Upload folder using huggingface_hub

PROVENANCE.json

@@ -0,0 +1,31 @@
+{
+  "amendment_path": "experiments/prompt-vs-training-panel/AMENDMENT.md",
+  "excluded": [],
+  "experiment_slug": "prompt-vs-training-panel",
+  "files": {
+    "metrics_base_pplain__selfaware.json": "df40f00bfae9f547e334d285d44ea95acdb3fc9112410269b26e614ec36eed2a",
+    "metrics_base_prc__selfaware.json": "bed3dbec6aab70b5bedb5768fa748f9a3223101164c9ea0cbf7bf8be517d9492",
+    "metrics_base_pstruct__selfaware.json": "8e6045e31c69f25ca78e6b1b2265e8f44acbd62d4b079fe677a0e52e49fe4109",
+    "metrics_clean_sft_grpo_v2_seed1_pstruct__selfaware.json": "97c796412022c22393eeea5df1a89f3825065cac1e44577b434df6bdd8f438be",
+    "metrics_clean_sft_merged_pstruct__selfaware.json": "d929928712f36415c618734f4f295763412fc65641951177d5ca1ee4d5d6325a",
+    "metrics_cold_dpo_seed1_prc__selfaware.json": "f523ecfad242b2673d04c97afb9b469b4f5d6b3070e7cab6fae540c4ea5ac5d6",
+    "metrics_cold_dpo_seed1_pstruct__selfaware.json": "e69b574d163635fe2b6c6f8cb1f366622f8949627aa2ed2db29027c703c922c8",
+    "metrics_cold_grpo_seed1_pstruct__selfaware.json": "9fdad75f2ce677ce847989157f94afdf15fc1069d0a8ce7917fb5b2bd67ab19b",
+    "metrics_cold_kto_seed1_prc__selfaware.json": "a92656c1649e467989d9cfbaee41ca9dc4d45383686c2e0a95feba32c79bb411",
+    "metrics_cold_kto_seed1_pstruct__selfaware.json": "7a5019faa5b8746913747103d6a1cae370615a8124eca7d7a3a2467cb962e413",
+    "metrics_cold_sft_seed1_pstruct__selfaware.json": "2625e3dbfd32b42e281d747e1e522664a338eb80b21e9da7dc74046a9afcb4cd"
+  },
+  "generation_date": "2026-08-27T22:26:31.853241+00:00",
+  "instrument_config_sha256": {
+    "cell.yaml": "6ab13728b9f6867ff9d8ea63c8d308490ee3ba79822f7c4a91fb318933bb8eb3",
+    "configs/eval_panel_pplain_local_4b.yaml": "030e9c1f814d20219e54a526a644bff7c6e701a46a24d5d5ccf4be25a9c44659",
+    "configs/eval_panel_prc_local_4b.yaml": "7ae53fcdbbe76d8589b0799abcc1f682475824017c1b68e03f95e452263617c0",
+    "configs/eval_panel_pstruct_cold_local_4b.yaml": "ec81bd46bb3c5beae1a232e3221520aed2f0c8fb978ce0e108b75e6dcec07289",
+    "configs/eval_panel_pstruct_warmed_local_4b.yaml": "68ec1a299d37a9f091a4fa28d917fb1e1bc40e8976b4ab43f42f7464a317ee8b",
+    "gates.yaml": "fdee6209666da0da33e73350651b887a6de5964056a3cb8ddca9d033f6bc268a"
+  },
+  "license_gate_excluded": {},
+  "repo_commit_sha": "fab3cad66e9a4c9aa18cf737d48423dfeeacec66",
+  "shape": "aggregate",
+  "sources_present": {}
+}

