Black-Box Auditing of Epistemic Reliability in Multi-Agent Debate Distillation

Despite theoretical promise, debate as a scalable oversight protocol has produced mixed empirical results: gains in some settings, and null effects in others, especially when the judge does not have information hidden from it. We study proposer-critic debate in a stronger-debater/weaker-judge setting on programmatically verifiable code and logic tasks. Debate helps the judge over a consultancy baseline when the critic provides a usable advantage: the critic's classification ability must exceed the judge's, and the judge must treat critic speeches as claims to verify rather than testimony to summarize. On the three of five pairings where the condition holds, proposer-critic debate's gains are statistically significant over consultancy, and these pairings are the most capable model pairings. On the two non-responder pairings in our set, debate produces null effects, and judge verification rates drop by tens of percentage points once a critic enters the transcript. In these cases the critic's binary-classification ability and the judge's are within noise of each other, and the critic's disagreement is parsed as testimony rather than a claim to check. Ablating rebuttal rounds from debate produces no measurable change in judge performance: a single independent critique recovers the bulk of debate's benefit at lower inference cost. These findings suggest a cheaper primitive for training-free scalable oversight in verifiable domains (answer, critique, judge) and a pre-deployment audit (does the critic beat the judge, and will the judge verify it?) that predicts when debate will help.

As AI agents surpass human performance, it becomes exceedingly hard for system designers to evaluate them directly and understand their failure modes. Consequently, agents themselves are being deployed extensively to evaluate, judge, and provide feedback on model traces. But this raises an important question: how can we trust the judge? In this work, we propose an agent-guided method to find weaknesses of agentic judges that expose interpretable failure mechanisms. Our method focuses on mathematical reasoning and proceeds in two stages: first, we deploy adversarial agents to mutate a set of sound proofs by introducing errors, attempting to misguide judges---in other words, injecting errors that judges are unable to catch. Then, we distill these attempts into a small set of mutation strategies which allow us to analyze the failure modes of the judges. To ensure that these strategies are not overfit to the initial set of proofs, we evaluate them by applying the mutation strategies to a held-out set of proofs and querying the same judge. We deploy our method on GPT-5.6-sol and Claude Opus 5, paired with their agent orchestrators, Codex and Claude Code, respectively. These are used both as mutators to introduce errors and as judges to evaluate correctness of mathematical reasoning. We find that across all the agentic judges, we are able to distill mutation strategies that consistently bypass their evaluations, thereby enabling us to ascertain actionable failure modes. Our analysis also reveals that judge reliability degrades at the frontier: errors in Olympiad-level proofs or graduate-level mathematical texts are detected more consistently, whereas flaws in research-level manuscripts are more likely to escape detection.

We introduce AgentRelBench, an environment-agnostic reliability instrument that computes ground-truth, severity-priced damage from database state diffs across repeated runs, with no LLM in the measurement path, demonstrated on EnterpriseOps-Gym. Across 2,128 evaluation runs spanning nine models in six families (four development, three pre-registered held-out, plus a frontier pass on two frontier-tier models that the pre-registration designates exploratory), we find: (1) damage on irreversible actions is universal across the families we measured and stochastic within them on pinned, single-provider stacks. (2) No task damaged on every run: zero always-fail cells across 42 confirmatory held-out damage events. A single clean run misses a damage-producing (model, task) pair 0.80 of the time on the development pool (13 pairs); the held-out pool is descriptively consistent (0.575 over 5 pairs, pair-weighted) but sits below our pre-registered power floor and is reported as underpowered, not as confirmation. (3) Damage-producing task count falls with model capability, from 7 of 20 tasks for an 8B model to 1 of 20 for the most capable; capability is confounded with family and training, so this is an observed gradient, not a causal claim. The residual damage does not change in character: in the exploratory frontier pass, the most capable model's one damaging task damages at $\hat{p} = 0.16$ per run, inside the same demonstrably-stochastic band, and a single audit misses it 84% of the time. (4) One model family committed the gated irreversible change while declaring it had refused: transcript- and judge-based grading scores those runs as safe refusals, only state diffs as damage. All confirmatory findings were pre-registered with per-claim demote criteria; one demoted our own initially favored finding, which we report.