Separating Verification-Regime Effects from Verifier Stochasticity in Coding-Agent Evaluation

LLM-generated tests and behavioral probes are increasingly used as measurement instruments in coding-agent evaluation. This preprint studies whether candidate-patch visibility changes verification outcomes beyond the verifier's own stochastic self-disagreement. In a prospectively frozen Phase-0 study, four fixed coding-agent submissions were evaluated on 20 SWE-bench tasks (80 task-candidate cells). Three Candidate-Blind verifier artifacts were generated per task and reused across four candidates, while three Candidate-Conditioned artifacts were generated per task-candidate pair. Seven aggregate verdicts changed across regimes. Replicate-level between-regime disagreement was 0.1236, compared with a within-regime stochastic baseline of 0.1208, yielding excess disagreement of 0.0028 and a ratio of 1.023. On the 16-task common-determinate subset, the largest candidate score change was -0.0625 and no strict pairwise ranking reversal occurred. The prospectively frozen decision was STOP. The central methodological implication is that same-regime replication is needed as an identification baseline before cross-condition disagreement is interpreted as a treatment effect when stochastic LLMs generate verification evidence.