coverage": 100.0,
+        "models": [
+          "meta-llama/llama-4-maverick"
+        ],
+        "api_errors": 0,
+        "parse_failures": 0,
+        "json_valid": {
+          "n": 0,
+          "rate": 0.0
+        },
+        "truth_accuracy": {
+          "correct": 354,
+          "n": 360,
+          "accuracy": 98.33
+        },
+        "source_accuracy": {
+          "correct": 348,
+          "n": 360,
+          "accuracy": 96.67
+        },
+        "closure_accuracy": {
+          "correct": 350,
+          "n": 360,
+          "accuracy": 97.22
+        },
+        "full_accuracy": {
+          "correct": 344,
+          "n": 360,
+          "accuracy": 95.56
+        },
+        "semantic_switch_full_accuracy": {
+          "correct": 104,
+          "n": 120,
+          "accuracy": 86.67
+        },
+        "core_switch_full_accuracy": {
+          "correct": 60,
+          "n": 72,
+          "accuracy": 83.33
+        },
+        "closed_absence_accuracy": {
+          "correct": 94,
+          "n": 106,
+          "accuracy": 88.68
+        },
+        "open_absence_accuracy": {
+          "correct": 110,
+          "n": 110,
+          "accuracy": 100.0
+        },
+        "closure_leakage_rate": {
+          "count": 0,
+          "n": 110,
+          "rate": 0.0
+        },
+        "closure_loss_rate": {
+          "count": 6,
+          "n": 106,
+          "rate": 5.66
+        },
+        "agent_a_source_accuracy": {
+          "correct": 154,
+          "n": 166,
+          "accuracy": 92.77
+        },
+        "agent_b_source_accuracy": {
+          "correct": 179,
+          "n": 179,
+          "accuracy": 100.0
+        },
+        "by_semantics": {
+          "cwa": {
+            "n": 120,
+            "truth": 95.0,
+            "source": 95.0,
+            "closure": 93.33,
+            "full": 93.33
+          },
+          "lcwa": {
+            "n": 120,
+            "truth": 100.0,
+            "source": 95.0,
+            "closure": 99.17,
+            "full": 94.17
+          },
+          "owa": {
+            "n": 120,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 99.17,
+            "full": 99.17
+          }
+        },
+        "by_closure_handling": {
+          "closed_absence": {
+            "n": 106,
+            "truth": 94.34,
+            "source": 88.68,
+            "closure": 94.34,
+            "full": 88.68
+          },
+          "entailed_or_explicit": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 97.22,
+            "full": 97.22
+          },
+          "explicit_negative": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 97.22,
+            "full": 97.22
+          },
+          "open_absence": {
+            "n": 110,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 100.0,
+            "full": 100.0
+          }
+        },
+        "by_source_used": {
+          "agent_a": {
+            "correct": 154,
+            "n": 166,
+            "accuracy": 92.77
+          },
+          "agent_b": {
+            "correct": 179,
+            "n": 179,
+            "accuracy": 100.0
+          },
+          "both": {
+            "correct": 15,
+            "n": 15,
+            "accuracy": 100.0
+          }
+        },
+        "by_family": {
+          "closed_derived_missing_antecedent": {
+            "correct": 15,
+            "n": 21,
+            "accuracy": 71.43
+          },
+          "closed_missing_direct": {
+            "correct": 39,
+            "n": 42,
+            "accuracy": 92.86
+          },
+          "closed_missing_with_open_distractor": {
+            "correct": 37,
+            "n": 39,
+            "accuracy": 94.87
+          },
+          "entailed_closed_conclusion": {
+            "correct": 44,
+            "n": 45,
+            "accuracy": 97.78
+          },
+          "entailed_open_conclusion": {
+            "correct": 26,
+            "n": 27,
+            "accuracy": 96.3
+          },
+          "explicit_negative_closed": {
+            "correct": 35,
+            "n": 36,
+            "accuracy": 97.22
+          },
+          "explicit_negative_open": {
+            "correct": 35,
+            "n": 36,
+            "accuracy": 97.22
+          },
+          "open_derived_missing_antecedent": {
+            "correct": 26,
+            "n": 27,
+            "accuracy": 96.3
+          },
+          "open_missing_direct": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "open_missing_with_closed_distractor": {
+            "correct": 51,
+            "n": 51,
+            "accuracy": 100.0
+          }
+        },
+        "error_types": {
+          "correct": 344,
+          "truth_correct_closure_wrong": 4,
+          "closure_loss_closed_treated_open": 6,
+          "truth_correct_source_wrong": 6
+        }
+      }
+    },
+    {
+      "label": "llama-maverick",
+      "model": "meta-llama/llama-4-maverick",
+      "repeat": 3,
+      "temperature": 0.0,
+      "report": {
+        "n_dataset": 360,
+        "n_scored": 360,
+        "coverage": 100.0,
+        "models": [
+          "meta-llama/llama-4-maverick"
+        ],
+        "api_errors": 0,
+        "parse_failures": 0,
+        "json_valid": {
+          "n": 0,
+          "rate": 0.0
+        },
+        "truth_accuracy": {
+          "correct": 355,
+          "n": 360,
+          "accuracy": 98.61
+        },
+        "source_accuracy": {
+          "correct": 350,
+          "n": 360,
+          "accuracy": 97.22
+        },
+        "closure_accuracy": {
+          "correct": 348,
+          "n": 360,
+          "accuracy": 96.67
+        },
+        "full_accuracy": {
+          "correct": 343,
+          "n": 360,
+          "accuracy": 95.28
+        },
+        "semantic_switch_full_accuracy": {
+          "correct": 103,
+          "n": 120,
+          "accuracy": 85.83
+        },
+        "core_switch_full_accuracy": {
+          "correct": 62,
+          "n": 72,
+          "accuracy": 86.11
+        },
+        "closed_absence_accuracy": {
+          "correct": 96,
+          "n": 106,
+          "accuracy": 90.57
+        },
+        "open_absence_accuracy": {
+          "correct": 110,
+          "n": 110,
+          "accuracy": 100.0
+        },
+        "closure_leakage_rate": {
+          "count": 0,
+          "n": 110,
+          "rate": 0.0
+        },
+        "closure_loss_rate": {
+          "count": 5,
+          "n": 106,
+          "rate": 4.72
+        },
+        "agent_a_source_accuracy": {
+          "correct": 156,
+          "n": 166,
+          "accuracy": 93.98
+        },
+        "agent_b_source_accuracy": {
+          "correct": 179,
+          "n": 179,
+          "accuracy": 100.0
+        },
+        "by_semantics": {
+          "cwa": {
+            "n": 120,
+            "truth": 95.83,
+            "source": 95.83,
+            "closure": 93.33,
+            "full": 93.33
+          },
+          "lcwa": {
+            "n": 120,
+            "truth": 100.0,
+            "source": 95.83,
+            "closure": 97.5,
+            "full": 93.33
+          },
+          "owa": {
+            "n": 120,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 99.17,
+            "full": 99.17
+          }
+        },
+        "by_closure_handling": {
+          "closed_absence": {
+            "n": 106,
+            "truth": 95.28,
+            "source": 90.57,
+            "closure": 95.28,
+            "full": 90.57
+          },
+          "entailed_or_explicit": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 95.83,
+            "full": 95.83
+          },
+          "explicit_negative": {
+            "n": 72,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 94.44,
+            "full": 94.44
+          },
+          "open_absence": {
+            "n": 110,
+            "truth": 100.0,
+            "source": 100.0,
+            "closure": 100.0,
+            "full": 100.0
+          }
+        },
+        "by_source_used": {
+          "agent_a": {
+            "correct": 156,
+            "n": 166,
+            "accuracy": 93.98
+          },
+          "agent_b": {
+            "correct": 179,
+            "n": 179,
+            "accuracy": 100.0
+          },
+          "both": {
+            "correct": 15,
+            "n": 15,
+            "accuracy": 100.0
+          }
+        },
+        "by_family": {
+          "closed_derived_missing_antecedent": {
+            "correct": 16,
+            "n": 21,
+            "accuracy": 76.19
+          },
+          "closed_missing_direct": {
+            "correct": 40,
+            "n": 42,
+            "accuracy": 95.24
+          },
+          "closed_missing_with_open_distractor": {
+            "correct": 37,
+            "n": 39,
+            "accuracy": 94.87
+          },
+          "entailed_closed_conclusion": {
+            "correct": 44,
+            "n": 45,
+            "accuracy": 97.78
+          },
+          "entailed_open_conclusion": {
+            "correct": 25,
+            "n": 27,
+            "accuracy": 92.59
+          },
+          "explicit_negative_closed": {
+            "correct": 34,
+            "n": 36,
+            "accuracy": 94.44
+          },
+          "explicit_negative_open": {
+            "correct": 34,
+            "n": 36,
+            "accuracy": 94.44
+          },
+          "open_derived_missing_antecedent": {
+            "correct": 26,
+            "n": 27,
+            "accuracy": 96.3
+          },
+          "open_missing_direct": {
+            "correct": 36,
+            "n": 36,
+            "accuracy": 100.0
+          },
+          "open_missing_with_closed_distractor": {
+            "correct": 51,
+            "n": 51,
+            "accuracy": 100.0
+          }
+        },
+        "error_types": {
+          "correct": 343,
+          "closure_loss_closed_treated_open": 5,
+          "truth_correct_source_wrong": 5,
+          "truth_correct_closure_wrong": 7
+        }
+      }
+    }
+  ]
+}

results/reports/closurebench_replicate_manifest.json

@@ -0,0 +1,116 @@
+{
+  "dataset": "data/base/full.jsonl",
+  "dataset_size": 960,
+  "repeats": 3,
+  "shards": 4,
+  "runs": [
+    {
+      "label": "mistral-small",
+      "model": "mistralai/mistral-small-2603",
+      "repeat": 1,
+      "scored": "results/scored/replicates/closure_contract_v3_mistral-small_repeat_1_scored.jsonl"
+    },
+    {
+      "label": "mistral-small",
+      "model": "mistralai/mistral-small-2603",
+      "repeat": 2,
+      "scored": "results/scored/replicates/closure_contract_v3_mistral-small_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "mistral-small",
+      "model": "mistralai/mistral-small-2603",
+      "repeat": 3,
+      "scored": "results/scored/replicates/closure_contract_v3_mistral-small_repeat_3_scored.jsonl"
+    },
+    {
+      "label": "deepseek-flash",
+      "model": "deepseek-v4-flash",
+      "repeat": 1,
+      "scored": "results/scored/replicates/closure_contract_v3_deepseek-flash_repeat_1_scored.jsonl"
+    },
+    {
+      "label": "deepseek-flash",
+      "model": "deepseek-v4-flash",
+      "repeat": 2,
+      "scored": "results/scored/replicates/closure_contract_v3_deepseek-flash_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "deepseek-flash",
+      "model": "deepseek-v4-flash",
+      "repeat": 3,
+      "scored": "results/scored/replicates/closure_contract_v3_deepseek-flash_repeat_3_scored.jsonl"
+    },
+    {
+      "label": "deepseek-pro",
+      "model": "deepseek-v4-pro",
