Figure 1. Exact full-coverage selection success for two preexisting candidates, exactly one correct with equiprobable correct index. Each query of each judge on each candidate costs one abstract unit; generation is excluded. Both judges share false-accept probability 1/10 and false-reject probability 1/5. The three categorical laws differ only in within-candidate joint errors; observations are conditionally independent across candidates. No-query success is 50%; known-law Bayes selection after A costs 2 and succeeds 85%; after A+B it costs 4 and succeeds 96%, 94.1%, or 85%. Points are discrete authored calculations, with no interpolation or empirical scaling law. Full coverage makes selected and unconditional success coincide.

Figure 2. Known-law joint-minus-single information gain for the same laws: 11, 9.1, and 0 percentage points. This compares cost-4 joint observations to cost-2 single observations, before any utility penalty; it is not a cost-matched net-benefit estimate. Class-conditional error correlations (incorrect, correct) are (-1/9,-1/4), (0,0), and (1,1). They are not one common scalar dependence coordinate. There is no sampling uncertainty for these exact specified laws; no confidence intervals are implied.

Both figures are synthetic finite-law illustrations, not learned-policy performance, model measurements, installed-serving parity, or H1 study results. Minimum feasible joint error is a categorical extremal law, not a universal optimal verifier construction.