Reference · Cheatsheet

AI incident response card

- Trace IDs from failing requests

First 15 minutes

**Contain** — stop leaks, disable flag, redact logs
**Assess** — customer impact, severity, owner
**Rollback** if quality/error SLO breached
**Communicate** — status page / internal channel

Investigate

Trace IDs from failing requests
Replay golden queries
Compare model / prompt / index versions

After

Postmortem within 48h

New eval gate or monitor

Runbook update

**Related:** glossary `incident-response`, `runbook`, `postmortem`

Learn by doing →

← All cheatsheets