When Verification Fails: what AI evaluations cannot certify

When Verification Fails: measuring the safety properties that AI evaluations cannot certify

Frontier AI risk turns on a measurement problem. Decisions about deployment, scaling and governance rely on evaluations and trust signals (benchmarks, watermarks, safety fine-tuning checks) that can fail silently: fine-tuning removes safety behavior that evals then certify as intact, agent trajectories hide reward hacking behind passing scores, and some verification goals are provably unattainable. This project runs preregistered stress tests of these trust signals on frontier and open-weight systems, publishes exactly which safety claims each instrument can and cannot support, and releases the harnesses openly so any lab or auditor can reproduce the limits. Reducing catastrophic risk requires knowing where our instruments are blind; this work maps the blind spots before they are load-bearing.