Does model diversity actually reduce validator agreement failures?

The Honest Quorum paper (2607.16109) says agentic validators may share model weights or lineage, training distributions, prompts, retrieval sources, toolchains, evidence, reasoning scaffolds and provider infrastructure, and that this yields correlated epistemic faults (Can a quorum of validators really provide independent judgment?). The cheapest remedy anyone reaches for is to use models from different families. The vault has a positive result for that remedy in a neighbouring setting: Can a panel of smaller judges outperform one large judge? reduces intra-model bias by drawing judges from disjoint families. The vault also holds a caution from open-ended generation: Do different AI models actually produce diverse outputs? finds different models producing strikingly similar outputs and points to shared alignment and training priors, which suggests the channels family diversity acts on may themselves be only partly diversified. That study measures generation on questions with no single correct answer, not approval of transitions, so it bears on this question by analogy.

What is unknown is how far that carries to validators approving state transitions. Three sub-questions stand out.

1. Which channels dominate? Family diversity acts on weights and training distributions. If most correlated error runs through the prompt, the retrieval index or the evidence package, a family-diverse panel could still fail together. The lineage channel has one measured instance elsewhere in the vault: Do frontier AI models favor their own company? reports an own-company tilt whose direction differs by family, on single models answering and grading and not on validators. Does grading expose company bias that answering hides? leaves open whether the grader role itself surfaces the tilt, which matters here because a validator is closer to a grader than to an answerer.
2. Can some sharing be removed at all? Validators asked to judge the same transition have to see the same request. The paper's example gives them the same request, snapshot, policy context and evidence package. Diversifying evidence means having validators gather it independently, which costs more and may itself be a source of disagreement about facts rather than about the transition.
3. Do the channels add up? Whether removing each shared channel lowers correlation roughly independently, or whether one shared channel sets a floor, is not something the excerpt says.

One way to settle it, which is this note's proposal and not the paper's: assemble a labelled set of valid and invalid transitions in one domain, run validators that differ on exactly one channel at a time, and compare pairwise error correlation across the variants. If varying family barely moves the correlation while varying evidence source moves it a lot, the eight-channel list is telling us where to spend effort. The monitoring side of the vault states the same separation problem and names an intervention of the same kind as one option: How do we tell coordination apart from shared causes? frames closing a channel and seeing whether shared behavior returns as an interventional route to telling a shared cause from influence.

Until something like that is measured, the safe reading is the paper's: a quorum's size and its model mix are not evidence of independence, and the scope statement on any guarantee (What scope should a validator's statistical guarantee actually state?) should say how independence was established.

What the excerpt does not give. Any empirical correlation between validators, or any result for a diversified quorum.