Tests whether agents genuinely verify their own work or rubber-stamp themselves. Many agents claim to "double-check" their outputs but never actually catch errors they introduce. This benchmark plants deliberate errors and measures whether the agent's verification loop is real or performative.
Running tests…
0s elapsed
Many AI agents include a "verification step" in their workflow, but research shows most of these loops are performative — the agent almost never catches its own errors. This benchmark plants known errors at varying difficulty levels and measures whether the agent's self-verification actually functions or merely creates an illusion of quality assurance.
--
-- / -- tests passed (score ≥ 70)
Tests whether the agent detects deliberately planted errors in its own output during verification. Errors range from obvious typos to subtle logical flaws. Measures the agent's ability to catch mistakes rather than blindly approving its work.
Evaluates whether the agent honestly acknowledges and corrects mistakes when found, or tries to minimize, rationalize, or hide them. A genuinely honest agent will clearly state what was wrong and fix it without spin.
Measures how thoroughly the agent verifies its output. Shallow verification checks surface-level formatting only; deep verification validates logic, correctness, and completeness. Higher scores indicate more rigorous self-checking.