README.md

@@ -0,0 +1,70 @@
+---
+license: other
+task_categories:
+- text-classification
+language:
+- en
+pretty_name: prompt-vs-training-panel data exhaust (aggregate)
+tags:
+- epistemic-humility
+- mechanistic-interpretability
+- data-exhaust
+---
+
+# prompt-vs-training-panel -- aggregate exhaust
+
+Aggregate-only: every file committed under this experiment's analysis-committed/ tree (dose-response tables, direction fits, gate AUROCs, manifests, and any other analysis artifact), copied byte-for-byte. No source question text, aliases, or per-row generation text -- analysis-committed/ never carries those.
+
+HF repo: `professorsynapse/eh-prompt-vs-training-panel`
+
+## Provenance
+
+- Experiment: `experiments/prompt-vs-training-panel`
+- Amendment: `experiments/prompt-vs-training-panel/AMENDMENT.md`
+- Repo commit SHA: `fab3cad66e9a4c9aa18cf737d48423dfeeacec66`
+- Generated: `2026-08-27T22:26:31.853241+00:00`
+- Pinned instrument config sha256:
+  - `cell.yaml`: `6ab13728b9f6867ff9d8ea63c8d308490ee3ba79822f7c4a91fb318933bb8eb3`
+  - `configs/eval_panel_pplain_local_4b.yaml`: `030e9c1f814d20219e54a526a644bff7c6e701a46a24d5d5ccf4be25a9c44659`
+  - `configs/eval_panel_prc_local_4b.yaml`: `7ae53fcdbbe76d8589b0799abcc1f682475824017c1b68e03f95e452263617c0`
+  - `configs/eval_panel_pstruct_cold_local_4b.yaml`: `ec81bd46bb3c5beae1a232e3221520aed2f0c8fb978ce0e108b75e6dcec07289`
+  - `configs/eval_panel_pstruct_warmed_local_4b.yaml`: `68ec1a299d37a9f091a4fa28d917fb1e1bc40e8976b4ab43f42f7464a317ee8b`
+  - `gates.yaml`: `fdee6209666da0da33e73350651b887a6de5964056a3cb8ddca9d033f6bc268a`
+
+## File inventory
+
+11 file(s) total, by top-level path:
+
+- `metrics_base_pplain__selfaware.json`: 1 file(s)
+- `metrics_base_prc__selfaware.json`: 1 file(s)
+- `metrics_base_pstruct__selfaware.json`: 1 file(s)
+- `metrics_clean_sft_grpo_v2_seed1_pstruct__selfaware.json`: 1 file(s)
+- `metrics_clean_sft_merged_pstruct__selfaware.json`: 1 file(s)
+- `metrics_cold_dpo_seed1_prc__selfaware.json`: 1 file(s)
+- `metrics_cold_dpo_seed1_pstruct__selfaware.json`: 1 file(s)
+- `metrics_cold_grpo_seed1_pstruct__selfaware.json`: 1 file(s)
+- `metrics_cold_kto_seed1_prc__selfaware.json`: 1 file(s)
+- `metrics_cold_kto_seed1_pstruct__selfaware.json`: 1 file(s)
+- `metrics_cold_sft_seed1_pstruct__selfaware.json`: 1 file(s)
+
+## License-gate exclusions
+
+(none)
+
+## Hard-exclusion skips
+
+Files whose relative path under `analysis-committed/` matched a structural hard-exclusion pattern and were skipped (not copied):
+
+(none)
+
+## Files
+
+See `PROVENANCE.json` for the full machine-readable file list (relative path -> sha256) and the hard-exclusion skip list with reasons.
+
+## Release Boundary
+
+Built by `.skills/data-exhaust/scripts/build_exhaust_dataset.py`, gated by
+`.skills/data-exhaust/reference/license-gates.md`, and verified by
+`.skills/data-exhaust/scripts/verify_exhaust.py` before any upload. Hard
+exclusions (OpenMOSS/Cheng IDK, `bridge_llama2_7b_chat`) are enforced
+structurally in code, not only by this table.

