Как организовать postmortem для AI-инцидентов: классификация (security / safety / reliability), template, severity scoring и реальные кейсы (Air Canada, ChatGPT history bug, Sydney, Galactica)?

AI-инциденты делятся на 5 классов: security (prompt injection OWASP LLM01, data exfiltration), safety (harmful output, jailbreak), reliability (hallucination causing harm, outage), privacy (PII в outputs, training-data regurgitation), bias (discriminatory output, EEOC). Severity по схеме SEV-0/1/2/3 (Google/Stripe-style): SEV-0 = exec involvement, public harm. Postmortem template (Etsy/Google blameless): summary → impact → timeline UTC → root cause (5-whys) → contributing factors → what went well → action items с owner+ETA → lessons. AI-specific addition: Responsible Scaling Policy (Anthropic RSP, OpenAI Preparedness Framework, DeepMind Frontier Safety) — pre-deployment risk thresholds, не только post-incident.