Your AI Root-Cause Is Wrong 23% of the Time: Testing an AI Observability Component Before On-Call Trusts It

Your incident starts with 47 alerts: twelve services, four databases, two queues, all degraded within 90 seconds. A causal-reasoning layer promises to collapse that storm into a single root cause in under four minutes - and in production it does, cutting median time-to-root-cause from 22 minutes to 4. Then comes the number nobody puts on a slide: it is wrong 23% of the time.

That headline improvement is survivorship-contaminated. The engine solves easy, repetitive cascades instantly and stalls on novel ones, so the median flatters it. To find the real value we had to measure the cost of its failures - what happens to an incident when an on-call engineer follows a confidently wrong answer.

This talk is about testing a system whose entire job is to interpret the telemetry you already depend on. Standard software testing does not apply: it reasons probabilistically over dynamic graphs, ground truth is scarce, and being wrong carries direct operational cost. We cover three ground-truth strategies, the failure-mode taxonomy behind that 23%, and the detection queries built for each. The core argument: detectability, not accuracy, is the safety property to design for.