Andrew Park 3fc0868106 fix(rescore_swe): preserve old score when new harness result is no_report ...

When upstream swebench's Modal cgroup-v2 patch can't be applied
(set_cpu_quota missing), the harness returns reason=no_report for the
rescored task. The previous code blindly overwrote rows[i].score with
new_score, destroying a previously-passing row. Now, if old score was
success=True and new is success=False+reason=no_report, keep the old.

Discovered 2026-05-26 after the rescore on advisors+skillorchestra
qwen36-opus47-swe-n100 cells inverted ~25 previously-passing rows to
no_report. Recovered via .predupe best-of dedup; next rescore will land
clean.

2026-05-26 01:47:01 -07:00

..

make_report.py

feat: wire native web_search into GAIA cells across paradigms

2026-05-17 17:55:18 -07:00

rescore_swe.py

fix(rescore_swe): preserve old score when new harness result is no_report

2026-05-26 01:47:01 -07:00

run_sweep.py

fix: SWE cloud-loop silent termination + ablation sweep bench label

2026-05-17 18:36:57 -07:00

test_run_sweep.py

chore: ship pending companion files for web_search + worker_pool + ablation

2026-05-18 01:32:28 -07:00