Can a quorum of validators really provide independent judgment?

The Honest Quorum Problem (2607.16109) does not stop at saying one validator can be wrong. It says several validators can be wrong for the same reason. The abstract lists what agentic validators "may share": model weights or lineage, training distributions, prompts, retrieval sources, toolchains, evidence, reasoning scaffolds, and provider infrastructure. The result, in the paper's phrase, is correlated epistemic faults, and the practical upshot is that a quorum's votes need not be independent.

The list is worth reading as a list. Model family, the axis most multi-model designs diversify, is the first item and the most visible one. The other seven are quieter and are usually chosen once for the whole system: one prompt template, one retrieval index, one tool wrapper, one cloud region. The paper's own example shows how ordinary this is. The validators there "inspect the same canonical request, state snapshot, policy context, and evidence package." Even validators built on different models would share that last item, so a misleading evidence package is a common cause by construction.

Why correlation matters more for a quorum than for a single reasoner is this note's reading, not something the excerpt spells out. A quorum threshold counts faults: it tolerates up to some number of bad participants and relies on any two quorums overlapping in a good one. A shared cause turns that count into an event. One flawed policy summary, one blind spot in a training distribution or one bad retrieval hit can push many validators the same way at once, and the votes that result look like independent confirmation while adding little beyond the first. Agreement then measures how alike the validators are, which is why Can a quorum of honest validators certify an invalid transition? cannot be solved by raising the quorum threshold alone. The coordination-defence work meets the same confound from the other side: How do we tell coordination apart from shared causes? says agents behaving alike may share a model, instructions or environment and not influence one another, and validators voting alike may likewise share a cause and not confirm one another.

This is the failure side of two vault findings about diversity. Can models trained on many imperfect experts outperform everyone? makes uncorrelated biases an explicit condition for aggregation to help: majority voting denoises only when the errors are diverse rather than correlated. Can a panel of smaller judges outperform one large judge? gets its gain by drawing judges from disjoint families, which addresses the weights-and-lineage and training-distribution channels. The Honest Quorum list says that is two of eight; prompts, retrieval, evidence, scaffolds and infrastructure can still be common.

The claim is conditional. The paper says validators may share these things, and the excerpt gives no measurement of how strongly any channel correlates faults in practice. That gap is the subject of Does model diversity actually reduce validator agreement failures?.

What the excerpt does not give. No correlation estimates, no experiment, and no statement of how the paper's thresholds account for correlation. Only the list of shared elements and the conclusion that faults are correlated are in the excerpt.