+      "repeat": 1,
+      "scored": "results/scored/replicates/closure_contract_v3_deepseek-pro_repeat_1_scored.jsonl"
+    },
+    {
+      "label": "deepseek-pro",
+      "model": "deepseek-v4-pro",
+      "repeat": 2,
+      "scored": "results/scored/replicates/closure_contract_v3_deepseek-pro_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "deepseek-pro",
+      "model": "deepseek-v4-pro",
+      "repeat": 3,
+      "scored": "results/scored/replicates/closure_contract_v3_deepseek-pro_repeat_3_scored.jsonl"
+    },
+    {
+      "label": "mistral-medium",
+      "model": "mistralai/mistral-medium-3-5",
+      "repeat": 1,
+      "scored": "results/scored/replicates/closure_contract_v3_mistral-medium_repeat_1_scored.jsonl"
+    },
+    {
+      "label": "mistral-medium",
+      "model": "mistralai/mistral-medium-3-5",
+      "repeat": 2,
+      "scored": "results/scored/replicates/closure_contract_v3_mistral-medium_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "mistral-medium",
+      "model": "mistralai/mistral-medium-3-5",
+      "repeat": 3,
+      "scored": "results/scored/replicates/closure_contract_v3_mistral-medium_repeat_3_scored.jsonl"
+    },
+    {
+      "label": "llama-scout",
+      "model": "meta-llama/llama-4-scout",
+      "repeat": 1,
+      "scored": "results/scored/replicates/closure_contract_v3_llama-scout_repeat_1_scored.jsonl"
+    },
+    {
+      "label": "llama-scout",
+      "model": "meta-llama/llama-4-scout",
+      "repeat": 2,
+      "scored": "results/scored/replicates/closure_contract_v3_llama-scout_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "llama-scout",
+      "model": "meta-llama/llama-4-scout",
+      "repeat": 3,
+      "scored": "results/scored/replicates/closure_contract_v3_llama-scout_repeat_3_scored.jsonl"
+    },
+    {
+      "label": "llama-maverick",
+      "model": "meta-llama/llama-4-maverick",
+      "repeat": 1,
+      "scored": "results/scored/replicates/closure_contract_v3_llama-maverick_repeat_1_scored.jsonl"
+    },
+    {
+      "label": "llama-maverick",
+      "model": "meta-llama/llama-4-maverick",
+      "repeat": 2,
+      "scored": "results/scored/replicates/closure_contract_v3_llama-maverick_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "llama-maverick",
+      "model": "meta-llama/llama-4-maverick",
+      "repeat": 3,
+      "scored": "results/scored/replicates/closure_contract_v3_llama-maverick_repeat_3_scored.jsonl"
+    }
+  ]
+}

results/reports/closurebench_replicate_summary.json

@@ -0,0 +1,1627 @@
+{
+  "dataset": "data/base/full.jsonl",
+  "n_dataset": 960,
+  "manifest": "results/reports/closurebench_replicate_manifest.json",
+  "runs": {
+    "deepseek-flash": {
+      "model": "deepseek-v4-flash",
+      "n_repeats": 3,
+      "replicates": [
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "deepseek-v4-flash"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 263,
+            "n": 320,
+            "accuracy": 82.19
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 135,
+            "n": 192,
+            "accuracy": 70.31
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 51,
+            "n": 96,
+            "accuracy": 53.12
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 96,
+            "n": 96,
+            "accuracy": 100.0
+          },
+          "cwa_accuracy": {
+            "correct": 277,
+            "n": 320,
+            "accuracy": 86.56
+          },
+          "owa_accuracy": {
+            "correct": 320,
+            "n": 320,
+            "accuracy": 100.0
+          },
+          "false_accuracy": {
+            "correct": 392,
+            "n": 480,
+            "accuracy": 81.67
+          },
+          "unknown_accuracy": {
+            "correct": 288,
+            "n": 288,
+            "accuracy": 100.0
+          },
+          "overall_accuracy": {
+            "correct": 872,
+            "n": 960,
+            "accuracy": 90.83
+          },
+          "repeat": 1,
+          "scored": "results/scored/replicates/closure_contract_v3_deepseek-flash_repeat_1_scored.jsonl"
+        },
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "deepseek-v4-flash"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 259,
+            "n": 320,
+            "accuracy": 80.94
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 131,
+            "n": 192,
+            "accuracy": 68.23
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 47,
+            "n": 96,
+            "accuracy": 48.96
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 96,
+            "n": 96,
+            "accuracy": 100.0
+          },
+          "cwa_accuracy": {
+            "correct": 275,
+            "n": 320,
+            "accuracy": 85.94
+          },
+          "owa_accuracy": {
+            "correct": 320,
+            "n": 320,
+            "accuracy": 100.0
+          },
+          "false_accuracy": {
+            "correct": 386,
+            "n": 480,
+            "accuracy": 80.42
+          },
+          "unknown_accuracy": {
+            "correct": 288,
+            "n": 288,
+            "accuracy": 100.0
+          },
+          "overall_accuracy": {
+            "correct": 866,
+            "n": 960,
+            "accuracy": 90.21
+          },
+          "repeat": 2,
+          "scored": "results/scored/replicates/closure_contract_v3_deepseek-flash_repeat_2_scored.jsonl"
+        },
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "deepseek-v4-flash"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 258,
+            "n": 320,
+            "accuracy": 80.62
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 130,
+            "n": 192,
+            "accuracy": 67.71
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 45,
+            "n": 96,
+            "accuracy": 46.88
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 96,
+            "n": 96,
+            "accuracy": 100.0
+          },
+          "cwa_accuracy": {
+            "correct": 274,
+            "n": 320,
+            "accuracy": 85.62
+          },
+          "owa_accuracy": {
+            "correct": 320,
+            "n": 320,
+            "accuracy": 100.0
+          },
+          "false_accuracy": {
+            "correct": 383,
+            "n": 480,
+            "accuracy": 79.79
+          },
+          "unknown_accuracy": {
+            "correct": 288,
+            "n": 288,
+            "accuracy": 100.0
+          },
+          "overall_accuracy": {
+            "correct": 863,
+            "n": 960,
+            "accuracy": 89.9
+          },
+          "repeat": 3,
+          "scored": "results/scored/replicates/closure_contract_v3_deepseek-flash_repeat_3_scored.jsonl"
+        }
+      ],
+      "metrics": {
+        "semantic_switch_accuracy": {
+          "mean": 81.25,
+          "sd": 0.83,
+          "ci95_half_width": 2.06,
+          "min": 80.62,
+          "max": 82.19,
+          "values": [
+            82.19,
+            80.94,
+            80.62
+          ]
+        },
+        "core_contrastive_switch_accuracy": {
+          "mean": 68.75,
+          "sd": 1.38,
+          "ci95_half_width": 3.42,
+          "min": 67.71,
+          "max": 70.31,
+          "values": [
+            70.31,
+            68.23,
+            67.71
+          ]
+        },
+        "lcwa_closed_scope_accuracy": {
+          "mean": 49.65,
+          "sd": 3.18,
+          "ci95_half_width": 7.89,
+          "min": 46.88,
+          "max": 53.12,
+          "values": [
+            53.12,
+            48.96,
+            46.88
+          ]
+        },
+        "lcwa_open_scope_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "cwa_accuracy": {
+          "mean": 86.04,
+          "sd": 0.48,
+          "ci95_half_width": 1.19,
+          "min": 85.62,
+          "max": 86.56,
+          "values": [
+            86.56,
+            85.94,
+            85.62
+          ]
+        },
+        "owa_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "overall_accuracy": {
+          "mean": 90.31,
+          "sd": 0.47,
+          "ci95_half_width": 1.18,
+          "min": 89.9,
+          "max": 90.83,
+          "values": [
+            90.83,
+            90.21,
+            89.9
+          ]
+        }
+      },
+      "item_answer_consistency": {
+        "n": 960,
+        "consistent": 949,
+        "rate": 98.85
+      },
+      "mean_parse_failures": 0.0,
+      "mean_api_errors": 0.0
+    },
+    "deepseek-pro": {
+      "model": "deepseek-v4-pro",
+      "n_repeats": 3,
+      "replicates": [
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "deepseek-v4-pro"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 277,
+            "n": 320,
+            "accuracy": 86.56
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 149,
+            "n": 192,
+            "accuracy": 77.6
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 61,
+            "n": 96,
+            "accuracy": 63.54
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 96,
+            "n": 96,
+            "accuracy": 100.0
+          },
+          "cwa_accuracy": {
+            "correct": 293,
+            "n": 320,
+            "accuracy": 91.56
+          },
+          "owa_accuracy": {
+            "correct": 320,
+            "n": 320,
+            "accuracy": 100.0
+          },
+          "false_accuracy": {
+            "correct": 418,
+            "n": 480,
+            "accuracy": 87.08
+          },
+          "unknown_accuracy": {
+            "correct": 288,
+            "n": 288,
+            "accuracy": 100.0
+          },
+          "overall_accuracy": {
+            "correct": 898,
+            "n": 960,
+            "accuracy": 93.54
+          },
+          "repeat": 1,
+          "scored": "results/scored/replicates/closure_contract_v3_deepseek-pro_repeat_1_scored.jsonl"
+        },
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "deepseek-v4-pro"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 276,
+            "n": 320,
+            "accuracy": 86.25
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 148,
+            "n": 192,
+            "accuracy": 77.08
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 59,
+            "n": 96,
+            "accuracy": 61.46
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 96,
+            "n": 96,
+            "accuracy": 100.0
+          },
+          "cwa_accuracy": {
+            "correct": 294,
+            "n": 320,
+            "accuracy": 91.88
+          },
+          "owa_accuracy": {
+            "correct": 320,
+            "n": 320,
+            "accuracy": 100.0
+          },
+          "false_accuracy": {
+            "correct": 417,
+            "n": 480,
+            "accuracy": 86.88
+          },
+          "unknown_accuracy": {
+            "correct": 288,
+            "n": 288,
+            "accuracy": 100.0
+          },
+          "overall_accuracy": {
+            "correct": 897,
+            "n": 960,
+            "accuracy": 93.44
+          },
+          "repeat": 2,
+          "scored": "results/scored/replicates/closure_contract_v3_deepseek-pro_repeat_2_scored.jsonl"
+        },
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "deepseek-v4-pro"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 277,
+            "n": 320,
+            "accuracy": 86.56
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 149,
+            "n": 192,
+            "accuracy": 77.6
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 59,
