When several AI checkers all agree, are they really checking independently — or just repeating the same hidden mistake?

This explores what you would have to observe to conclude that several AI validators agreeing on a verdict are checking independently, as opposed to all echoing something they share upstream. The corpus's first answer is that agreement can't tell you. Can a quorum of honest validators certify an invalid transition? shows that a quorum can sign the right messages, reach consensus, and still endorse an invalid result, because agreement proves the protocol ran, not that the answer is right. Can a quorum of validators really provide independent judgment? explains why: validators can share up to eight inputs, from model weights to provider infrastructure. One fault then hits several validators at once, and the vote counts it as several independent confirmations.

So the evidence has to come from causal tests, not vote counts. How do we tell coordination apart from shared causes? describes two kinds. Observational evidence looks for signs that something was actually transferred between agents. Interventional evidence closes a suspected channel and checks whether the shared behavior comes back. For validators, this means changing one shared input for one validator, such as its retrieval source, and seeing whether its mistakes stop matching the others. If they stop matching, that input was a common cause. If they keep matching, the cause is somewhere else. You would only call validators independent after every channel has been varied and their errors still line up no more than chance would predict.

Does model diversity actually reduce validator agreement failures? turns this into a concrete design: measure error correlation across validators that differ on one channel at a time. It also warns against the obvious shortcut. Using different model families touches only two of the eight channels, so a GPT, Claude and Gemini panel can still share prompts, retrieval sources and infrastructure. A test that varied only the model would certify independence that isn't there. The correlation that matters is in the errors, not the correct answers, so you need cases where you know the right answer. That is why Can validator consensus guarantee both agreement and semantic correctness? matters here. The protocol guarantees agreement by itself, but semantic correctness holds only under statistical bounds on validator behavior that the protocol can't enforce. Independence is therefore an empirical claim that someone has to measure.

Three neighboring ideas suggest cheaper signals, though the corpus doesn't test them on validators directly. First, look at reasoning as well as verdicts. Where do reasoning agents actually fail during long traces? finds that most failures are process violations visible in intermediate steps. Two validators making the same wrong intermediate move would then be a stronger fingerprint of a shared cause than matching final answers. Second, look at whether they ever disagree. Can disagreement in reasoning traces signal legitimate value conflicts? argues that trace divergence carries real information, so a panel that never diverges, even on contested cases, deserves suspicion. Third, borrow the record-keeping idea from Can infrastructure evidence replace terminal scores in benchmark validation?, which grounds claims in recorded infrastructure evidence. Logging which weights, prompts, retrieval sources and provider each validator used would let you audit the shared channels directly instead of inferring them from behavior.