Anthropic's 3 percent drop, from a system prompt change
A published postmortem where the model was not the cause. We read what a product-layer regression looks like and why a change log would never have shown it.
Failures
Public, sourced post-mortems of AI feature regressions: what was reported, what the operator said, and what evaluation would and would not have seen.
Public post-mortems of AI feature regressions, with sources.
4 posts
When a provider's own evals missed the degradation
A published provider postmortem where the evals did not reproduce what users were reporting. We read why a fixed suite can miss a fault that only some requests hit.
Incident 1152: what the AI Incident Database records
The AI Incident Database's record of a July 2025 agent failure, read as a trajectory problem rather than an answer-quality problem, with the hedge the registry itself uses.
What The Markup reported about the MyCity chatbot
A journalistic investigation found a public chatbot giving unlawful answers. We read the failure class it points at, and we mark clearly what the source does not tell you.