+            "n": 96,
+            "accuracy": 61.46
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 96,
+            "n": 96,
+            "accuracy": 100.0
+          },
+          "cwa_accuracy": {
+            "correct": 298,
+            "n": 320,
+            "accuracy": 93.12
+          },
+          "owa_accuracy": {
+            "correct": 320,
+            "n": 320,
+            "accuracy": 100.0
+          },
+          "false_accuracy": {
+            "correct": 421,
+            "n": 480,
+            "accuracy": 87.71
+          },
+          "unknown_accuracy": {
+            "correct": 288,
+            "n": 288,
+            "accuracy": 100.0
+          },
+          "overall_accuracy": {
+            "correct": 901,
+            "n": 960,
+            "accuracy": 93.85
+          },
+          "repeat": 3,
+          "scored": "results/scored/replicates/closure_contract_v3_deepseek-pro_repeat_3_scored.jsonl"
+        }
+      ],
+      "metrics": {
+        "semantic_switch_accuracy": {
+          "mean": 86.46,
+          "sd": 0.18,
+          "ci95_half_width": 0.44,
+          "min": 86.25,
+          "max": 86.56,
+          "values": [
+            86.56,
+            86.25,
+            86.56
+          ]
+        },
+        "core_contrastive_switch_accuracy": {
+          "mean": 77.43,
+          "sd": 0.3,
+          "ci95_half_width": 0.75,
+          "min": 77.08,
+          "max": 77.6,
+          "values": [
+            77.6,
+            77.08,
+            77.6
+          ]
+        },
+        "lcwa_closed_scope_accuracy": {
+          "mean": 62.15,
+          "sd": 1.2,
+          "ci95_half_width": 2.98,
+          "min": 61.46,
+          "max": 63.54,
+          "values": [
+            63.54,
+            61.46,
+            61.46
+          ]
+        },
+        "lcwa_open_scope_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "cwa_accuracy": {
+          "mean": 92.19,
+          "sd": 0.82,
+          "ci95_half_width": 2.05,
+          "min": 91.56,
+          "max": 93.12,
+          "values": [
+            91.56,
+            91.88,
+            93.12
+          ]
+        },
+        "owa_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "overall_accuracy": {
+          "mean": 93.61,
+          "sd": 0.21,
+          "ci95_half_width": 0.53,
+          "min": 93.44,
+          "max": 93.85,
+          "values": [
+            93.54,
+            93.44,
+            93.85
+          ]
+        }
+      },
+      "item_answer_consistency": {
+        "n": 960,
+        "consistent": 946,
+        "rate": 98.54
+      },
+      "mean_parse_failures": 0.0,
+      "mean_api_errors": 0.0
+    },
+    "llama-maverick": {
+      "model": "meta-llama/llama-4-maverick",
+      "n_repeats": 3,
+      "replicates": [
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "meta-llama/llama-4-maverick"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 318,
+            "n": 320,
+            "accuracy": 99.38
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 190,
+            "n": 192,
+            "accuracy": 98.96
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 95,
+            "n": 96,
+            "accuracy": 98.96
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 96,
+            "n": 96,
+            "accuracy": 100.0
+          },
+          "cwa_accuracy": {
+            "correct": 319,
+            "n": 320,
+            "accuracy": 99.69
+          },
+          "owa_accuracy": {
+            "correct": 320,
+            "n": 320,
+            "accuracy": 100.0
+          },
+          "false_accuracy": {
+            "correct": 478,
+            "n": 480,
+            "accuracy": 99.58
+          },
+          "unknown_accuracy": {
+            "correct": 288,
+            "n": 288,
+            "accuracy": 100.0
+          },
+          "overall_accuracy": {
+            "correct": 958,
+            "n": 960,
+            "accuracy": 99.79
+          },
+          "repeat": 1,
+          "scored": "results/scored/replicates/closure_contract_v3_llama-maverick_repeat_1_scored.jsonl"
+        },
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "meta-llama/llama-4-maverick"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 319,
+            "n": 320,
+            "accuracy": 99.69
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 191,
+            "n": 192,
+            "accuracy": 99.48
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 96,
+            "n": 96,
+            "accuracy": 100.0
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 96,
+            "n": 96,
+            "accuracy": 100.0
+          },
+          "cwa_accuracy": {
+            "correct": 319,
+            "n": 320,
+            "accuracy": 99.69
+          },
+          "owa_accuracy": {
+            "correct": 320,
+            "n": 320,
+            "accuracy": 100.0
+          },
+          "false_accuracy": {
+            "correct": 479,
+            "n": 480,
+            "accuracy": 99.79
+          },
+          "unknown_accuracy": {
+            "correct": 288,
+            "n": 288,
+            "accuracy": 100.0
+          },
+          "overall_accuracy": {
+            "correct": 959,
+            "n": 960,
+            "accuracy": 99.9
+          },
+          "repeat": 2,
+          "scored": "results/scored/replicates/closure_contract_v3_llama-maverick_repeat_2_scored.jsonl"
+        },
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "meta-llama/llama-4-maverick"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 319,
+            "n": 320,
+            "accuracy": 99.69
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 191,
+            "n": 192,
+            "accuracy": 99.48
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 96,
+            "n": 96,
+            "accuracy": 100.0
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 96,
+            "n": 96,
+            "accuracy": 100.0
+          },
+          "cwa_accuracy": {
+            "correct": 319,
+            "n": 320,
+            "accuracy": 99.69
+          },
+          "owa_accuracy": {
+            "correct": 320,
+            "n": 320,
+            "accuracy": 100.0
+          },
+          "false_accuracy": {
+            "correct": 479,
+            "n": 480,
+            "accuracy": 99.79
+          },
+          "unknown_accuracy": {
+            "correct": 288,
+            "n": 288,
+            "accuracy": 100.0
+          },
+          "overall_accuracy": {
+            "correct": 959,
+            "n": 960,
+            "accuracy": 99.9
+          },
+          "repeat": 3,
+          "scored": "results/scored/replicates/closure_contract_v3_llama-maverick_repeat_3_scored.jsonl"
+        }
+      ],
+      "metrics": {
+        "semantic_switch_accuracy": {
+          "mean": 99.59,
+          "sd": 0.18,
+          "ci95_half_width": 0.44,
+          "min": 99.38,
+          "max": 99.69,
+          "values": [
+            99.38,
+            99.69,
+            99.69
+          ]
+        },
+        "core_contrastive_switch_accuracy": {
+          "mean": 99.31,
+          "sd": 0.3,
+          "ci95_half_width": 0.75,
+          "min": 98.96,
+          "max": 99.48,
+          "values": [
+            98.96,
+            99.48,
+            99.48
+          ]
+        },
+        "lcwa_closed_scope_accuracy": {
+          "mean": 99.65,
+          "sd": 0.6,
+          "ci95_half_width": 1.49,
+          "min": 98.96,
+          "max": 100.0,
+          "values": [
+            98.96,
+            100.0,
+            100.0
+          ]
+        },
+        "lcwa_open_scope_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "cwa_accuracy": {
+          "mean": 99.69,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 99.69,
+          "max": 99.69,
+          "values": [
+            99.69,
+            99.69,
+            99.69
+          ]
+        },
+        "owa_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "overall_accuracy": {
+          "mean": 99.86,
+          "sd": 0.06,
+          "ci95_half_width": 0.16,
+          "min": 99.79,
+          "max": 99.9,
+          "values": [
+            99.79,
+            99.9,
+            99.9
+          ]
+        }
+      },
+      "item_answer_consistency": {
+        "n": 960,
+        "consistent": 959,
+        "rate": 99.9
+      },
+      "mean_parse_failures": 0.0,
+      "mean_api_errors": 0.0
+    },
+    "llama-scout": {
+      "model": "meta-llama/llama-4-scout",
+      "n_repeats": 3,
+      "replicates": [
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "meta-llama/llama-4-scout"
+          ],
+          "api_errors": 0,
+          "parse_failures": 1,
+          "semantic_switch_accuracy": {
+            "correct": 162,
+            "n": 320,
+            "accuracy": 50.62
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 35,
+            "n": 192,
+            "accuracy": 18.23
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 4,
+            "n": 96,
+            "accuracy": 4.17
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 52,
+            "n": 96,
+            "accuracy": 54.17
+          },
+          "cwa_accuracy": {
+            "correct": 223,
+            "n": 320,
+            "accuracy": 69.69
+          },
+          "owa_accuracy": {
+            "correct": 281,
+            "n": 320,
+            "accuracy": 87.81
+          },
+          "false_accuracy": {
+            "correct": 290,
+            "n": 480,
+            "accuracy": 60.42
+          },
+          "unknown_accuracy": {
+            "correct": 206,
+            "n": 288,
+            "accuracy": 71.53
+          },
+          "overall_accuracy": {
+            "correct": 688,
+            "n": 960,
+            "accuracy": 71.67
+          },
+          "repeat": 1,
+          "scored": "results/scored/replicates/closure_contract_v3_llama-scout_repeat_1_scored.jsonl"
+        },
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "meta-llama/llama-4-scout"
+          ],
+          "api_errors": 0,
+          "parse_failures": 1,
+          "semantic_switch_accuracy": {
+            "correct": 164,
+            "n": 320,
+            "accuracy": 51.25
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 37,
+            "n": 192,
+            "accuracy": 19.27
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 6,
+            "n": 96,
+            "accuracy": 6.25
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 52,
+            "n": 96,
+            "accuracy": 54.17
+          },
+          "cwa_accuracy": {
+            "correct": 220,
+            "n": 320,
+            "accuracy": 68.75
+          },
+          "owa_accuracy": {
+            "correct": 278,
+            "n": 320,
+            "accuracy": 86.88
+          },
+          "false_accuracy": {
+            "correct": 290,
+            "n": 480,
+            "accuracy": 60.42
+          },
+          "unknown_accuracy": {
+            "correct": 202,
+            "n": 288,
+            "accuracy": 70.14
+          },
+          "overall_accuracy": {
+            "correct": 684,
+            "n": 960,
+            "accuracy": 71.25
+          },
+          "repeat": 2,
+          "scored": "results/scored/replicates/closure_contract_v3_llama-scout_repeat_2_scored.jsonl"
+        },
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "meta-llama/llama-4-scout"
+          ],
+          "api_errors": 0,
+          "parse_failures": 2,
