A group of AI checkers can follow every rule and fully agree with each other, yet still sign off on a wrong answer.

This explores why a group of AI validators can all agree with each other, and follow the protocol perfectly, while still endorsing an answer that's wrong. The corpus's answer is that agreement and correctness are two different kinds of guarantee. In the Honest Quorum analysis, agreement follows from the protocol's rules alone, so it holds deterministically. Semantic validity, meaning the answer is actually right, depends on statistical assumptions about how validators behave, and the protocol has no way to enforce those Can validator consensus guarantee both agreement and semantic correctness?. A quorum can authenticate correctly, sign the expected messages, reach consensus, and still endorse an invalid transition. In the paper's words, agreement proves the protocol ran, not that the answer is right Can a quorum of honest validators certify an invalid transition?.

The hidden assumption is independence. Voting only counts as confirmation if each validator errs on its own. The Honest Quorum paper lists eight things validators might share, from model weights and training lineage to retrieval sources and provider infrastructure. One shared fault can then hit many validators at once, and the vote counts this as several independent confirmations when it is really one mistake repeated Can a quorum of validators really provide independent judgment?. Swapping in different model families looks like the obvious fix, but it only addresses two of the eight channels. Errors can still correlate through shared prompts, retrieval, and infrastructure, and the paper suggests measuring this rather than assuming diversity works Does model diversity actually reduce validator agreement failures?.

The corpus has a second, more basic finding: LLM agents often can't reach agreement at all. In hundreds of simulations, groups failed mainly by stalling or timing out, not by quietly corrupting a value, and reliability dropped as groups grew, even with no adversaries present Can LLM agent groups reliably reach consensus together?. So the guarantees are weak on both sides. Agreement is fragile in practice for LLM agents, and when it does happen it says little about truth. Making agreement detection better helps with stalling and premature convergence Can AI systems detect when they've genuinely reached agreement?, but a debate that settles cleanly can still settle on the wrong answer.

A further wrinkle is that treating disagreement as noise can throw away useful information. When agents share the same facts but reach different conclusions, that split often marks territory where reasonable values conflict. Forcing consensus there erases the signal that a human should be looking at the case Can disagreement in reasoning traces signal legitimate value conflicts?. Training has a similar effect. Optimizing models toward a single verifiable answer makes them worse at predicting where people legitimately disagree Why do reasoning models fail at predicting disagreement?. A validator set built from such models may be more uniform, and therefore more agreeable, without being any more correct.

The practical lesson is that a consensus certificate should be read as a statement about process, and semantic correctness needs evidence from somewhere else. That evidence could be measured error correlation between validators, deliberately varying prompts and sources as well as models, or a route to human review when validators split. Consensus tells you the machinery worked. It says nothing on whether the validators were independent enough for their agreement to count as evidence.