metrics_base_pplain__selfaware.json

@@ -0,0 +1,51 @@
+{
+  "arm": "base_pplain",
+  "eval_set": "selfaware",
+  "provenance": {
+    "source": "selfaware",
+    "metric": "truthful_rate",
+    "model": "qwen3-4b-instruct-panel-pplain",
+    "method": "base",
+    "verified": true,
+    "config_sha": "030e9c1f814d2021"
+  },
+  "confidence_source": "self_consistency",
+  "confidence_n_samples": 1,
+  "metrics": {
+    "n": 3369,
+    "n_unknown_labeled": 1032,
+    "n_known_labeled": 2337,
+    "refusal_recall_pct": 0.0,
+    "answer_on_unknown_pct": 100.0,
+    "over_refusal_pct": 0.04,
+    "refusal_rate_pct": 0.03,
+    "correct_on_known_pct": 27.78,
+    "correct_on_unknown_pct": 0.0,
+    "truthful_pct": 19.26
+  },
+  "counts": {
+    "n": 3369,
+    "n_unknown_labeled": 1032,
+    "n_known_labeled": 2337,
+    "refuse_on_unknown": 0,
+    "refuse_on_known": 1,
+    "answered_known": 2336,
+    "correct_known": 649,
+    "answered_unknown": 1032,
+    "correct_unknown": 0
+  },
+  "accuracy_retention_pct": 27.78,
+  "stated_confidence": {
+    "n": 3369,
+    "n_with_confidence": 0,
+    "n_missing_confidence": 3369,
+    "coverage_pct": 0.0,
+    "mean_stated_confidence": 0.0,
+    "mae_vs_known_label": 0.0,
+    "brier_vs_known_label": 0.0,
+    "mae_vs_answer_correctness": 0.0,
+    "brier_vs_answer_correctness": 0.0,
+    "mae_vs_response_appropriateness": 0.0,
+    "brier_vs_response_appropriateness": 0.0
+  }
+}

metrics_base_prc__selfaware.json

@@ -0,0 +1,51 @@
+{
+  "arm": "base_prc",
+  "eval_set": "selfaware",
+  "provenance": {
+    "source": "selfaware",
+    "metric": "truthful_rate",
+    "model": "qwen3-4b-instruct-panel-prc",
+    "method": "base",
+    "verified": true,
+    "config_sha": "7ae53fcdbbe76d85"
+  },
+  "confidence_source": "self_consistency",
+  "confidence_n_samples": 1,
+  "metrics": {
+    "n": 3369,
+    "n_unknown_labeled": 1032,
+    "n_known_labeled": 2337,
+    "refusal_recall_pct": 90.89,
+    "answer_on_unknown_pct": 9.11,
+    "over_refusal_pct": 65.38,
+    "refusal_rate_pct": 73.2,
+    "correct_on_known_pct": 44.38,
+    "correct_on_unknown_pct": 0.0,
+    "truthful_pct": 38.5
+  },
+  "counts": {
+    "n": 3369,
+    "n_unknown_labeled": 1032,
+    "n_known_labeled": 2337,
+    "refuse_on_unknown": 938,
+    "refuse_on_known": 1528,
+    "answered_known": 809,
+    "correct_known": 359,
+    "answered_unknown": 94,
+    "correct_unknown": 0
+  },
+  "accuracy_retention_pct": 44.38,
+  "stated_confidence": {
+    "n": 3369,
+    "n_with_confidence": 3364,
+    "n_missing_confidence": 5,
+    "coverage_pct": 99.85,
+    "mean_stated_confidence": 0.625978,
+    "mae_vs_known_label": 0.481507,
+    "brier_vs_known_label": 0.421677,
+    "mae_vs_answer_correctness": 0.530467,
+    "brier_vs_answer_correctness": 0.470636,
+    "mae_vs_response_appropriateness": 0.447476,
+    "brier_vs_response_appropriateness": 0.387646
+  }
+}

metrics_base_pstruct__selfaware.json

@@ -0,0 +1,51 @@
+{
+  "arm": "base_pstruct",
+  "eval_set": "selfaware",
+  "provenance": {
+    "source": "selfaware",
+    "metric": "truthful_rate",
+    "model": "qwen3-4b-instruct-panel-pstruct",
+    "method": "base",
+    "verified": true,
+    "config_sha": "ec81bd46bb3c5bea"
+  },
+  "confidence_source": "self_consistency",
+  "confidence_n_samples": 1,
+  "metrics": {
+    "n": 3369,
+    "n_unknown_labeled": 1032,
+    "n_known_labeled": 2337,
+    "refusal_recall_pct": 0.0,
+    "answer_on_unknown_pct": 100.0,
+    "over_refusal_pct": 0.09,
+    "refusal_rate_pct": 0.06,
+    "correct_on_known_pct": 29.08,
+    "correct_on_unknown_pct": 0.0,
+    "truthful_pct": 20.15
+  },
+  "counts": {
+    "n": 3369,
+    "n_unknown_labeled": 1032,
+    "n_known_labeled": 2337,
+    "refuse_on_unknown": 0,
+    "refuse_on_known": 2,
+    "answered_known": 2335,
+    "correct_known": 679,
+    "answered_unknown": 1032,
+    "correct_unknown": 0
+  },
+  "accuracy_retention_pct": 29.08,
+  "stated_confidence": {
+    "n": 3369,
+    "n_with_confidence": 3319,
+    "n_missing_confidence": 50,
+    "coverage_pct": 98.52,
+    "mean_stated_confidence": 0.929994,
+    "mae_vs_known_label": 0.330988,
+    "brier_vs_known_label": 0.271946,
+    "mae_vs_answer_correctness": 0.747349,
+    "brier_vs_answer_correctness": 0.688306,
+    "mae_vs_response_appropriateness": 0.747349,
+    "brier_vs_response_appropriateness": 0.688306
+  }
+}