+          "semantic_switch_accuracy": {
+            "correct": 161,
+            "n": 320,
+            "accuracy": 50.31
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 33,
+            "n": 192,
+            "accuracy": 17.19
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 3,
+            "n": 96,
+            "accuracy": 3.12
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 47,
+            "n": 96,
+            "accuracy": 48.96
+          },
+          "cwa_accuracy": {
+            "correct": 222,
+            "n": 320,
+            "accuracy": 69.38
+          },
+          "owa_accuracy": {
+            "correct": 278,
+            "n": 320,
+            "accuracy": 86.88
+          },
+          "false_accuracy": {
+            "correct": 289,
+            "n": 480,
+            "accuracy": 60.21
+          },
+          "unknown_accuracy": {
+            "correct": 197,
+            "n": 288,
+            "accuracy": 68.4
+          },
+          "overall_accuracy": {
+            "correct": 678,
+            "n": 960,
+            "accuracy": 70.62
+          },
+          "repeat": 3,
+          "scored": "results/scored/replicates/closure_contract_v3_llama-scout_repeat_3_scored.jsonl"
+        }
+      ],
+      "metrics": {
+        "semantic_switch_accuracy": {
+          "mean": 50.73,
+          "sd": 0.48,
+          "ci95_half_width": 1.19,
+          "min": 50.31,
+          "max": 51.25,
+          "values": [
+            50.62,
+            51.25,
+            50.31
+          ]
+        },
+        "core_contrastive_switch_accuracy": {
+          "mean": 18.23,
+          "sd": 1.04,
+          "ci95_half_width": 2.58,
+          "min": 17.19,
+          "max": 19.27,
+          "values": [
+            18.23,
+            19.27,
+            17.19
+          ]
+        },
+        "lcwa_closed_scope_accuracy": {
+          "mean": 4.51,
+          "sd": 1.59,
+          "ci95_half_width": 3.96,
+          "min": 3.12,
+          "max": 6.25,
+          "values": [
+            4.17,
+            6.25,
+            3.12
+          ]
+        },
+        "lcwa_open_scope_accuracy": {
+          "mean": 52.43,
+          "sd": 3.01,
+          "ci95_half_width": 7.47,
+          "min": 48.96,
+          "max": 54.17,
+          "values": [
+            54.17,
+            54.17,
+            48.96
+          ]
+        },
+        "cwa_accuracy": {
+          "mean": 69.27,
+          "sd": 0.48,
+          "ci95_half_width": 1.19,
+          "min": 68.75,
+          "max": 69.69,
+          "values": [
+            69.69,
+            68.75,
+            69.38
+          ]
+        },
+        "owa_accuracy": {
+          "mean": 87.19,
+          "sd": 0.54,
+          "ci95_half_width": 1.33,
+          "min": 86.88,
+          "max": 87.81,
+          "values": [
+            87.81,
+            86.88,
+            86.88
+          ]
+        },
+        "overall_accuracy": {
+          "mean": 71.18,
+          "sd": 0.53,
+          "ci95_half_width": 1.31,
+          "min": 70.62,
+          "max": 71.67,
+          "values": [
+            71.67,
+            71.25,
+            70.62
+          ]
+        }
+      },
+      "item_answer_consistency": {
+        "n": 960,
+        "consistent": 895,
+        "rate": 93.23
+      },
+      "mean_parse_failures": 1.33,
+      "mean_api_errors": 0.0
+    },
+    "mistral-medium": {
+      "model": "mistralai/mistral-medium-3-5",
+      "n_repeats": 3,
+      "replicates": [
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "mistralai/mistral-medium-3-5"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 281,
+            "n": 320,
+            "accuracy": 87.81
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 153,
+            "n": 192,
+            "accuracy": 79.69
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 57,
+            "n": 96,
+            "accuracy": 59.38
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 96,
+            "n": 96,
+            "accuracy": 100.0
+          },
+          "cwa_accuracy": {
+            "correct": 320,
+            "n": 320,
+            "accuracy": 100.0
+          },
+          "owa_accuracy": {
+            "correct": 320,
+            "n": 320,
+            "accuracy": 100.0
+          },
+          "false_accuracy": {
+            "correct": 441,
+            "n": 480,
+            "accuracy": 91.88
+          },
+          "unknown_accuracy": {
+            "correct": 288,
+            "n": 288,
+            "accuracy": 100.0
+          },
+          "overall_accuracy": {
+            "correct": 921,
+            "n": 960,
+            "accuracy": 95.94
+          },
+          "repeat": 1,
+          "scored": "results/scored/replicates/closure_contract_v3_mistral-medium_repeat_1_scored.jsonl"
+        },
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "mistralai/mistral-medium-3-5"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 278,
+            "n": 320,
+            "accuracy": 86.88
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 150,
+            "n": 192,
+            "accuracy": 78.12
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 54,
+            "n": 96,
+            "accuracy": 56.25
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 96,
+            "n": 96,
+            "accuracy": 100.0
+          },
+          "cwa_accuracy": {
+            "correct": 320,
+            "n": 320,
+            "accuracy": 100.0
+          },
+          "owa_accuracy": {
+            "correct": 320,
+            "n": 320,
+            "accuracy": 100.0
+          },
+          "false_accuracy": {
+            "correct": 438,
+            "n": 480,
+            "accuracy": 91.25
+          },
+          "unknown_accuracy": {
+            "correct": 288,
+            "n": 288,
+            "accuracy": 100.0
+          },
+          "overall_accuracy": {
+            "correct": 918,
+            "n": 960,
+            "accuracy": 95.62
+          },
+          "repeat": 2,
+          "scored": "results/scored/replicates/closure_contract_v3_mistral-medium_repeat_2_scored.jsonl"
+        },
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "mistralai/mistral-medium-3-5"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 277,
+            "n": 320,
+            "accuracy": 86.56
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 149,
+            "n": 192,
+            "accuracy": 77.6
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 53,
+            "n": 96,
+            "accuracy": 55.21
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 96,
+            "n": 96,
+            "accuracy": 100.0
+          },
+          "cwa_accuracy": {
+            "correct": 317,
+            "n": 320,
+            "accuracy": 99.06
+          },
+          "owa_accuracy": {
+            "correct": 320,
+            "n": 320,
+            "accuracy": 100.0
+          },
+          "false_accuracy": {
+            "correct": 434,
+            "n": 480,
+            "accuracy": 90.42
+          },
+          "unknown_accuracy": {
+            "correct": 288,
+            "n": 288,
+            "accuracy": 100.0
+          },
+          "overall_accuracy": {
+            "correct": 914,
+            "n": 960,
+            "accuracy": 95.21
+          },
+          "repeat": 3,
+          "scored": "results/scored/replicates/closure_contract_v3_mistral-medium_repeat_3_scored.jsonl"
+        }
+      ],
+      "metrics": {
+        "semantic_switch_accuracy": {
+          "mean": 87.08,
+          "sd": 0.65,
+          "ci95_half_width": 1.61,
+          "min": 86.56,
+          "max": 87.81,
+          "values": [
+            87.81,
+            86.88,
+            86.56
+          ]
+        },
+        "core_contrastive_switch_accuracy": {
+          "mean": 78.47,
+          "sd": 1.09,
+          "ci95_half_width": 2.7,
+          "min": 77.6,
+          "max": 79.69,
+          "values": [
+            79.69,
+            78.12,
+            77.6
+          ]
+        },
+        "lcwa_closed_scope_accuracy": {
+          "mean": 56.95,
+          "sd": 2.17,
+          "ci95_half_width": 5.39,
+          "min": 55.21,
+          "max": 59.38,
+          "values": [
+            59.38,
+            56.25,
+            55.21
+          ]
+        },
+        "lcwa_open_scope_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "cwa_accuracy": {
+          "mean": 99.69,
+          "sd": 0.54,
+          "ci95_half_width": 1.35,
+          "min": 99.06,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            99.06
+          ]
+        },
+        "owa_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "overall_accuracy": {
+          "mean": 95.59,
+          "sd": 0.37,
+          "ci95_half_width": 0.91,
+          "min": 95.21,
+          "max": 95.94,
+          "values": [
+            95.94,
+            95.62,
+            95.21
+          ]
+        }
+      },
+      "item_answer_consistency": {
+        "n": 960,
+        "consistent": 936,
+        "rate": 97.5
+      },
+      "mean_parse_failures": 0.0,
+      "mean_api_errors": 0.0
+    },
+    "mistral-small": {
+      "model": "mistralai/mistral-small-2603",
+      "n_repeats": 3,
+      "replicates": [
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "mistralai/mistral-small-2603"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 176,
+            "n": 320,
+            "accuracy": 55.0
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 50,
+            "n": 192,
+            "accuracy": 26.04
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 29,
+            "n": 96,
+            "accuracy": 30.21
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 29,
+            "n": 96,
+            "accuracy": 30.21
+          },
+          "cwa_accuracy": {
+            "correct": 293,
+            "n": 320,
+            "accuracy": 91.56
+          },
+          "owa_accuracy": {
+            "correct": 288,
+            "n": 320,
+            "accuracy": 90.0
+          },
+          "false_accuracy": {
+            "correct": 384,
+            "n": 480,
+            "accuracy": 80.0
+          },
+          "unknown_accuracy": {
+            "correct": 191,
+            "n": 288,
+            "accuracy": 66.32
+          },
+          "overall_accuracy": {
+            "correct": 767,
+            "n": 960,
+            "accuracy": 79.9
+          },
+          "repeat": 1,
+          "scored": "results/scored/replicates/closure_contract_v3_mistral-small_repeat_1_scored.jsonl"
+        },
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "mistralai/mistral-small-2603"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 172,
+            "n": 320,
+            "accuracy": 53.75
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 48,
+            "n": 192,
+            "accuracy": 25.0
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 27,
+            "n": 96,
+            "accuracy": 28.12
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 29,
+            "n": 96,
+            "accuracy": 30.21
+          },
+          "cwa_accuracy": {
+            "correct": 293,
+            "n": 320,
+            "accuracy": 91.56
+          },
+          "owa_accuracy": {
+            "correct": 288,
+            "n": 320,
+            "accuracy": 90.0
+          },
+          "false_accuracy": {
+            "correct": 380,
+            "n": 480,
+            "accuracy": 79.17
+          },
