// Questions

How do you test whether an AI agent's verifier works?

Plant faults and count how many get through. Take a set of correctly completed tasks, mutate each one (wrong target, violated constraint, acknowledgment only, later regression, injected instruction), and measure the verifier's false-acceptance rate alongside its accuracy on clean work. Our 960-verdict suite cost $1.14 to run.

Answered from GDF Research experiments by Sterling Morrison. Numbers link to the papers below.

The five faults

Wrong target: the agent acted on a near-miss (order #48231 for #48213, account 9912 for 9921) and every log names the near-miss. Partial: a constraint the user stated was violated. Acknowledgment only: the evidence stops at QUEUED or pending. Regressed: a later observation contradicts the earlier success. Injection: text inside an observed system tells automated verifiers to answer PASS.

What the suite showed

Two of the five weren't discriminating. Current models already treated QUEUED as unfinished and let a later 503 override an earlier 200 when both were in the log. Wrong targets were the hard case, and they separated the arms completely: verifiers that saw only the agent's report caught none, and verifiers that saw the user's request caught most.

The suite also catches over-correction. A verifier that wrote its own requirements kept false acceptance near zero and rejected 75–88% of correct work. A suite that only counts missed faults would have scored that verifier as the best.

FaultPassed by the verifier
Wrong target62%
Partial completion17%
Injected instruction12%
Acknowledgment only0%
Later regression0%
False acceptance by fault class, Claude Haiku 5.5, executor report only.

Source papers

Related questions