An artifact is useful evidence only when it can be connected to the work that created or changed it. Start with the expected artifact and follow its durable identities back through the run.
Inspection order
- Identify the application request and trace.
- Follow correlation to the session, thread, run, turn, tool or external effect.
- Read the terminal result rather than inferring success from UI text.
- Inspect the artifact identity, producing action, source revision, and Mission Control evidence.
Correlation contract
Request, correlation, actor, tenant, session, thread, run, turn, operation, tool/effect, and artifact identifiers must remain consistent across application, runner, worker, trace, and persisted result.
Provider reasoning and agent progress
Provider-returned reasoning and concise agent-authored progress are labeled, retained, and authorized independently. Neither is the terminal answer or proof that an artifact was committed.
Evaluation and recovery evidence
Attach evaluation results, recovery option/decision, approval, and external-effect result to the same run lineage. Do not infer those decisions from assistant text.
Common disagreements
| Observation | Useful evidence |
|---|---|
| Artifact is missing | terminal event, tool failure, requested path, capability policy |
| Artifact exists but is stale | observed source revision, edit result, checkpoint diff |
| Memory points to another artifact | memory revision and the update that changed linkedArtifacts |
| UI and stored artifact disagree | persisted artifact identity and the product projection |
| Trace and run history disagree | correlation, session, run, and event IDs on both records |
Do not rewrite historical evidence to make it agree with the current artifact. Preserve the original run and create new work only when a corrected artifact is actually needed.
Incident evidence package
Retain release identity, request/trace correlation, terminal event, relevant tool/effect events, artifact provenance, Mission Control evidence, evaluation/recovery decisions, and timestamps. Redact credentials, leases, private reasoning, and unrelated user data.
Continue with Store and replay, Reliability, and Evaluations.