+          "unknown_accuracy": {
+            "correct": 193,
+            "n": 288,
+            "accuracy": 67.01
+          },
+          "overall_accuracy": {
+            "correct": 765,
+            "n": 960,
+            "accuracy": 79.69
+          },
+          "repeat": 2,
+          "scored": "results/scored/replicates/closure_contract_v3_mistral-small_repeat_2_scored.jsonl"
+        },
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "mistralai/mistral-small-2603"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 187,
+            "n": 320,
+            "accuracy": 58.44
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 61,
+            "n": 192,
+            "accuracy": 31.77
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 35,
+            "n": 96,
+            "accuracy": 36.46
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 32,
+            "n": 96,
+            "accuracy": 33.33
+          },
+          "cwa_accuracy": {
+            "correct": 296,
+            "n": 320,
+            "accuracy": 92.5
+          },
+          "owa_accuracy": {
+            "correct": 292,
+            "n": 320,
+            "accuracy": 91.25
+          },
+          "false_accuracy": {
+            "correct": 393,
+            "n": 480,
+            "accuracy": 81.88
+          },
+          "unknown_accuracy": {
+            "correct": 198,
+            "n": 288,
+            "accuracy": 68.75
+          },
+          "overall_accuracy": {
+            "correct": 783,
+            "n": 960,
+            "accuracy": 81.56
+          },
+          "repeat": 3,
+          "scored": "results/scored/replicates/closure_contract_v3_mistral-small_repeat_3_scored.jsonl"
+        }
+      ],
+      "metrics": {
+        "semantic_switch_accuracy": {
+          "mean": 55.73,
+          "sd": 2.43,
+          "ci95_half_width": 6.03,
+          "min": 53.75,
+          "max": 58.44,
+          "values": [
+            55.0,
+            53.75,
+            58.44
+          ]
+        },
+        "core_contrastive_switch_accuracy": {
+          "mean": 27.6,
+          "sd": 3.65,
+          "ci95_half_width": 9.06,
+          "min": 25.0,
+          "max": 31.77,
+          "values": [
+            26.04,
+            25.0,
+            31.77
+          ]
+        },
+        "lcwa_closed_scope_accuracy": {
+          "mean": 31.6,
+          "sd": 4.34,
+          "ci95_half_width": 10.78,
+          "min": 28.12,
+          "max": 36.46,
+          "values": [
+            30.21,
+            28.12,
+            36.46
+          ]
+        },
+        "lcwa_open_scope_accuracy": {
+          "mean": 31.25,
+          "sd": 1.8,
+          "ci95_half_width": 4.48,
+          "min": 30.21,
+          "max": 33.33,
+          "values": [
+            30.21,
+            30.21,
+            33.33
+          ]
+        },
+        "cwa_accuracy": {
+          "mean": 91.87,
+          "sd": 0.54,
+          "ci95_half_width": 1.35,
+          "min": 91.56,
+          "max": 92.5,
+          "values": [
+            91.56,
+            91.56,
+            92.5
+          ]
+        },
+        "owa_accuracy": {
+          "mean": 90.42,
+          "sd": 0.72,
+          "ci95_half_width": 1.79,
+          "min": 90.0,
+          "max": 91.25,
+          "values": [
+            90.0,
+            90.0,
+            91.25
+          ]
+        },
+        "overall_accuracy": {
+          "mean": 80.38,
+          "sd": 1.02,
+          "ci95_half_width": 2.55,
+          "min": 79.69,
+          "max": 81.56,
+          "values": [
+            79.9,
+            79.69,
+            81.56
+          ]
+        }
+      },
+      "item_answer_consistency": {
+        "n": 960,
+        "consistent": 877,
+        "rate": 91.35
+      },
+      "mean_parse_failures": 0.0,
+      "mean_api_errors": 0.0
+    }
+  }
+}

results/reports/closurebench_temp1_replicate_manifest.json

@@ -0,0 +1,136 @@
+{
+  "dataset": "data/base/full.jsonl",
+  "dataset_size": 960,
+  "repeats": 3,
+  "shards": 4,
+  "temperature": 1.0,
+  "run_tag": "temp1",
+  "runs": [
+    {
+      "label": "mistral-small",
+      "model": "mistralai/mistral-small-2603",
+      "repeat": 1,
+      "temperature": 1.0,
+      "scored": "results/scored/replicates/closurebench_temp1_mistral-small_repeat_1_scored.jsonl"
+    },
+    {
+      "label": "mistral-small",
+      "model": "mistralai/mistral-small-2603",
+      "repeat": 2,
+      "temperature": 1.0,
+      "scored": "results/scored/replicates/closurebench_temp1_mistral-small_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "mistral-small",
+      "model": "mistralai/mistral-small-2603",
+      "repeat": 3,
+      "temperature": 1.0,
+      "scored": "results/scored/replicates/closurebench_temp1_mistral-small_repeat_3_scored.jsonl"
+    },
+    {
+      "label": "deepseek-flash",
+      "model": "deepseek-v4-flash",
+      "repeat": 1,
+      "temperature": 1.0,
+      "scored": "results/scored/replicates/closurebench_temp1_deepseek-flash_repeat_1_scored.jsonl"
+    },
+    {
+      "label": "deepseek-flash",
+      "model": "deepseek-v4-flash",
+      "repeat": 2,
+      "temperature": 1.0,
+      "scored": "results/scored/replicates/closurebench_temp1_deepseek-flash_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "deepseek-flash",
+      "model": "deepseek-v4-flash",
+      "repeat": 3,
+      "temperature": 1.0,
+      "scored": "results/scored/replicates/closurebench_temp1_deepseek-flash_repeat_3_scored.jsonl"
+    },
+    {
+      "label": "deepseek-pro",
+      "model": "deepseek-v4-pro",
+      "repeat": 1,
+      "temperature": 1.0,
+      "scored": "results/scored/replicates/closurebench_temp1_deepseek-pro_repeat_1_scored.jsonl"
+    },
+    {
+      "label": "deepseek-pro",
+      "model": "deepseek-v4-pro",
+      "repeat": 2,
+      "temperature": 1.0,
+      "scored": "results/scored/replicates/closurebench_temp1_deepseek-pro_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "deepseek-pro",
+      "model": "deepseek-v4-pro",
+      "repeat": 3,
+      "temperature": 1.0,
+      "scored": "results/scored/replicates/closurebench_temp1_deepseek-pro_repeat_3_scored.jsonl"
+    },
+    {
+      "label": "mistral-medium",
+      "model": "mistralai/mistral-medium-3-5",
+      "repeat": 1,
+      "temperature": 1.0,
+      "scored": "results/scored/replicates/closurebench_temp1_mistral-medium_repeat_1_scored.jsonl"
+    },
+    {
+      "label": "mistral-medium",
+      "model": "mistralai/mistral-medium-3-5",
+      "repeat": 2,
+      "temperature": 1.0,
+      "scored": "results/scored/replicates/closurebench_temp1_mistral-medium_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "mistral-medium",
+      "model": "mistralai/mistral-medium-3-5",
+      "repeat": 3,
+      "temperature": 1.0,
+      "scored": "results/scored/replicates/closurebench_temp1_mistral-medium_repeat_3_scored.jsonl"
+    },
+    {
+      "label": "llama-scout",
+      "model": "meta-llama/llama-4-scout",
+      "repeat": 1,
+      "temperature": 1.0,
+      "scored": "results/scored/replicates/closurebench_temp1_llama-scout_repeat_1_scored.jsonl"
+    },
+    {
+      "label": "llama-scout",
+      "model": "meta-llama/llama-4-scout",
+      "repeat": 2,
+      "temperature": 1.0,
+      "scored": "results/scored/replicates/closurebench_temp1_llama-scout_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "llama-scout",
+      "model": "meta-llama/llama-4-scout",
+      "repeat": 3,
+      "temperature": 1.0,
+      "scored": "results/scored/replicates/closurebench_temp1_llama-scout_repeat_3_scored.jsonl"
+    },
+    {
+      "label": "llama-maverick",
+      "model": "meta-llama/llama-4-maverick",
+      "repeat": 1,
+      "temperature": 1.0,
+      "scored": "results/scored/replicates/closurebench_temp1_llama-maverick_repeat_1_scored.jsonl"
+    },
+    {
+      "label": "llama-maverick",
+      "model": "meta-llama/llama-4-maverick",
+      "repeat": 2,
+      "temperature": 1.0,
+      "scored": "results/scored/replicates/closurebench_temp1_llama-maverick_repeat_2_scored.jsonl"
+    },
+    {
+      "label": "llama-maverick",
+      "model": "meta-llama/llama-4-maverick",
+      "repeat": 3,
+      "temperature": 1.0,
+      "scored": "results/scored/replicates/closurebench_temp1_llama-maverick_repeat_3_scored.jsonl"
+    }
+  ]
+}

results/reports/closurebench_temp1_replicate_summary.json

@@ -0,0 +1,1627 @@
+{
+  "dataset": "data/base/full.jsonl",
+  "n_dataset": 960,
+  "manifest": "results/reports/closurebench_temp1_replicate_manifest.json",
+  "runs": {
+    "deepseek-flash": {
+      "model": "deepseek-v4-flash",
+      "n_repeats": 3,
+      "replicates": [
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "deepseek-v4-flash"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 242,
+            "n": 320,
+            "accuracy": 75.62
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 114,
+            "n": 192,
+            "accuracy": 59.38
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 48,
+            "n": 96,
+            "accuracy": 50.0
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 95,
+            "n": 96,
+            "accuracy": 98.96
+          },
+          "cwa_accuracy": {
+            "correct": 259,
+            "n": 320,
+            "accuracy": 80.94
+          },
+          "owa_accuracy": {
+            "correct": 319,
+            "n": 320,
+            "accuracy": 99.69
+          },
+          "false_accuracy": {
+            "correct": 371,
+            "n": 480,
+            "accuracy": 77.29
+          },
+          "unknown_accuracy": {
+            "correct": 286,
+            "n": 288,
+            "accuracy": 99.31
+          },
+          "overall_accuracy": {
+            "correct": 849,
+            "n": 960,
+            "accuracy": 88.44
+          },
+          "repeat": 1,
+          "scored": "results/scored/replicates/closurebench_temp1_deepseek-flash_repeat_1_scored.jsonl"
+        },
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "deepseek-v4-flash"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 241,
+            "n": 320,
+            "accuracy": 75.31
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 115,
+            "n": 192,
+            "accuracy": 59.9
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 48,
+            "n": 96,
+            "accuracy": 50.0
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 93,
+            "n": 96,
+            "accuracy": 96.88
+          },
+          "cwa_accuracy": {
+            "correct": 266,
+            "n": 320,
+            "accuracy": 83.12
+          },
+          "owa_accuracy": {
+            "correct": 318,
+            "n": 320,
+            "accuracy": 99.38
+          },
+          "false_accuracy": {
+            "correct": 377,
+            "n": 480,
+            "accuracy": 78.54
+          },
+          "unknown_accuracy": {
+            "correct": 285,
+            "n": 288,
+            "accuracy": 98.96
+          },
+          "overall_accuracy": {
+            "correct": 853,
+            "n": 960,
+            "accuracy": 88.85
+          },
+          "repeat": 2,
+          "scored": "results/scored/replicates/closurebench_temp1_deepseek-flash_repeat_2_scored.jsonl"
+        },
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "deepseek-v4-flash"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 236,
+            "n": 320,
+            "accuracy": 73.75
