Watch it investigate.

We ran a real online shop and broke one part of it on purpose, 14 times. Pick an incident and step through what TypeRCA looked at, what it suspected, and what it decided. Then see what actually broke. These are replays of the recorded runs, not live, and the misses are included.

The shop is Online Boutique, eleven services in Docker. Faults were picked by a seeded random draw before each round and written down in advance. Round 1 measured each service's latency only from the inside; rounds 2 and 3 also measured it from the callers. The AI decision model is TypeSafe's Jev; plain code decides when the answer counts as verified. Full reports: round 1, round 2, round 3.