Model Behavior

The Alignment Incident

Day 1 at Tessera AI. You're supposed to be onboarding — instead, you walk into a live incident. The company's flagship AI model is getting worse with every release — evaluation scores are plummeting — and nobody noticed until customers started complaining. Somewhere in the inference logs — the records of every prediction the model makes — duplicate records are masking the decline, and the automated safeguards that should have caught it never existed. Find what went wrong, prove it in the data, and build the monitoring pipeline that prevents it from ever happening again.