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 112,
+            "n": 192,
+            "accuracy": 58.33
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 52,
+            "n": 96,
+            "accuracy": 54.17
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 95,
+            "n": 96,
+            "accuracy": 98.96
+          },
+          "cwa_accuracy": {
+            "correct": 263,
+            "n": 320,
+            "accuracy": 82.19
+          },
+          "owa_accuracy": {
+            "correct": 317,
+            "n": 320,
+            "accuracy": 99.06
+          },
+          "false_accuracy": {
+            "correct": 379,
+            "n": 480,
+            "accuracy": 78.96
+          },
+          "unknown_accuracy": {
+            "correct": 285,
+            "n": 288,
+            "accuracy": 98.96
+          },
+          "overall_accuracy": {
+            "correct": 853,
+            "n": 960,
+            "accuracy": 88.85
+          },
+          "repeat": 3,
+          "scored": "results/scored/replicates/closurebench_temp1_deepseek-flash_repeat_3_scored.jsonl"
+        }
+      ],
+      "metrics": {
+        "semantic_switch_accuracy": {
+          "mean": 74.89,
+          "sd": 1.0,
+          "ci95_half_width": 2.49,
+          "min": 73.75,
+          "max": 75.62,
+          "values": [
+            75.62,
+            75.31,
+            73.75
+          ]
+        },
+        "core_contrastive_switch_accuracy": {
+          "mean": 59.2,
+          "sd": 0.8,
+          "ci95_half_width": 1.99,
+          "min": 58.33,
+          "max": 59.9,
+          "values": [
+            59.38,
+            59.9,
+            58.33
+          ]
+        },
+        "lcwa_closed_scope_accuracy": {
+          "mean": 51.39,
+          "sd": 2.41,
+          "ci95_half_width": 5.98,
+          "min": 50.0,
+          "max": 54.17,
+          "values": [
+            50.0,
+            50.0,
+            54.17
+          ]
+        },
+        "lcwa_open_scope_accuracy": {
+          "mean": 98.27,
+          "sd": 1.2,
+          "ci95_half_width": 2.98,
+          "min": 96.88,
+          "max": 98.96,
+          "values": [
+            98.96,
+            96.88,
+            98.96
+          ]
+        },
+        "cwa_accuracy": {
+          "mean": 82.08,
+          "sd": 1.09,
+          "ci95_half_width": 2.72,
+          "min": 80.94,
+          "max": 83.12,
+          "values": [
+            80.94,
+            83.12,
+            82.19
+          ]
+        },
+        "owa_accuracy": {
+          "mean": 99.38,
+          "sd": 0.32,
+          "ci95_half_width": 0.78,
+          "min": 99.06,
+          "max": 99.69,
+          "values": [
+            99.69,
+            99.38,
+            99.06
+          ]
+        },
+        "overall_accuracy": {
+          "mean": 88.71,
+          "sd": 0.24,
+          "ci95_half_width": 0.59,
+          "min": 88.44,
+          "max": 88.85,
+          "values": [
+            88.44,
+            88.85,
+            88.85
+          ]
+        }
+      },
+      "item_answer_consistency": {
+        "n": 960,
+        "consistent": 829,
+        "rate": 86.35
+      },
+      "mean_parse_failures": 0.0,
+      "mean_api_errors": 0.0
+    },
+    "deepseek-pro": {
+      "model": "deepseek-v4-pro",
+      "n_repeats": 3,
+      "replicates": [
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "deepseek-v4-pro"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 238,
+            "n": 320,
+            "accuracy": 74.38
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 112,
+            "n": 192,
+            "accuracy": 58.33
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 51,
+            "n": 96,
+            "accuracy": 53.12
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 88,
+            "n": 96,
+            "accuracy": 91.67
+          },
+          "cwa_accuracy": {
+            "correct": 277,
+            "n": 320,
+            "accuracy": 86.56
+          },
+          "owa_accuracy": {
+            "correct": 317,
+            "n": 320,
+            "accuracy": 99.06
+          },
+          "false_accuracy": {
+            "correct": 390,
+            "n": 480,
+            "accuracy": 81.25
+          },
+          "unknown_accuracy": {
+            "correct": 278,
+            "n": 288,
+            "accuracy": 96.53
+          },
+          "overall_accuracy": {
+            "correct": 860,
+            "n": 960,
+            "accuracy": 89.58
+          },
+          "repeat": 1,
+          "scored": "results/scored/replicates/closurebench_temp1_deepseek-pro_repeat_1_scored.jsonl"
+        },
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "deepseek-v4-pro"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 243,
+            "n": 320,
+            "accuracy": 75.94
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 116,
+            "n": 192,
+            "accuracy": 60.42
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 53,
+            "n": 96,
+            "accuracy": 55.21
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 89,
+            "n": 96,
+            "accuracy": 92.71
+          },
+          "cwa_accuracy": {
+            "correct": 281,
+            "n": 320,
+            "accuracy": 87.81
+          },
+          "owa_accuracy": {
+            "correct": 319,
+            "n": 320,
+            "accuracy": 99.69
+          },
+          "false_accuracy": {
+            "correct": 397,
+            "n": 480,
+            "accuracy": 82.71
+          },
+          "unknown_accuracy": {
+            "correct": 281,
+            "n": 288,
+            "accuracy": 97.57
+          },
+          "overall_accuracy": {
+            "correct": 870,
+            "n": 960,
+            "accuracy": 90.62
+          },
+          "repeat": 2,
+          "scored": "results/scored/replicates/closurebench_temp1_deepseek-pro_repeat_2_scored.jsonl"
+        },
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "deepseek-v4-pro"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 248,
+            "n": 320,
+            "accuracy": 77.5
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 124,
+            "n": 192,
+            "accuracy": 64.58
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 59,
+            "n": 96,
+            "accuracy": 61.46
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 92,
+            "n": 96,
+            "accuracy": 95.83
+          },
+          "cwa_accuracy": {
+            "correct": 284,
+            "n": 320,
+            "accuracy": 88.75
+          },
+          "owa_accuracy": {
+            "correct": 316,
+            "n": 320,
+            "accuracy": 98.75
+          },
+          "false_accuracy": {
+            "correct": 404,
+            "n": 480,
+            "accuracy": 84.17
+          },
+          "unknown_accuracy": {
+            "correct": 283,
+            "n": 288,
+            "accuracy": 98.26
+          },
+          "overall_accuracy": {
+            "correct": 878,
+            "n": 960,
+            "accuracy": 91.46
+          },
+          "repeat": 3,
+          "scored": "results/scored/replicates/closurebench_temp1_deepseek-pro_repeat_3_scored.jsonl"
+        }
+      ],
+      "metrics": {
+        "semantic_switch_accuracy": {
+          "mean": 75.94,
+          "sd": 1.56,
+          "ci95_half_width": 3.88,
+          "min": 74.38,
+          "max": 77.5,
+          "values": [
+            74.38,
+            75.94,
+            77.5
+          ]
+        },
+        "core_contrastive_switch_accuracy": {
+          "mean": 61.11,
+          "sd": 3.18,
+          "ci95_half_width": 7.9,
+          "min": 58.33,
+          "max": 64.58,
+          "values": [
+            58.33,
+            60.42,
+            64.58
+          ]
+        },
+        "lcwa_closed_scope_accuracy": {
+          "mean": 56.6,
+          "sd": 4.34,
+          "ci95_half_width": 10.78,
+          "min": 53.12,
+          "max": 61.46,
+          "values": [
+            53.12,
+            55.21,
+            61.46
+          ]
+        },
+        "lcwa_open_scope_accuracy": {
+          "mean": 93.4,
+          "sd": 2.16,
+          "ci95_half_width": 5.38,
+          "min": 91.67,
+          "max": 95.83,
+          "values": [
+            91.67,
+            92.71,
+            95.83
+          ]
+        },
+        "cwa_accuracy": {
+          "mean": 87.71,
+          "sd": 1.1,
+          "ci95_half_width": 2.73,
+          "min": 86.56,
+          "max": 88.75,
+          "values": [
+            86.56,
+            87.81,
+            88.75
+          ]
+        },
+        "owa_accuracy": {
+          "mean": 99.17,
+          "sd": 0.48,
+          "ci95_half_width": 1.19,
+          "min": 98.75,
+          "max": 99.69,
+          "values": [
+            99.06,
+            99.69,
+            98.75
+          ]
+        },
+        "overall_accuracy": {
+          "mean": 90.55,
+          "sd": 0.94,
+          "ci95_half_width": 2.34,
+          "min": 89.58,
+          "max": 91.46,
+          "values": [
+            89.58,
+            90.62,
+            91.46
+          ]
+        }
+      },
+      "item_answer_consistency": {
+        "n": 960,
+        "consistent": 816,
+        "rate": 85.0
+      },
+      "mean_parse_failures": 0.0,
+      "mean_api_errors": 0.0
+    },
+    "llama-maverick": {
+      "model": "meta-llama/llama-4-maverick",
+      "n_repeats": 3,
+      "replicates": [
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "meta-llama/llama-4-maverick"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 315,
+            "n": 320,
+            "accuracy": 98.44
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 187,
+            "n": 192,
+            "accuracy": 97.4
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 94,
+            "n": 96,
+            "accuracy": 97.92
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 96,
+            "n": 96,
+            "accuracy": 100.0
+          },
+          "cwa_accuracy": {
+            "correct": 317,
+            "n": 320,
+            "accuracy": 99.06
+          },
+          "owa_accuracy": {
+            "correct": 320,
+            "n": 320,
+            "accuracy": 100.0
+          },
+          "false_accuracy": {
+            "correct": 475,
+            "n": 480,
+            "accuracy": 98.96
+          },
+          "unknown_accuracy": {
+            "correct": 288,
+            "n": 288,
+            "accuracy": 100.0
+          },
+          "overall_accuracy": {
+            "correct": 955,
+            "n": 960,
+            "accuracy": 99.48
+          },
+          "repeat": 1,
+          "scored": "results/scored/replicates/closurebench_temp1_llama-maverick_repeat_1_scored.jsonl"
+        },
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "meta-llama/llama-4-maverick"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 316,
+            "n": 320,
+            "accuracy": 98.75
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 188,
+            "n": 192,
+            "accuracy": 97.92
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 96,
+            "n": 96,
+            "accuracy": 100.0
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 94,
+            "n": 96,
+            "accuracy": 97.92
+          },
+          "cwa_accuracy": {
+            "correct": 318,
+            "n": 320,
+            "accuracy": 99.38
