Verdict Substitution Impact on Self-Consistency Verifiers

Analyzing $k=7$ self-consistency panels on MATH-500 shows that replacing a single ballot with a different-model verifier yields a $+24.2$pp pivotal gain, whereas role-reversal results in $-11.2$pp. The study demonstrates that sign-based gains are driven by the verifier's state-specific accuracy rather than global model provenance.