Blameless, structured postmortems for AI incidents — hallucinations, runaway agents, cost blowouts, bad tool calls.