Verifier

A check that evaluates one stated property of an AI output, such as schema validity, test behavior, source support, policy compliance, or a rubric score.

A verifier turns an acceptance criterion into an observable result. A schema validator can establish that required fields have the right types. A test can exercise a code path and compare the result with an expected outcome. A source check can determine whether a claim is supported by a cited passage. Model-based graders can judge properties that resist simple rules, but their scores remain probabilistic and can share the generator's blind spots. Each verifier covers the property it tests; passing one does not certify the entire output.

Builder example

Verifiers let a workflow route results instead of treating every generation alike. A failed schema can trigger deterministic repair, a failed test can send the error back to the coding agent, and an uncertain policy judgment can escalate to a specialist. Generator and verifier quality interact, so compare complete systems on representative cases rather than assuming a weaker generator plus checks or a stronger generator alone will win.

Common confusion: Verification is not synonymous with safety or truth. A test suite can omit an important behavior, a rubric can encode the wrong preference, and a second model can confidently agree with the first. Combine independent checks around the actual failure modes and measure both false passes and false failures.