Incident response scenarios
Three production AI workloads with known failure modes. Each runs the full five-agent pipeline — coordinator, trace analyzer, eval runner, root cause, remediation, postmortem — against seeded watched-system traces and produces a validated Google-SRE postmortem.
False-positive burst on transaction classifier
FP rate spiked 3× in 90 seconds on the electronics merchant category. 1,247 legitimate transactions blocked, 312 customer accounts frozen, $84.3k in revenue at risk.
Sanctions-list hallucination on PEP screener
LLM-based PEP screener returned 7 fabricated OFAC matches in 120 seconds, 0 real matches in the same window. Regulatory disclosure threshold breached (FCA SUP 15.3, EU 5MLD Art. 33).
Latency regression after model deploy
Underwriting model p99 jumped 280ms to 4,234ms within 8 minutes of deploy. 89 underwriting decisions delayed, 12 user-facing timeouts, SLA breached (target 800ms).