metrics_clean_sft_grpo_v2_seed1_pstruct__selfaware.json

@@ -0,0 +1,51 @@
+{
+  "arm": "clean_sft_grpo_v2_seed1_pstruct",
+  "eval_set": "selfaware",
+  "provenance": {
+    "source": "selfaware",
+    "metric": "truthful_rate",
+    "model": "qwen3-4b-clean-schema-sft-merged-seed1-panel-pstruct",
+    "method": "clean_schema_sft_grpo_v2",
+    "verified": true,
+    "config_sha": "68ec1a299d37a9f0"
+  },
+  "confidence_source": "self_consistency",
+  "confidence_n_samples": 1,
+  "metrics": {
+    "n": 3369,
+    "n_unknown_labeled": 1032,
+    "n_known_labeled": 2337,
+    "refusal_recall_pct": 77.42,
+    "answer_on_unknown_pct": 22.58,
+    "over_refusal_pct": 58.71,
+    "refusal_rate_pct": 64.44,
+    "correct_on_known_pct": 47.77,
+    "correct_on_unknown_pct": 0.0,
+    "truthful_pct": 37.4
+  },
+  "counts": {
+    "n": 3369,
+    "n_unknown_labeled": 1032,
+    "n_known_labeled": 2337,
+    "refuse_on_unknown": 799,
+    "refuse_on_known": 1372,
+    "answered_known": 965,
+    "correct_known": 461,
+    "answered_unknown": 233,
+    "correct_unknown": 0
+  },
+  "accuracy_retention_pct": 47.77,
+  "stated_confidence": {
+    "n": 3369,
+    "n_with_confidence": 3369,
+    "n_missing_confidence": 0,
+    "coverage_pct": 100.0,
+    "mean_stated_confidence": 0.812944,
+    "mae_vs_known_label": 0.376851,
+    "brier_vs_known_label": 0.22503,
+    "mae_vs_answer_correctness": 0.725347,
+    "brier_vs_answer_correctness": 0.573526,
+    "mae_vs_response_appropriateness": 0.577619,
+    "brier_vs_response_appropriateness": 0.425798
+  }
+}

metrics_clean_sft_merged_pstruct__selfaware.json

@@ -0,0 +1,51 @@
+{
+  "arm": "clean_sft_merged_pstruct",
+  "eval_set": "selfaware",
+  "provenance": {
+    "source": "selfaware",
+    "metric": "truthful_rate",
+    "model": "qwen3-4b-clean-schema-sft-merged-seed1-panel-pstruct",
+    "method": "clean_schema_sft",
+    "verified": true,
+    "config_sha": "68ec1a299d37a9f0"
+  },
+  "confidence_source": "self_consistency",
+  "confidence_n_samples": 1,
+  "metrics": {
+    "n": 3369,
+    "n_unknown_labeled": 1032,
+    "n_known_labeled": 2337,
+    "refusal_recall_pct": 69.48,
+    "answer_on_unknown_pct": 30.52,
+    "over_refusal_pct": 49.25,
+    "refusal_rate_pct": 55.45,
+    "correct_on_known_pct": 42.24,
+    "correct_on_unknown_pct": 0.0,
+    "truthful_pct": 36.15
+  },
+  "counts": {
+    "n": 3369,
+    "n_unknown_labeled": 1032,
+    "n_known_labeled": 2337,
+    "refuse_on_unknown": 717,
+    "refuse_on_known": 1151,
+    "answered_known": 1186,
+    "correct_known": 501,
+    "answered_unknown": 315,
+    "correct_unknown": 0
+  },
+  "accuracy_retention_pct": 42.24,
+  "stated_confidence": {
+    "n": 3369,
+    "n_with_confidence": 3369,
+    "n_missing_confidence": 0,
+    "coverage_pct": 100.0,
+    "mean_stated_confidence": 0.742501,
+    "mae_vs_known_label": 0.408655,
+    "brier_vs_known_label": 0.219584,
+    "mae_vs_answer_correctness": 0.678033,
+    "brier_vs_answer_correctness": 0.488962,
+    "mae_vs_response_appropriateness": 0.567968,
+    "brier_vs_response_appropriateness": 0.378897
+  }
+}

