Adversarial verification of agent findings
Does routing an agent's findings through independent skeptics catch more false positives than a single confidence score?
Spawning several independent verifiers, each prompted to refute a finding, removes more plausible-but-wrong results than asking the original agent how sure it is.
An agent that reviews code or audits a system produces a list of findings. Some are real; some are confident fiction. Trusting the agent's own certainty is how false positives reach a human as if they were facts.
The loop under test
For each finding, spawn a small panel of independent verifiers, each told to refute it and to default to "refuted" when unsure. Keep the finding only if a majority fails to refute it. The original agent never grades its own homework.
Early signal
The adversarial pass removes a meaningful share of findings that read as plausible but do not survive a second look. The open question is cost: how small the panel can be while still catching the fictions.
This is running. The goal is a verification step cheap enough that skipping it is never the right call.