Add calibforge_v4: unfiltered CalibForge tasks with LLM judge annotations
Keep all 5,317 unchanged source tasks. Publish 4,911 saved judgments and explicit error records for 406 tasks. Judge outcomes do not remove tasks. Include per-criterion results, saved attempt history, rubric, and provenance.
5 days ago
CrossCodeEval: note that v1 and v2 are not tested for offline use
Publish the incremental FACET v3 release: 109 thinking-on successes plus 619 thinking-off recoveries, with separate cohorts, settings, pass@2 results, trace statistics and all recorded conversations. Preserve evaluated task payloads and image manifests. Document online host networking, incomplete checks for two unsolved tasks, and the user-requested deviation from standard v3 settings.
FineEnvs v3: 515 newly solved tasks and all 1,998 Qwen/Qwen3.8-27B attempts (#31)
Verify all 27 rerun payload hashes against the 14,564 already-published tasks. Update current discovery status and README coverage; preserve underlying run evidence and unchanged task payloads.
Restore recorded Tmax validation timings and publish timing evidence
Restore three infrastructure build cases and seven reviewed no-op cases. Publish the v1 README and preserve acceptance decisions and original diagnostic evidence in metadata.
Record user-approved AI-detection exclusion for Tmax v1
