Definition
A repeated execute, observe, compare, and correct cycle that withholds completion until an attempted result satisfies explicit evidence or acceptance criteria. A useful implementation separates the producer from the observation or grading step, records false alarms as well as caught defects, and limits retries when the same failure recurs.
Each pass should add information: a test result, comparison with a source, invariant violation, human decision, or other observation not available to the original attempt. Asking the same model to reconsider the same context may improve an answer, but it is weaker than verification when no new evidence enters the loop.
Bound the loop with retry and cost limits, a repeated-failure detector, and an escalation path. Track verifier false positives as well as caught defects because a bad correction step can damage work that was already correct.
Distinguish it from nearby terms
A grader emits a judgment, an oracle establishes truth for a defined property, and a verification gate enforces admission. The verification loop connects those checks to another bounded attempt. A retry repeats work after failure; it becomes a verification loop only when evidence is compared with a criterion.
Check your understanding
A formatter check rejects valid code because it uses the wrong configuration. The agent rewrites the logic on every pass until the code fails. More verification attempts reduced reliability because the verifier was wrong.