Verifiers need their own evaluations
Agents can exploit missing constraints to satisfy a literal objective while doing the work incorrectly. Tests should include taking no action, reusing stale evidence, exhausting the budget, and submitting superficially correct outputs.
Record what changed after the fix
Preserve the strategy, verifier version, added checks, and rerun results. Blocking a case shows that the check works on that case; it does not establish that all future attacks will fail.
Separate explanation from measurement
Illustrations explain a mechanism. Measured reports require executable cases, task seeds, and run logs. Label each clearly so readers know which conclusions they can reproduce.
This is a draft. Adversarial evaluation results from production environments have not been published.