slug: scoring-recovery
title: "When a grader fails, can the model's response be scored later?"
kind: probe  # docs/framework.md#probes: mock models only, no pre-registration
source:
  - https://github.com/UKGovernmentBEIS/inspect_evals/issues/2599
  - https://github.com/UKGovernmentBEIS/inspect_evals/pull/2708
stage: report
gates:
  signoff: null  # {approved_by, date, commit}
budget:
  ceiling_usd: 0
  spent_usd: 0
pins:
  inspect_ai: "0.3.283"  # also run on 0.3.285 on 2026-10-28, with the same results