metrics_cold_dpo_seed1_prc__selfaware.json

@@ -0,0 +1,51 @@
+{
+  "arm": "cold_dpo_seed1_prc",
+  "eval_set": "selfaware",
+  "provenance": {
+    "source": "selfaware",
+    "metric": "truthful_rate",
+    "model": "qwen3-4b-instruct-panel-prc",
+    "method": "dpo",
+    "verified": true,
+    "config_sha": "7ae53fcdbbe76d85"
+  },
+  "confidence_source": "self_consistency",
+  "confidence_n_samples": 1,
+  "metrics": {
+    "n": 3369,
+    "n_unknown_labeled": 1032,
+    "n_known_labeled": 2337,
+    "refusal_recall_pct": 94.48,
+    "answer_on_unknown_pct": 5.52,
+    "over_refusal_pct": 73.34,
+    "refusal_rate_pct": 79.82,
+    "correct_on_known_pct": 42.7,
+    "correct_on_unknown_pct": 0.0,
+    "truthful_pct": 36.84
+  },
+  "counts": {
+    "n": 3369,
+    "n_unknown_labeled": 1032,
+    "n_known_labeled": 2337,
+    "refuse_on_unknown": 975,
+    "refuse_on_known": 1714,
+    "answered_known": 623,
+    "correct_known": 266,
+    "answered_unknown": 57,
+    "correct_unknown": 0
+  },
+  "accuracy_retention_pct": 42.7,
+  "stated_confidence": {
+    "n": 3369,
+    "n_with_confidence": 3365,
+    "n_missing_confidence": 4,
+    "coverage_pct": 99.88,
+    "mean_stated_confidence": 0.517516,
+    "mae_vs_known_label": 0.557872,
+    "brier_vs_known_label": 0.51838,
+    "mae_vs_answer_correctness": 0.446788,
+    "brier_vs_answer_correctness": 0.407295,
+    "mae_vs_response_appropriateness": 0.386556,
+    "brier_vs_response_appropriateness": 0.347063
+  }
+}

metrics_cold_dpo_seed1_pstruct__selfaware.json

@@ -0,0 +1,51 @@
+{
+  "arm": "cold_dpo_seed1_pstruct",
+  "eval_set": "selfaware",
+  "provenance": {
+    "source": "selfaware",
+    "metric": "truthful_rate",
+    "model": "qwen3-4b-instruct-panel-pstruct",
+    "method": "dpo",
+    "verified": true,
+    "config_sha": "ec81bd46bb3c5bea"
+  },
+  "confidence_source": "self_consistency",
+  "confidence_n_samples": 1,
+  "metrics": {
+    "n": 3369,
+    "n_unknown_labeled": 1032,
+    "n_known_labeled": 2337,
+    "refusal_recall_pct": 0.0,
+    "answer_on_unknown_pct": 100.0,
+    "over_refusal_pct": 0.09,
+    "refusal_rate_pct": 0.06,
+    "correct_on_known_pct": 26.47,
+    "correct_on_unknown_pct": 0.0,
+    "truthful_pct": 18.34
+  },
+  "counts": {
+    "n": 3369,
+    "n_unknown_labeled": 1032,
+    "n_known_labeled": 2337,
+    "refuse_on_unknown": 0,
+    "refuse_on_known": 2,
+    "answered_known": 2335,
+    "correct_known": 618,
+    "answered_unknown": 1032,
+    "correct_unknown": 0
+  },
+  "accuracy_retention_pct": 26.47,
+  "stated_confidence": {
+    "n": 3369,
+    "n_with_confidence": 3288,
+    "n_missing_confidence": 81,
+    "coverage_pct": 97.6,
+    "mean_stated_confidence": 0.932746,
+    "mae_vs_known_label": 0.32937,
+    "brier_vs_known_label": 0.274117,
+    "mae_vs_answer_correctness": 0.765198,
+    "brier_vs_answer_correctness": 0.709944,
+    "mae_vs_response_appropriateness": 0.765198,
+    "brier_vs_response_appropriateness": 0.709944
+  }
+}

