An engineer working on AI evaluation and control - verifying whether checks actually work. Senior engineer, London; React/TypeScript and full-stack underneath.
- trusted-monitoring - self-hosted AI-control evaluation on open-weight models. First result at small scale: an 8B trusted monitor flagged 63% of attack-mode trajectories at a 4.1% false-positive rate, but only 9 of the 20 working backdoors.
- trace-eval - an evals harness that measures its LLM judge before trusting it (10 of 10 on a clear-cut gold set) and mutation-tests its own scorers (2 of 4 killed; the reference scorer failed its own baseline).
- sceptic - adversarial code review held under control measures; named the real defect in 6 of 6 plausible-but-wrong modules with passing tests. Proof of approach, not a usable reviewer.
More at danwalsh.co · LinkedIn · Bluesky