verifier-module

Answer verification audits a candidate against the original problem by extracting and checking claims for factual accuracy, logical support, and completeness to catch issues before acceptance. It decomposes claims into subclaims and classifies each as verified, contradicted, or uncheckable.

A conservative verdict contract blocks full approval when key checks are missing or claims remain uncheckable, reducing hallucination-driven false positives. It requires exactly one label (解决, 部分解决, or 未解决) to support consistent downstream decisioning.

To audit an answer, load the verifier role, provide the original problem statement plus the candidate answer, and ask it to output a single verdict label with a detailed verification report. The report lists defects, omissions, contradictions, and names checks run or blocked.

Yes, agent evaluation applies to solution review scenarios where answers may include text, plans, analyses, or patches, and where task materials and runnable checks are available for independent verification of factual accuracy and logical consistency.

When key checks are missing or claims remain uncheckable, conservative acceptance logic blocks full approval to reduce false positives. The verification report details which checks were run, which could not be completed, and any resulting defects or omissions.