Published August 1, 2026
| Version 0.2
Preprint
Open
Agreement Is Not Truth: How LLM Juries Converge Confidently on Legally False Answers
Description
Using a panel of Large Language Model judges to score and select answers — "LLM-as-jury", increasingly the default stopping criterion for agent loops and multi-model consensus pipelines in reward-free domains — rests on an unexamined assumption: that when diverse judges agree, their agreement tracks correctness. We test this on a legally-anchored gold set (N=16 Brazilian tax and labor cases, good/bad labels grounded in verified official statutes, not opinion), scored by three judges from distinct model families (GLM-4.6, GPT-4o-mini, Llama-3.3-70B). We report three findings. (A1) The panel discriminates well overall — accuracy 0.938, Youden 0.875 — and beats every single judge scored on the full set (best solo 0.875 over n=16), confirming the classic ensemble benefit. (A2) Yet the panel remains blind or inconsistent on exactly the two most dangerous error types in law: it gave a top score of 10.0 to an answer that applied real statutes to the wrong case (wrong subsumption), and nearly passed (9.0) an answer that cited a statute which does not exist (non-existent norm). The failure is systematic, not stochastic: both error types are lexically fluent and legally plausible, and plausibility is what an LLM judge measures. Critically, judge diversity does not rescue the panel, because the three families share a training corpus (the legal web) and therefore share the same blind spots — architectural diversity reduces idiosyncratic variance but not corpus-level correlated error. We conclude that jury agreement is a valid proxy for form and completeness but not for factual truth, that the top-compression of LLM judges pushes the admissible acceptance threshold against the ceiling of the scale — so far that the threshold is not identifiable (any > 9.0 induces the same partition; see AL-P4) — and that any pipeline delegating a correctness decision to a jury inherits a floor of confident, high-scoring error. The practical consequence: factual verification must be an external gate, not a jury vote. : LLM-as-judge, LLM jury, multi-model consensus, ensemble evaluation, reward-free evaluation, confident error, hallucination detection, legal NLP, calibration, Youden threshold, corpus correlation, agent loops.
Errata (2026-08-01). Version of 2026-08-01. Errata: the acceptance threshold reported in v0.1 as a point estimate (theta* = 9.5) was an artefact of a search grid that stopped at 9.5 and never evaluated 10.0, which yields an identical confusion matrix. No observed score falls in (9.0, 10.0), so the threshold is not identifiable on these data; the admissible region is '> 9.0'. Perfect sensitivity is likewise an arithmetic consequence of ceiling saturation (all positives tied at the scale maximum), not evidence of discrimination. The previous version remains permanently resolvable at its own version DOI.