Troubleshoot IO safely

Start with the boundary that failed. Do not repeat a run until you know whether the previous attempt changed a target.

Triage sequence

1. Protect: stop dependent work, restrict diagnostics, and preserve the minimal classified record.
2. Identify: record the request, content, runtime, and policy versions without copying secrets or real target data into public channels.
3. Locate: decide whether failure occurred at request, authorization, input, execution, dependency, observation, or target-verification boundary.
4. Compare: check the observed state against the selected release's documented contract and a known-good, equivalent non-sensitive case.
5. Verify impact: determine whether any target-side state changed.
6. Choose recovery: let the authorized owner select retry, repair, rollback, restore, or forward recovery.
7. Confirm: independently verify the resulting target and runtime state.

Symptom guide

Symptom	First boundary to inspect	Safe evidence
Request rejected	Authentication, policy, content identity, or scope	Redacted decision code and policy version
Run never begins	Dependency readiness, capacity, or authorization lifetime	Timestamped state transitions and health summary
Run ends unexpectedly	Runtime resources, dependency loss, content error, or policy action	Redacted termination class and immutable content identity
Run reports success but outcome is wrong	Target verification and content assumptions	Independent observed-state check
Repeated run changes more state	Content convergence, external drift, or mutable input	Input schema version and redacted change summary
Evidence is missing	Observation path, retention, access policy, or time source	Evidence-policy version and collection health

Escalation packet

Share only the product and release identity, UTC time window, boundary, redacted error class, reproducibility, business impact class, and checks already performed. Keep hostnames, addresses, customer names, tokens, inventories, raw logs, and internal topology in the approved private support channel.

For the cross-product method, see Troubleshooting principles.