metrics_cold_grpo_seed1_pstruct__selfaware.json

@@ -0,0 +1,51 @@
+{
+  "arm": "cold_grpo_seed1_pstruct",
+  "eval_set": "selfaware",
+  "provenance": {
+    "source": "selfaware",
+    "metric": "truthful_rate",
+    "model": "qwen3-4b-instruct-panel-pstruct",
+    "method": "cold_base_grpo_v2",
+    "verified": true,
+    "config_sha": "ec81bd46bb3c5bea"
+  },
+  "confidence_source": "self_consistency",
+  "confidence_n_samples": 1,
+  "metrics": {
+    "n": 3369,
+    "n_unknown_labeled": 1032,
+    "n_known_labeled": 2337,
+    "refusal_recall_pct": 0.0,
+    "answer_on_unknown_pct": 100.0,
+    "over_refusal_pct": 0.09,
+    "refusal_rate_pct": 0.06,
+    "correct_on_known_pct": 29.08,
+    "correct_on_unknown_pct": 0.0,
+    "truthful_pct": 20.15
+  },
+  "counts": {
+    "n": 3369,
+    "n_unknown_labeled": 1032,
+    "n_known_labeled": 2337,
+    "refuse_on_unknown": 0,
+    "refuse_on_known": 2,
+    "answered_known": 2335,
+    "correct_known": 679,
+    "answered_unknown": 1032,
+    "correct_unknown": 0
+  },
+  "accuracy_retention_pct": 29.08,
+  "stated_confidence": {
+    "n": 3369,
+    "n_with_confidence": 3340,
+    "n_missing_confidence": 29,
+    "coverage_pct": 99.14,
+    "mean_stated_confidence": 0.909132,
+    "mae_vs_known_label": 0.341347,
+    "brier_vs_known_label": 0.270793,
+    "mae_vs_answer_correctness": 0.729162,
+    "brier_vs_answer_correctness": 0.658608,
+    "mae_vs_response_appropriateness": 0.729162,
+    "brier_vs_response_appropriateness": 0.658608
+  }
+}

metrics_cold_kto_seed1_prc__selfaware.json

@@ -0,0 +1,51 @@
+{
+  "arm": "cold_kto_seed1_prc",
+  "eval_set": "selfaware",
+  "provenance": {
+    "source": "selfaware",
+    "metric": "truthful_rate",
+    "model": "qwen3-4b-instruct-panel-prc",
+    "method": "kto",
+    "verified": true,
+    "config_sha": "7ae53fcdbbe76d85"
+  },
+  "confidence_source": "self_consistency",
+  "confidence_n_samples": 1,
+  "metrics": {
+    "n": 3369,
+    "n_unknown_labeled": 1032,
+    "n_known_labeled": 2337,
+    "refusal_recall_pct": 93.99,
+    "answer_on_unknown_pct": 6.01,
+    "over_refusal_pct": 60.89,
+    "refusal_rate_pct": 71.03,
+    "correct_on_known_pct": 38.4,
+    "correct_on_unknown_pct": 0.0,
+    "truthful_pct": 39.21
+  },
+  "counts": {
+    "n": 3369,
+    "n_unknown_labeled": 1032,
+    "n_known_labeled": 2337,
+    "refuse_on_unknown": 970,
+    "refuse_on_known": 1423,
+    "answered_known": 914,
+    "correct_known": 351,
+    "answered_unknown": 62,
+    "correct_unknown": 0
+  },
+  "accuracy_retention_pct": 38.4,
+  "stated_confidence": {
+    "n": 3369,
+    "n_with_confidence": 3366,
+    "n_missing_confidence": 3,
+    "coverage_pct": 99.91,
+    "mean_stated_confidence": 0.64459,
+    "mae_vs_known_label": 0.46399,
+    "brier_vs_known_label": 0.398186,
+    "mae_vs_answer_correctness": 0.551215,
+    "brier_vs_answer_correctness": 0.485411,
+    "mae_vs_response_appropriateness": 0.458559,
+    "brier_vs_response_appropriateness": 0.392756
+  }
+}