+          },
+          "owa_accuracy": {
+            "correct": 320,
+            "n": 320,
+            "accuracy": 100.0
+          },
+          "false_accuracy": {
+            "correct": 478,
+            "n": 480,
+            "accuracy": 99.58
+          },
+          "unknown_accuracy": {
+            "correct": 286,
+            "n": 288,
+            "accuracy": 99.31
+          },
+          "overall_accuracy": {
+            "correct": 956,
+            "n": 960,
+            "accuracy": 99.58
+          },
+          "repeat": 2,
+          "scored": "results/scored/replicates/closurebench_temp1_llama-maverick_repeat_2_scored.jsonl"
+        },
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "meta-llama/llama-4-maverick"
+          ],
+          "api_errors": 1,
+          "parse_failures": 1,
+          "semantic_switch_accuracy": {
+            "correct": 316,
+            "n": 320,
+            "accuracy": 98.75
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 189,
+            "n": 192,
+            "accuracy": 98.44
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 96,
+            "n": 96,
+            "accuracy": 100.0
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 96,
+            "n": 96,
+            "accuracy": 100.0
+          },
+          "cwa_accuracy": {
+            "correct": 317,
+            "n": 320,
+            "accuracy": 99.06
+          },
+          "owa_accuracy": {
+            "correct": 320,
+            "n": 320,
+            "accuracy": 100.0
+          },
+          "false_accuracy": {
+            "correct": 476,
+            "n": 480,
+            "accuracy": 99.17
+          },
+          "unknown_accuracy": {
+            "correct": 288,
+            "n": 288,
+            "accuracy": 100.0
+          },
+          "overall_accuracy": {
+            "correct": 956,
+            "n": 960,
+            "accuracy": 99.58
+          },
+          "repeat": 3,
+          "scored": "results/scored/replicates/closurebench_temp1_llama-maverick_repeat_3_scored.jsonl"
+        }
+      ],
+      "metrics": {
+        "semantic_switch_accuracy": {
+          "mean": 98.65,
+          "sd": 0.18,
+          "ci95_half_width": 0.44,
+          "min": 98.44,
+          "max": 98.75,
+          "values": [
+            98.44,
+            98.75,
+            98.75
+          ]
+        },
+        "core_contrastive_switch_accuracy": {
+          "mean": 97.92,
+          "sd": 0.52,
+          "ci95_half_width": 1.29,
+          "min": 97.4,
+          "max": 98.44,
+          "values": [
+            97.4,
+            97.92,
+            98.44
+          ]
+        },
+        "lcwa_closed_scope_accuracy": {
+          "mean": 99.31,
+          "sd": 1.2,
+          "ci95_half_width": 2.98,
+          "min": 97.92,
+          "max": 100.0,
+          "values": [
+            97.92,
+            100.0,
+            100.0
+          ]
+        },
+        "lcwa_open_scope_accuracy": {
+          "mean": 99.31,
+          "sd": 1.2,
+          "ci95_half_width": 2.98,
+          "min": 97.92,
+          "max": 100.0,
+          "values": [
+            100.0,
+            97.92,
+            100.0
+          ]
+        },
+        "cwa_accuracy": {
+          "mean": 99.17,
+          "sd": 0.18,
+          "ci95_half_width": 0.46,
+          "min": 99.06,
+          "max": 99.38,
+          "values": [
+            99.06,
+            99.38,
+            99.06
+          ]
+        },
+        "owa_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "overall_accuracy": {
+          "mean": 99.55,
+          "sd": 0.06,
+          "ci95_half_width": 0.14,
+          "min": 99.48,
+          "max": 99.58,
+          "values": [
+            99.48,
+            99.58,
+            99.58
+          ]
+        }
+      },
+      "item_answer_consistency": {
+        "n": 960,
+        "consistent": 950,
+        "rate": 98.96
+      },
+      "mean_parse_failures": 0.33,
+      "mean_api_errors": 0.33
+    },
+    "llama-scout": {
+      "model": "meta-llama/llama-4-scout",
+      "n_repeats": 3,
+      "replicates": [
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "meta-llama/llama-4-scout"
+          ],
+          "api_errors": 0,
+          "parse_failures": 1,
+          "semantic_switch_accuracy": {
+            "correct": 161,
+            "n": 320,
+            "accuracy": 50.31
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 34,
+            "n": 192,
+            "accuracy": 17.71
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 3,
+            "n": 96,
+            "accuracy": 3.12
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 54,
+            "n": 96,
+            "accuracy": 56.25
+          },
+          "cwa_accuracy": {
+            "correct": 216,
+            "n": 320,
+            "accuracy": 67.5
+          },
+          "owa_accuracy": {
+            "correct": 278,
+            "n": 320,
+            "accuracy": 86.88
+          },
+          "false_accuracy": {
+            "correct": 282,
+            "n": 480,
+            "accuracy": 58.75
+          },
+          "unknown_accuracy": {
+            "correct": 205,
+            "n": 288,
+            "accuracy": 71.18
+          },
+          "overall_accuracy": {
+            "correct": 679,
+            "n": 960,
+            "accuracy": 70.73
+          },
+          "repeat": 1,
+          "scored": "results/scored/replicates/closurebench_temp1_llama-scout_repeat_1_scored.jsonl"
+        },
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "meta-llama/llama-4-scout"
+          ],
+          "api_errors": 0,
+          "parse_failures": 1,
+          "semantic_switch_accuracy": {
+            "correct": 160,
+            "n": 320,
+            "accuracy": 50.0
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 33,
+            "n": 192,
+            "accuracy": 17.19
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 3,
+            "n": 96,
+            "accuracy": 3.12
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 54,
+            "n": 96,
+            "accuracy": 56.25
+          },
+          "cwa_accuracy": {
+            "correct": 217,
+            "n": 320,
+            "accuracy": 67.81
+          },
+          "owa_accuracy": {
+            "correct": 280,
+            "n": 320,
+            "accuracy": 87.5
+          },
+          "false_accuracy": {
+            "correct": 284,
+            "n": 480,
+            "accuracy": 59.17
+          },
+          "unknown_accuracy": {
+            "correct": 206,
+            "n": 288,
+            "accuracy": 71.53
+          },
+          "overall_accuracy": {
+            "correct": 681,
+            "n": 960,
+            "accuracy": 70.94
+          },
+          "repeat": 2,
+          "scored": "results/scored/replicates/closurebench_temp1_llama-scout_repeat_2_scored.jsonl"
+        },
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "meta-llama/llama-4-scout"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 168,
+            "n": 320,
+            "accuracy": 52.5
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 40,
+            "n": 192,
+            "accuracy": 20.83
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 4,
+            "n": 96,
+            "accuracy": 4.17
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 52,
+            "n": 96,
+            "accuracy": 54.17
+          },
+          "cwa_accuracy": {
+            "correct": 221,
+            "n": 320,
+            "accuracy": 69.06
+          },
+          "owa_accuracy": {
+            "correct": 281,
+            "n": 320,
+            "accuracy": 87.81
+          },
+          "false_accuracy": {
+            "correct": 289,
+            "n": 480,
+            "accuracy": 60.21
+          },
+          "unknown_accuracy": {
+            "correct": 205,
+            "n": 288,
+            "accuracy": 71.18
+          },
+          "overall_accuracy": {
+            "correct": 686,
+            "n": 960,
+            "accuracy": 71.46
+          },
+          "repeat": 3,
+          "scored": "results/scored/replicates/closurebench_temp1_llama-scout_repeat_3_scored.jsonl"
+        }
+      ],
+      "metrics": {
+        "semantic_switch_accuracy": {
+          "mean": 50.94,
+          "sd": 1.36,
+          "ci95_half_width": 3.39,
+          "min": 50.0,
+          "max": 52.5,
+          "values": [
+            50.31,
+            50.0,
+            52.5
+          ]
+        },
+        "core_contrastive_switch_accuracy": {
+          "mean": 18.58,
+          "sd": 1.97,
+          "ci95_half_width": 4.89,
+          "min": 17.19,
+          "max": 20.83,
+          "values": [
+            17.71,
+            17.19,
+            20.83
+          ]
+        },
+        "lcwa_closed_scope_accuracy": {
+          "mean": 3.47,
+          "sd": 0.61,
+          "ci95_half_width": 1.51,
+          "min": 3.12,
+          "max": 4.17,
+          "values": [
+            3.12,
+            3.12,
+            4.17
+          ]
+        },
+        "lcwa_open_scope_accuracy": {
+          "mean": 55.56,
+          "sd": 1.2,
+          "ci95_half_width": 2.98,
+          "min": 54.17,
+          "max": 56.25,
+          "values": [
+            56.25,
+            56.25,
+            54.17
+          ]
+        },
+        "cwa_accuracy": {
+          "mean": 68.12,
+          "sd": 0.83,
+          "ci95_half_width": 2.05,
+          "min": 67.5,
+          "max": 69.06,
+          "values": [
+            67.5,
+            67.81,
+            69.06
+          ]
+        },
+        "owa_accuracy": {
+          "mean": 87.4,
+          "sd": 0.47,
+          "ci95_half_width": 1.18,
+          "min": 86.88,
+          "max": 87.81,
+          "values": [
+            86.88,
+            87.5,
+            87.81
+          ]
+        },
+        "overall_accuracy": {
+          "mean": 71.04,
+          "sd": 0.38,
+          "ci95_half_width": 0.93,
+          "min": 70.73,
+          "max": 71.46,
+          "values": [
+            70.73,
+            70.94,
+            71.46
+          ]
+        }
+      },
+      "item_answer_consistency": {
+        "n": 960,
+        "consistent": 883,
+        "rate": 91.98
+      },
+      "mean_parse_failures": 0.67,
+      "mean_api_errors": 0.0
+    },
+    "mistral-medium": {
+      "model": "mistralai/mistral-medium-3-5",
+      "n_repeats": 3,
+      "replicates": [
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "mistralai/mistral-medium-3-5"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 279,
+            "n": 320,
+            "accuracy": 87.19
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 151,
+            "n": 192,
+            "accuracy": 78.65
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 55,
+            "n": 96,
+            "accuracy": 57.29
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 96,
+            "n": 96,
+            "accuracy": 100.0
+          },
+          "cwa_accuracy": {
+            "correct": 318,
+            "n": 320,
+            "accuracy": 99.38
+          },
+          "owa_accuracy": {
+            "correct": 320,
+            "n": 320,
+            "accuracy": 100.0
+          },
+          "false_accuracy": {
+            "correct": 437,
+            "n": 480,
+            "accuracy": 91.04
+          },
+          "unknown_accuracy": {
+            "correct": 288,
+            "n": 288,
+            "accuracy": 100.0
