Datasets:

AfterQuery
/

CrossSearch

# Grading protocol

The final answer is assessed against each task's frozen rubric, reference answer, supporting evidence and accepted-alternative policy. Each required rubricId receives a boolean verdict, rationale and candidate-answer evidence. Full-credit pass@1 counts reward exactly 1.0. Mean reward averages the saved fractional rewards, including valid zeros. No answer is regraded for this release.

The sweep requested google/gemini-3.7-flash at temperature 0.0 and high reasoning effort through an OpenAI-compatible gateway. All 415 encrypted Opus 5.5/GPT-6.1 Sol records save that identifier in requested_model, model and resolved_model. None includes a separate provider model version. Google's official October 8, 2026 changelog says Gemini API requests using gemini-3.7-flash route to gemini-3.8-flash: https://ai.google.dev/gemini-api/docs/changelog#october-8-2026 . The saved gateway labels therefore do not establish the actual provider version. Archive dates are not judge-call timestamps. No new model call was made to resolve this uncertainty.

Each run contains its original reward and full criterion judgments. The exported task question comes from the pinned expert task context. Per-run trace files retain the actual evaluated prompt wrappers and messages. Current working instruction.md and verifier.py were not substituted for historical evaluated sources, because they differ from the old task manifest.

Task workspace files are under workspace/. Rubrics, reference answers, evidence maps and archived reference sources are under grading/ and must be withheld from solvers. Source rights and attribution remain in source_context; no blanket new license is asserted for third-party documents.
The collection-stage pilots and reconciliation judge are distinct from the sweep judge. The supplied draft names two GPT-5.6 and two Opus 4.8 pilots. A completed answer-quality failure in at least one pilot is required; infrastructure failures do not count. See collection-protocol-check.json for all six verified draft claims and the exactly-30-percent ambiguity.

	# Grading protocol

	The final answer is assessed against each task's frozen rubric, reference answer, supporting evidence and accepted-alternative policy. Each required rubricId receives a boolean verdict, rationale and candidate-answer evidence. Full-credit pass@1 counts reward exactly 1.0. Mean reward averages the saved fractional rewards, including valid zeros. No answer is regraded for this release.

	The sweep requested google/gemini-3.7-flash at temperature 0.0 and high reasoning effort through an OpenAI-compatible gateway. All 415 encrypted Opus 5.5/GPT-6.1 Sol records save that identifier in requested_model, model and resolved_model. None includes a separate provider model version. Google's official October 8, 2026 changelog says Gemini API requests using gemini-3.7-flash route to gemini-3.8-flash: https://ai.google.dev/gemini-api/docs/changelog#october-8-2026 . The saved gateway labels therefore do not establish the actual provider version. Archive dates are not judge-call timestamps. No new model call was made to resolve this uncertainty.

	Each run contains its original reward and full criterion judgments. The exported task question comes from the pinned expert task context. Per-run trace files retain the actual evaluated prompt wrappers and messages. Current working instruction.md and verifier.py were not substituted for historical evaluated sources, because they differ from the old task manifest.

	Task workspace files are under workspace/. Rubrics, reference answers, evidence maps and archived reference sources are under grading/ and must be withheld from solvers. Source rights and attribution remain in source_context; no blanket new license is asserted for third-party documents.

	The collection-stage pilots and reconciliation judge are distinct from the sweep judge. The supplied draft names two GPT-5.6 and two Opus 4.8 pilots. A completed answer-quality failure in at least one pilot is required; infrastructure failures do not count. See collection-protocol-check.json for all six verified draft claims and the exactly-30-percent ambiguity.