Luna endpoint incident: re-run 5 tau2/cro/sol cells, rescore 10 tests
#1
by SoFTbOiU - opened
Corrections for the ~08-20 gpt-5.6-luna endpoint degradation (the "-tools" model-id cells).
- 5 tau2/cro/sol cells whose SEARCH ran against the degraded endpoint are fully re-run at
identical method config (cro, 24/24 proposer sessions, same model ids, api.openai.com
serving). Every cell roughly doubles dev_best; test +0.18-0.32. Old tarballs retained
under cells/cro/superseded/2026-09-08/. - 10 cells with healthy search but sick sealed-test scoring are test-rescored (3 reps,
archived final candidate, real CRO runtime); backing data in rescores/luna_reeval_20260906.json. - RESULTS.md rows amended (flags rerun-20260908 / rescored,luna-20260908), SUPERSEDED.md
extended, full incident writeup + deviations + caveats in LUNA_INCIDENT.md. - MANIFEST.json deliberately untouched; recommend regenerating post-merge.
From the MA-bench replication effort; happy to provide raw traces/logs for any cell.