status and its next_action, classify
the state, and identify its spend before executing a recorded recovery.
Four-state triage
Spend boundary for recovery
Status reads, resume, patch repair, and rederive-only work preserve existing evidence and do not launch a new coding-agent attempt. Grader backfill can spend grader tokens only. Restarting or relaunching an arm runs agents again and can spend model tokens. You approve each spend boundary.Diagnose ambiguous symptoms
Systematic early zero-byte patches often indicate credentials or provider setup. Scattered empty patches with timeout or no-patch reasons may be genuine agent outcomes. An empty transcript with an exception usually means the container build failed before the agent ran. A base-URL configuration mistake can score a different provider than intended. Ask the agent to name the artifact-backed cause before treating any of these as model signal.Handle contamination without deleting evidence
Contamination flags annotate a run; they do not justify deleting a row. Stop superiority or rollout claims, verify the flag at trajectory level, and retain the cell in the denominator while it is unresolved. Follow the status or Trial Result’s supported emitted next action. If no remediation is emitted, stop and report the blocker; do not invent a command or configuration.Forbidden recovery moves
- Do not restart a healthy wait; wait for it to resume.
- Do not silently delete denominator rows; report them and repair or rerun.
- Do not rewrite result artifacts by hand; use the recorded recovery path.
- Do not promote stale or inspect-only evidence; read the current Trial Result.
- Do not spend on a new model run without approval.
- Do not hand-pair compare directories; use the frozen-baseline identity.
- Do not retry a terminal arm failure in a repair loop; inspect it and relaunch only with approval.