docs: publish GPT-OSS GPQA-main pruning results

[jayzou3773]

What and why

This PR publishes the completed GPT-OSS-120B held-out GPQA-main comparison after pruning on the same 64-example GPQA-main calibration split. It also teaches the frozen evaluator to preserve gpqa_main as a distinct dataset configuration instead of labeling every Idavidrein/gpqa manifest as Diamond.

Checkpoint	Mean pass@1 (avg@8)	Delta from base	Unparsed
BF16 baseline	76.99%	—	3
IntDim-E 50%	25.36%	-51.63 pp	188
IntDim-L 50%	62.99%	-14.00 pp	12
IntDim-G 50%	58.46%	-18.52 pp	44

Each model completed 384 held-out questions × 8 independent samples, or 3,072 completions. The metric is mean pass@1 over all completions; unparsed answers count as incorrect. All four runs have zero length truncations.

The exact 64-row calibration and 384-row held-out split is linked at jayzou3773/less-is-moe-gpqa-main-calibration-64, revision 7134dfef5af4605eae0706c30efa9226f49aed96. The dataset remains access-controlled because GPQA asks users not to reveal examples online.

The artifact directory contains complete metrics, run settings, grading audits, and all 12,288 per-sample predictions. Public predictions are sanitized: they retain UUIDs, gold/predicted letters, correctness, finish status, token counts, and question hashes while omitting question text, options, answer text, raw generations, and final-answer prose. The artifact manifest records SHA-256 hashes for the private raw and graded prediction files.

This is stacked on #35 because the experiment record follows the GPQA artifact format introduced there, and #35 is stacked on the evaluator in #34. After those PRs merge, this branch can be retargeted to main without changing this commit.

Verification

• uv run --with pytest --python 3.12 python -m pytest -q tests/test_supergpqa.py tests/test_registry_and_protocols.py — 28 passed
• validated all 12,288 sanitized prediction rows: unique sample IDs, gpqa_main labels, exact correct/unparsed/token sums, and absence of restricted fields
• verified every committed artifact against artifact-manifest.json SHA-256 values
• python3 -m json.tool on the experiment record, artifact manifest, and split manifest
• git diff --cached --check
• full B200 runs: four NVIDIA B200 GPUs per model, TP=4, final-answer-v4 grading, 3,072 completions per checkpoint

AI assistance

Codex collected the frozen B200 outputs, generated question-safe prediction artifacts, added explicit GPQA-main routing, wrote the experiment records, and performed consistency and privacy checks. Model generations and grading values come from the recorded B200 runs.

Checklist

• The change is focused and contains no unrelated cleanup.
• pytest passes for the affected evaluator and registry tests.
• The complete protocol and before/after results are recorded.
• Documentation and machine-readable artifacts are included.
• No weights, licensed dataset rows, raw generations, logs, or secrets are committed.