| { |
| "primary_arms": [ |
| "qwen_base", |
| "qwen_final", |
| "gemma_base", |
| "gemma_final" |
| ], |
| "secondary_arms": [ |
| "qwen_domain", |
| "gemma_domain" |
| ], |
| "primary_checkpoint_policy": "Original weights versus fixed final domain104+semantic40 adapters, seed20261014. Do not select checkpoint using new test outputs. Qwen other seeds/step20 are supplementary only; do not pool them against one Gemma seed.", |
| "suites": { |
| "generation": { |
| "cases": 772, |
| "breakdown": { |
| "tool": 32, |
| "json": 732, |
| "text": 8 |
| }, |
| "history": "same canonical conversation contents and tools for all models; historical tool returns only, no new tool execution", |
| "tool_output_contract": "When a tool call is needed, output only one JSON object with a \"tool_calls\" array. Each element must have exactly \"name\" (a listed tool name) and \"arguments\" (a JSON object matching that tool schema). Do not wrap the JSON in Markdown or XML. Do not invent tool execution results.", |
| "output_budget": 2048, |
| "scoring": "Same frozen scorer for both families. Tool JSON validated against identical JSON schemas and exact reference call names/arguments; no XML-only requirement. Ordinary JSON/text rules frozen before generation. Pure format and pending-reference match are not scientific correctness." |
| }, |
| "source": { |
| "cases": 152, |
| "source_cards": 76, |
| "quote_locators": 76, |
| "source_dois": 19, |
| "output_budget": { |
| "source_card": 2048, |
| "quote_locator": 512 |
| }, |
| "input": "same complete source and deterministic index, no answer-derived hints", |
| "scorer_path": "scoring/source_benchmark_scoring.py", |
| "scorer_sha256": "9ffad5704b96998a50f38b6d92c02c785ea3f4b9f1f771c3a18a4b8bedd3512b", |
| "dev_test_pooling": false |
| }, |
| "workflow": { |
| "cases": 12, |
| "raw_input_sha256": "e93345326fb081fc281b820b1e77a4c4fd9a0409caccef806deb886fd1fdfaaf", |
| "coordination": "rule for all four arms; JEV disabled", |
| "limits": { |
| "context_tokens": 32768, |
| "max_new_tokens_per_call": 8192, |
| "max_llm_calls": 20, |
| "max_generated_tokens": 65536, |
| "max_seconds": 3600, |
| "max_corrections": 4, |
| "max_tool_requests": 8, |
| "max_graph_steps": 200 |
| }, |
| "tools": "same pinned tool versions, schemas, service assets and access for all arms; verify state after upstream restores services", |
| "decoding_seed": 20261018, |
| "order": "four blocks of three tasks, counterbalanced arm ordering", |
| "endpoint": "delivery and separately blinded scientific content review; failed/unfinished outputs retained; expert review remains false" |
| } |
| }, |
| "common_conditions": { |
| "dtype": "BF16", |
| "quantization": null, |
| "attention": "sdpa", |
| "batch_size": 1, |
| "do_sample": false, |
| "max_context_tokens": 32768, |
| "thinking": "Qwen enable_thinking=False; Gemma native no separate thinking mode", |
| "model_native_chat_templates": true, |
| "semantic_prompt_contents_identical": true, |
| "native_end_tokens_only_removed_for_scoring": true, |
| "output_repair": false, |
| "constrained_decoding": false, |
| "reference_in_model_input": false, |
| "exclusive_gpu": "one A80080GB per worker, no other process sharing", |
| "time_accounting": "record token counts and per-case elapsed time; separate model loading and CPU scoring; no claim equal tokenization or identical GPU wall time" |
| }, |
| "reuse_policy": "Reuse existing Qwen output only when exact rendered model-input hash, adapter hash, raw reference hash, decoding, budgets, model/runtime and task contract match. Rescore raw output with the common frozen evaluator; never reuse only aggregate metrics. Changed tool prompts/contracts require new inference for affected Qwen cases. Legacy Qwen LoRA+JEV workflows cannot be reused as LoRA+rule. Record provenance of every reused case.", |
| "analysis": "Pair by case; source/generalization confidence intervals cluster by original source/DOI, not assistant turns treated as independent. Keep all failures, timeouts and budget hits. Report within-family fine-tuning effects and between-family contrasts separately. No cross-tokenizer NLL ranking. Lock blinded content scores before unblinding.", |
| "scope_limits": [ |
| "Shared historical tests, not a fresh unseen independent test.", |
| "Native tokenizer, architecture, and chat wrappers differ; record lengths and protocol overhead.", |
| "Historical development effort differs; training sample exposure and budget are aligned.", |
| "Gemma final dev loss was worse than domain stage; fixed final checkpoint remains primary, domain checkpoint secondary.", |
| "Scientific references include pending review; mechanical pass alone is not scientific validity." |
| ], |
| "snapshot_version": "0.1", |
| "original_protocol_sha256": "355ee93579f7f5e096f5df605f50ed1bf5b754845d94347ba113e762a125cea9" |
| } |
|
|
