The Incident War Room
A production AI feature is failing right now, and you’re in the room. Make five containment calls in order — detect, read the signal, contain the leak, stop the blast radius, recover — and each one either buys time or bleeds it. The war room scores your run into a single minutes-to-recovery clock, a verdict from smooth to prolonged, and names the one wrong call that cost you the most. The moat is what it shows you along the way: the dashboards stay green the whole time, so you feel why “everything looked fine” is the most dangerous signal in an incident. Runs entirely in your browser (0 uploads, works offline).
How to use it
- Read the unfolding incident at each of the five steps and pick one action — there’s always a fastest right call and a few that waste minutes or make it worse.
- Answer every step, then press Run the postmortem to see your minutes-to-recovery clock and verdict.
- Open the X-ray panel to see the signal that looked fine, your costliest wrong call, and a note on every decision.
- Copy or share your run — the recovery time and “the dashboards were green the whole time” are the parts people argue about.
What this clears up (the fundamentals)
- Contain before you diagnose — the fastest recovery rolls back or flips the kill-switch first and investigates after. Every minute spent “adding logging” is a minute users keep getting the failure.
- Green dashboards are not health — uptime, latency and error-rate can all be fine while the model quietly returns garbage, leaks data, or loops. Trust the outputs, not just the infra metrics.
- Blast radius sets the priority — a data leak or a runaway agent grows every second it stays live, so you remove the capability first and reconcile the damage after.
- The fix is a guardrail, not “be careful” — you close an incident with a blameless postmortem and a new alert that catches the next silent drift, not by re-enabling autonomy and hoping.
Where it’s used
A free, top-of-funnel AI-engineering reliability drill — the fastest way to feel why an incident goes sideways before it happens to you. It’s a miatz build-lab concept playable: play the war room here, then learn to build the case-card scoring engine and the “looked fine at the time” reveal yourself in the AI Engineering track.
FAQ
Is this a real incident?
No — it’s an anonymized, generalized drill built from the common shapes of production AI failures (garbage output, prompt injection, data leakage, hallucinated actions, cost blowouts, over-automation). No real company or user data appears.
How is minutes-to-recovery calculated?
Each option carries a hand-set cost in minutes; the best call at every step is the cheapest. Your recovery time is the sum of the options you picked. An all-correct run hits the minimum (29 minutes) and scores smooth; wrong calls add their cost and push you toward bumpy or prolonged. It’s a deterministic rubric — no model call, no randomness.
Why do the dashboards stay green?
Because that’s the lesson. The drill deliberately shows you the same reassuring signal the real team saw. Infra dashboards measure whether the service is up, not whether its answers are good — and that gap is where most AI incidents hide.
What does “biggest mistake” mean?
Of the calls you got wrong, it’s the one that cost the most minutes. Fixing that single decision is where a real recovery gets faster, so the reveal names it instead of just giving you a score.
Is anything uploaded?
No. Your choices are scored entirely in your browser — nothing is transmitted, stored, or logged. Turn off your Wi-Fi and the war room still runs.
Limits
A teaching drill with five fixed steps and hand-set costs — it’s a fast, opinionated model of incident response, not a runbook or a real on-call tool. Actual incidents are messier: signals conflict, calls interact, and recovery time depends on your systems and team. The instincts it trains (contain first, distrust green dashboards, cut blast radius, close with a guardrail) are exactly the real practice.
Related
Part of the Demystify Playgrounds. Explore the rest from the Playgrounds home.
Bookmark this page (Ctrl+D, or ⌘D on Mac) — it works offline the next time you need it.