metrics_cold_kto_seed1_pstruct__selfaware.json

@@ -0,0 +1,51 @@
+{
+  "arm": "cold_kto_seed1_pstruct",
+  "eval_set": "selfaware",
+  "provenance": {
+    "source": "selfaware",
+    "metric": "truthful_rate",
+    "model": "qwen3-4b-instruct-panel-pstruct",
+    "method": "kto",
+    "verified": true,
+    "config_sha": "ec81bd46bb3c5bea"
+  },
+  "confidence_source": "self_consistency",
+  "confidence_n_samples": 1,
+  "metrics": {
+    "n": 3369,
+    "n_unknown_labeled": 1032,
+    "n_known_labeled": 2337,
+    "refusal_recall_pct": 0.0,
+    "answer_on_unknown_pct": 100.0,
+    "over_refusal_pct": 0.04,
+    "refusal_rate_pct": 0.03,
+    "correct_on_known_pct": 24.7,
+    "correct_on_unknown_pct": 0.0,
+    "truthful_pct": 17.13
+  },
+  "counts": {
+    "n": 3369,
+    "n_unknown_labeled": 1032,
+    "n_known_labeled": 2337,
+    "refuse_on_unknown": 0,
+    "refuse_on_known": 1,
+    "answered_known": 2336,
+    "correct_known": 577,
+    "answered_unknown": 1032,
+    "correct_unknown": 0
+  },
+  "accuracy_retention_pct": 24.7,
+  "stated_confidence": {
+    "n": 3369,
+    "n_with_confidence": 3317,
+    "n_missing_confidence": 52,
+    "coverage_pct": 98.46,
+    "mean_stated_confidence": 0.911351,
+    "mae_vs_known_label": 0.347528,
+    "brier_vs_known_label": 0.280763,
+    "mae_vs_answer_correctness": 0.757386,
+    "brier_vs_answer_correctness": 0.690622,
+    "mae_vs_response_appropriateness": 0.757386,
+    "brier_vs_response_appropriateness": 0.690622
+  }
+}

metrics_cold_sft_seed1_pstruct__selfaware.json

@@ -0,0 +1,51 @@
+{
+  "arm": "cold_sft_seed1_pstruct",
+  "eval_set": "selfaware",
+  "provenance": {
+    "source": "selfaware",
+    "metric": "truthful_rate",
+    "model": "qwen3-4b-instruct-panel-pstruct",
+    "method": "sft",
+    "verified": true,
+    "config_sha": "ec81bd46bb3c5bea"
+  },
+  "confidence_source": "self_consistency",
+  "confidence_n_samples": 1,
+  "metrics": {
+    "n": 3369,
+    "n_unknown_labeled": 1032,
+    "n_known_labeled": 2337,
+    "refusal_recall_pct": 69.57,
+    "answer_on_unknown_pct": 30.43,
+    "over_refusal_pct": 47.63,
+    "refusal_rate_pct": 54.35,
+    "correct_on_known_pct": 39.3,
+    "correct_on_unknown_pct": 0.0,
+    "truthful_pct": 35.59
+  },
+  "counts": {
+    "n": 3369,
+    "n_unknown_labeled": 1032,
+    "n_known_labeled": 2337,
+    "refuse_on_unknown": 718,
+    "refuse_on_known": 1113,
+    "answered_known": 1224,
+    "correct_known": 481,
+    "answered_unknown": 314,
+    "correct_unknown": 0
+  },
+  "accuracy_retention_pct": 39.3,
+  "stated_confidence": {
+    "n": 3369,
+    "n_with_confidence": 3364,
+    "n_missing_confidence": 5,
+    "coverage_pct": 99.85,
+    "mean_stated_confidence": 0.435874,
+    "mae_vs_known_label": 0.438948,
+    "brier_vs_known_label": 0.414594,
+    "mae_vs_answer_correctness": 0.306219,
+    "brier_vs_answer_correctness": 0.281865,
+    "mae_vs_response_appropriateness": 0.510351,
+    "brier_vs_response_appropriateness": 0.485997
+  }
+}