01Hackathon · Open source · 2026
Cascade
An on-call root-cause agent that separates cause from symptom, and knows when to say it doesn't know.
- Heuristic baseline, top-1
- 0.60
- Single-prompt baseline
- 0.30
- Hand-verified, incl. abstain
- 3/10
Triage → Hypothesis → Falsifier → Deterministic checks → Human gate
The bottleneck
The loudest alert is almost never the broken service. Retries amplify load on the way up the call chain, so a service three hops downstream of the real cause screams loudest while the origin is merely slow. Teams fix the loudest symptom and the incident recurs.
How it works
- A deterministic simulator generates 10 incidents over a synthetic 12-service system. Each has a planted fault whose ground truth is held out from every tool an agent can call.
- Fault effects propagate through each service's real timeout and retry policy, so retry-storm signatures fall out of the mechanics instead of being hand-authored.
- Triage narrows the blast radius, hypothesis proposes falsifiable claims, and the falsifier runs each falsification query against real telemetry.
- Two deterministic checks with no model call: a precedence checker rejects causes that happen after their symptoms, and an evidence validator resolves every citation back to the fixtures.
- Nothing is remediated automatically. Every fix waits behind an explicit human approval gate.
Results, honestly reported
Both baselines are fully measured over all 10 incidents (deploy-correlation heuristic: 0.60 top-1; single prompt: 0.30). Cascade is hand-verified correct on 3 representative incidents: a standard case, a deep cascade where the culprit never alerts, and a case where it correctly abstains. The full 10-incident number is deliberately left out rather than filled in unverified.
Incident Wall · live service topology · streaming triage · hypothesis board · approval gate · scoreboard