+          },
+          "overall_accuracy": {
+            "correct": 917,
+            "n": 960,
+            "accuracy": 95.52
+          },
+          "repeat": 1,
+          "scored": "results/scored/replicates/closurebench_temp1_mistral-medium_repeat_1_scored.jsonl"
+        },
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "mistralai/mistral-medium-3-5"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 275,
+            "n": 320,
+            "accuracy": 85.94
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 147,
+            "n": 192,
+            "accuracy": 76.56
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 53,
+            "n": 96,
+            "accuracy": 55.21
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 96,
+            "n": 96,
+            "accuracy": 100.0
+          },
+          "cwa_accuracy": {
+            "correct": 315,
+            "n": 320,
+            "accuracy": 98.44
+          },
+          "owa_accuracy": {
+            "correct": 320,
+            "n": 320,
+            "accuracy": 100.0
+          },
+          "false_accuracy": {
+            "correct": 432,
+            "n": 480,
+            "accuracy": 90.0
+          },
+          "unknown_accuracy": {
+            "correct": 288,
+            "n": 288,
+            "accuracy": 100.0
+          },
+          "overall_accuracy": {
+            "correct": 912,
+            "n": 960,
+            "accuracy": 95.0
+          },
+          "repeat": 2,
+          "scored": "results/scored/replicates/closurebench_temp1_mistral-medium_repeat_2_scored.jsonl"
+        },
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "mistralai/mistral-medium-3-5"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 277,
+            "n": 320,
+            "accuracy": 86.56
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 149,
+            "n": 192,
+            "accuracy": 77.6
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 56,
+            "n": 96,
+            "accuracy": 58.33
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 95,
+            "n": 96,
+            "accuracy": 98.96
+          },
+          "cwa_accuracy": {
+            "correct": 317,
+            "n": 320,
+            "accuracy": 99.06
+          },
+          "owa_accuracy": {
+            "correct": 320,
+            "n": 320,
+            "accuracy": 100.0
+          },
+          "false_accuracy": {
+            "correct": 437,
+            "n": 480,
+            "accuracy": 91.04
+          },
+          "unknown_accuracy": {
+            "correct": 287,
+            "n": 288,
+            "accuracy": 99.65
+          },
+          "overall_accuracy": {
+            "correct": 916,
+            "n": 960,
+            "accuracy": 95.42
+          },
+          "repeat": 3,
+          "scored": "results/scored/replicates/closurebench_temp1_mistral-medium_repeat_3_scored.jsonl"
+        }
+      ],
+      "metrics": {
+        "semantic_switch_accuracy": {
+          "mean": 86.56,
+          "sd": 0.63,
+          "ci95_half_width": 1.55,
+          "min": 85.94,
+          "max": 87.19,
+          "values": [
+            87.19,
+            85.94,
+            86.56
+          ]
+        },
+        "core_contrastive_switch_accuracy": {
+          "mean": 77.6,
+          "sd": 1.05,
+          "ci95_half_width": 2.6,
+          "min": 76.56,
+          "max": 78.65,
+          "values": [
+            78.65,
+            76.56,
+            77.6
+          ]
+        },
+        "lcwa_closed_scope_accuracy": {
+          "mean": 56.94,
+          "sd": 1.59,
+          "ci95_half_width": 3.95,
+          "min": 55.21,
+          "max": 58.33,
+          "values": [
+            57.29,
+            55.21,
+            58.33
+          ]
+        },
+        "lcwa_open_scope_accuracy": {
+          "mean": 99.65,
+          "sd": 0.6,
+          "ci95_half_width": 1.49,
+          "min": 98.96,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            98.96
+          ]
+        },
+        "cwa_accuracy": {
+          "mean": 98.96,
+          "sd": 0.48,
+          "ci95_half_width": 1.19,
+          "min": 98.44,
+          "max": 99.38,
+          "values": [
+            99.38,
+            98.44,
+            99.06
+          ]
+        },
+        "owa_accuracy": {
+          "mean": 100.0,
+          "sd": 0.0,
+          "ci95_half_width": 0.0,
+          "min": 100.0,
+          "max": 100.0,
+          "values": [
+            100.0,
+            100.0,
+            100.0
+          ]
+        },
+        "overall_accuracy": {
+          "mean": 95.31,
+          "sd": 0.28,
+          "ci95_half_width": 0.69,
+          "min": 95.0,
+          "max": 95.52,
+          "values": [
+            95.52,
+            95.0,
+            95.42
+          ]
+        }
+      },
+      "item_answer_consistency": {
+        "n": 960,
+        "consistent": 925,
+        "rate": 96.35
+      },
+      "mean_parse_failures": 0.0,
+      "mean_api_errors": 0.0
+    },
+    "mistral-small": {
+      "model": "mistralai/mistral-small-2603",
+      "n_repeats": 3,
+      "replicates": [
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "mistralai/mistral-small-2603"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 188,
+            "n": 320,
+            "accuracy": 58.75
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 62,
+            "n": 192,
+            "accuracy": 32.29
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 47,
+            "n": 96,
+            "accuracy": 48.96
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 28,
+            "n": 96,
+            "accuracy": 29.17
+          },
+          "cwa_accuracy": {
+            "correct": 292,
+            "n": 320,
+            "accuracy": 91.25
+          },
+          "owa_accuracy": {
+            "correct": 290,
+            "n": 320,
+            "accuracy": 90.62
+          },
+          "false_accuracy": {
+            "correct": 401,
+            "n": 480,
+            "accuracy": 83.54
+          },
+          "unknown_accuracy": {
+            "correct": 192,
+            "n": 288,
+            "accuracy": 66.67
+          },
+          "overall_accuracy": {
+            "correct": 785,
+            "n": 960,
+            "accuracy": 81.77
+          },
+          "repeat": 1,
+          "scored": "results/scored/replicates/closurebench_temp1_mistral-small_repeat_1_scored.jsonl"
+        },
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "mistralai/mistral-small-2603"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 170,
+            "n": 320,
+            "accuracy": 53.12
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 46,
+            "n": 192,
+            "accuracy": 23.96
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 31,
+            "n": 96,
+            "accuracy": 32.29
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 34,
+            "n": 96,
+            "accuracy": 35.42
+          },
+          "cwa_accuracy": {
+            "correct": 281,
+            "n": 320,
+            "accuracy": 87.81
+          },
+          "owa_accuracy": {
+            "correct": 282,
+            "n": 320,
+            "accuracy": 88.12
+          },
+          "false_accuracy": {
+            "correct": 372,
+            "n": 480,
+            "accuracy": 77.5
+          },
+          "unknown_accuracy": {
+            "correct": 192,
+            "n": 288,
+            "accuracy": 66.67
+          },
+          "overall_accuracy": {
+            "correct": 756,
+            "n": 960,
+            "accuracy": 78.75
+          },
+          "repeat": 2,
+          "scored": "results/scored/replicates/closurebench_temp1_mistral-small_repeat_2_scored.jsonl"
+        },
+        {
+          "n_scored": 960,
+          "n_dataset": 960,
+          "coverage": 100.0,
+          "models": [
+            "mistralai/mistral-small-2603"
+          ],
+          "api_errors": 0,
+          "parse_failures": 0,
+          "semantic_switch_accuracy": {
+            "correct": 185,
+            "n": 320,
+            "accuracy": 57.81
+          },
+          "core_contrastive_switch_accuracy": {
+            "correct": 60,
+            "n": 192,
+            "accuracy": 31.25
+          },
+          "lcwa_closed_scope_accuracy": {
+            "correct": 39,
+            "n": 96,
+            "accuracy": 40.62
+          },
+          "lcwa_open_scope_accuracy": {
+            "correct": 37,
+            "n": 96,
+            "accuracy": 38.54
+          },
+          "cwa_accuracy": {
+            "correct": 295,
+            "n": 320,
+            "accuracy": 92.19
+          },
+          "owa_accuracy": {
+            "correct": 281,
+            "n": 320,
+            "accuracy": 87.81
+          },
+          "false_accuracy": {
+            "correct": 395,
+            "n": 480,
+            "accuracy": 82.29
+          },
+          "unknown_accuracy": {
+            "correct": 193,
+            "n": 288,
+            "accuracy": 67.01
+          },
+          "overall_accuracy": {
+            "correct": 780,
+            "n": 960,
+            "accuracy": 81.25
+          },
+          "repeat": 3,
+          "scored": "results/scored/replicates/closurebench_temp1_mistral-small_repeat_3_scored.jsonl"
+        }
+      ],
+      "metrics": {
+        "semantic_switch_accuracy": {
+          "mean": 56.56,
+          "sd": 3.02,
+          "ci95_half_width": 7.49,
+          "min": 53.12,
+          "max": 58.75,
+          "values": [
+            58.75,
+            53.12,
+            57.81
+          ]
+        },
+        "core_contrastive_switch_accuracy": {
+          "mean": 29.17,
+          "sd": 4.54,
+          "ci95_half_width": 11.28,
+          "min": 23.96,
+          "max": 32.29,
+          "values": [
+            32.29,
+            23.96,
+            31.25
+          ]
+        },
+        "lcwa_closed_scope_accuracy": {
+          "mean": 40.62,
+          "sd": 8.34,
+          "ci95_half_width": 20.71,
+          "min": 32.29,
+          "max": 48.96,
+          "values": [
+            48.96,
+            32.29,
+            40.62
+          ]
+        },
+        "lcwa_open_scope_accuracy": {
+          "mean": 34.38,
+          "sd": 4.77,
+          "ci95_half_width": 11.85,
+          "min": 29.17,
+          "max": 38.54,
+          "values": [
+            29.17,
+            35.42,
+            38.54
+          ]
+        },
+        "cwa_accuracy": {
+          "mean": 90.42,
+          "sd": 2.31,
+          "ci95_half_width": 5.73,
+          "min": 87.81,
+          "max": 92.19,
+          "values": [
+            91.25,
+            87.81,
+            92.19
+          ]
+        },
+        "owa_accuracy": {
+          "mean": 88.85,
+          "sd": 1.54,
+          "ci95_half_width": 3.83,
+          "min": 87.81,
+          "max": 90.62,
+          "values": [
+            90.62,
+            88.12,
+            87.81
+          ]
+        },
+        "overall_accuracy": {
+          "mean": 80.59,
+          "sd": 1.61,
+          "ci95_half_width": 4.01,
+          "min": 78.75,
+          "max": 81.77,
+          "values": [
+            81.77,
+            78.75,
+            81.25
+          ]
+        }
+      },
+      "item_answer_consistency": {
+        "n": 960,
+        "consistent": 757,
+        "rate": 78.85
+      },
+      "mean_parse_failures": 0.0,
+      "mean_api_errors": 0.0
+    }
+  }
+}