Datasets:
Contribute benchmark observation openai-codex-gpt-6-sol-114f336cc2-low-2026-09-24T20-52-52.679Z-cec88bb5
Accepted community observation openai-codex-gpt-6-sol-114f336cc2-low-2026-09-24T20-52-52.679Z-cec88bb5 in Dataset commit bd91e0089d6c78f4274e36f135b632ee6009fcdc.
@alexshpunt Usage-only correction requested for this existing accepted observation—not a new run or leaderboard entry.
Our local Dirac adapter mistakenly returned null for token and cost metrics, although the original private CLI logs retained complete API-status usage. We fixed the extractor and reconstructed the raw execution metrics in a private copy, then re-exported with the benchmark exporter and passed its validator.
- Total tokens, including cache: 3,586,585.
- Token buckets: input 662219; output 65870; cache read 2858496; cache write 0.
- CLI-reported per-call cost: $0.000000000 (Codex subscription cost is not an invoice amount).
- Repeated status snapshots are deduplicated by request ID; previous-round replay is excluded; reasoning is not counted twice; DeepSeek cache misses are assigned to cache writes rather than duplicated in input.
- 19 extractor tests + 4 full-run/resolver integration tests passed. Every one of the 226 tasks is retained. Only usage fields and the rounds digest changed; scores, outcomes, recoveries, timings, model/harness identities, configuration hashes, and run/client IDs are unchanged. No inference reruns occurred.
Corrected manifest SHA-256: ec69f6c642ba2546095c4d69f1c25be312b185d2aed82fe11b165dce623a9fb6. Corrected rounds SHA-256: 9867ac457add0db80ed9036feab963fad44d49d44aa2f1f1bdbba67b5d856206.
The corrected normalized bundle is ready, but uploading to this closed PR branch was denied, and the Hub also forbids us from reopening a discussion closed by another account. The normal importer rejects changed evidence under an accepted run ID. Please provide/reopen an update path and replace the existing canonical evidence and derived usage views under the same IDs, without adding another observation. No corrected data has been committed to Dataset main and no new contribution PR has been created.
@dirac-run This PR is open again. Please upload the corrected normalized bundle to this existing PR branch, matching the manifest and rounds SHA-256 hashes in your update comment, and reply here when it is ready. We will review the usage-only correction for the existing observation, without adding a new run or leaderboard entry. The correction has not been applied yet; acceptance will be confirmed here after validation and the Dataset update.
Community observation openai-codex-gpt-6-sol-114f336cc2-low-2026-09-24T20-52-52.679Z-cec88bb5 was already accepted on Dataset main at bd91e0089d6c78f4274e36f135b632ee6009fcdc.
@alexshpunt The corrected normalized bundle is now uploaded and verified byte-for-byte on this existing PR branch. Correction commit: 5676ed56d1899844611a64af6df4c3a953bf0d9b. Manifest SHA-256: ec69f6c642ba2546095c4d69f1c25be312b185d2aed82fe11b165dce623a9fb6. Rounds SHA-256: 9867ac457add0db80ed9036feab963fad44d49d44aa2f1f1bdbba67b5d856206. These match our prior correction request. Total tokens: 3586585; CLI-reported USD cost: 0. Only usage metrics and the round digest changed; all identities, tasks, scores, timings and recovery outcomes are unchanged. 19 extractor tests and 4 full-run/resolver integration tests passed. Please replace the existing accepted observation and rebuild its derived usage views under the same IDs, without adding a run or leaderboard entry. Dataset main replacement remains pending your validation.
Accepted the usage-only correction for existing observation openai-codex-gpt-6-sol-114f336cc2-low-2026-09-24T20-52-52.679Z-cec88bb5 in Dataset commit 34077a0003a0000c121a2b02215e823bf696bc30. Total tokens including cache: 3586585; CLI-reported USD cost: 0.000000000. Validated the corrected bundle hashes and unchanged non-usage evidence. The existing run and submission IDs, task results, timings and recovery outcomes are preserved; no new observation was added. Canonical evidence and derived usage views have been updated. CLI-reported cost is not an invoice amount. This PR remains open under manual review.