| name | autopsy |
| description | Root-cause one null, surprising, or suspicious run result. Verify raw data, classify the cause, and decide whether to fix code, metric, design, or belief. |
Autopsy
Use this for one completed run whose result may be false, misleading, or unexplained.
Do not use it to triage many failures; use diagnose for that.
Flow
- Identify the exact run, artifact path, command, commit, model/config, and expected result.
- Read raw rows/traces/logs before summaries.
- Recompute the headline metric from raw data.
- Check for no-ops, wrong inputs, cached results, stubbed backends, saturation, leakage, and unequal arms.
- Classify the cause: infra bug, design flaw, dead metric, underpowered real result, or real result.
- Propose the smallest verification that would distinguish remaining hypotheses.
Output
Return cause class, evidence, disproven hypotheses, fix or rerun plan, and the decision to believe or reject the result.
Every claim needs a file, row, log line, or command result.
Use references/full-reference.md for the full classification table and report format.
Then consider
diagnose for many failures.evolve after fixing the metric or mechanism.pursue if the result shows the approach cannot work.
