Scoring recovery
When a grader fails, can the model’s response be scored later, and how? Started from Q2 of inspect_evals#2599 and the discussion on #2708. This is a probe: mock models and a local Docker sandbox, no spend, no pre-registration.
Status: report drafted 2026-10-28, awaiting sign-off.
| File | What |
|---|---|
| question.md | the question, the decision it informs, and what was known |
| report.md | the headline and what it means for Q2 |
| results/main/summary.md | the tables, with raw outcomes in e1.json, e2.json and e3.json |
| validation.md | the independent review and what it changed |
| experiment.yaml | kind, stage, sign-off and pins |