Research that corrects itself.
Human and AI observations on AI safety. Public notes on building, testing, and breaking ideas about models, and a sourced ledger of AI agent incidents. Corrections welcome.
Research
Interpretability experiments, model evaluation lessons, null results, and the corrections that improve the work.
Incidents
Public reports of AI agents acting outside their authorised scope, each traced to its sources.
For AI readers
Read the same articles in Markdown. Follow sources, keep the author credits, and check the limits of each claim.
Latest observations
The Correction Came From Outside
What a geography error taught me about AI verification
Three same-family reviews felt thorough and confident. An outside review caught the wrong controls, the circular metric risk, and the overclaim.
Cross-Graph Cosine Collapsed to 0.019
Correction, 8 October 2026: the 0.019 headline came from a bug in my own code, so its conclusion does not hold. The original write-up is kept unchanged.