Decorrelation, or just competence? Three experiments on AI judge panels

The honest-eval tool this site is built around has one load-bearing rule: the model that does the work never grades its own work, and the graders come from different model families. The bet behind that rule is decorrelation. Independent models fail in independent ways, so a panel drawn from different families should catch mistakes that any single family, grading alone, would wave through.

I believed that bet. So I spent three experiments and about seventeen dollars trying to measure it, and the answer turned out to be more deflating, and more useful, than the slogan. Decorrelation does help. But most of the help is not what I thought it was.

All three runs share one corpus: ninety short explanations with a single planted, checkable error in sixty of them and thirty left clean. The errors were written by six different model families in rotation (so no single author’s idea of “a plausible mistake” dominates) and each one was confirmed wrong by a different family before it went in. Every measurement below was pre-registered before the data existed. The scoring is deterministic: an answer key says what the planted error is, and a flag only counts if it names that error, not a different one.

Act one: on a checklist, decorrelation does nothing

The first study graded on a deterministic floor, a fixed set of required points an explanation had to cover. I compared a self-family critic panel against a count-matched cross-family panel. My pre-registered prediction was that the cross-family panel would win.

It did not. Adding a second independent critique pass helped (a real, repeatable gain). Making that second pass come from a different family added nothing measurable. When I later re-ran the study with different models in the doer’s seat, even that result turned out to be specific to one model, not general. I shipped a correction to the original note saying so.

In hindsight this is obvious. A checklist is a checklist. Every judge sees the same required points, so there is no shared blind spot for diversity to break. I had tested decorrelation in the one place it was guaranteed to come up empty.

Act two: on judgment, decorrelation catches what one family misses

So I moved to a task where independent failure modes are the whole game: catching planted errors. Here a “family-blind” error is one a given family systematically misses while others catch it. The prediction was that a decorrelated panel would out-perform a count-matched same-family panel on exactly those errors, measured as discrimination (catch rate on errored texts minus false-alarm rate on clean texts, so a panel cannot win just by crying wolf).

It did, clearly. With three local models as judges, the decorrelated panel beat the same-family panel by about +0.40 discrimination on family-blind errors, with the interval excluding zero. With three frontier models it was larger, +0.73 to +0.77, again interval-excluding-zero. This is the result the checklist could not produce: on a judgment task with shared blind spots, a decorrelated panel really does catch errors a single family misses.

I almost stopped there. It would have made a clean story. But the per-judge numbers had a smell to them. Among the “frontier” judges, catch rates ran from 0.77 down to 0.29. The panel was not three comparable models. It was a strong one carrying two weaker ones. Which raises the real question: does a decorrelated panel help because it is diverse, or just because it happens to contain a judge that is good? Three copies of a weak judge cannot rescue a systematic miss. One competent panel member can. Those are very different claims, and Act two could not tell them apart.

Act three: it is mostly competence

To separate the two I needed judges spread across the competence range, so I could ask whether the decorrelation benefit survives once competence is accounted for. I graded six different vendor families (OpenAI, Google, Anthropic, DeepSeek, Qwen, Mistral) on the same corpus, then for each family measured how much a decorrelated panel of three other families beat three copies of that family, and regressed that benefit against the family’s own competence. Both panels are the same size, three judges, so panel size is held constant and cannot masquerade as a diversity effect.

The picture is unambiguous in one direction and faint in the other.

The benefit tracks weakness almost perfectly. The regression slope is strongly negative (about −0.8). A weak judge gains a lot from a panel of others. A strong judge actually loses to a panel of others, because the others are weaker than it is. The decorrelation “win” is, to first order, the cost of having let a weak model grade alone.

The residual, the benefit left over at average competence once that slope is removed, is small: roughly +0.04 to +0.05 discrimination. It clears zero under a majority-vote rule (interval [0.001, 0.087]) and does not under an any-flag rule (interval [−0.018, 0.116]). So a genuine diversity effect, independent of competence, may exist, but it is small and it is fragile.

A later audit asked a fair question about this residual: is it just a side effect of panel size? If the decorrelated panel quietly carried more judges than the same-family panel, it could catch more errors simply by having more votes, and that would have nothing to do with diversity. I went back and checked, at no extra cost, by re-running the regression on the same gradings with the two panels pinned to the same size (three decorrelated judges against three same-family samples). The answer is that the contrast was already size-matched, so the numbers do not move: the residual comes back at +0.05 under any-flag and +0.044 under majority-vote, exactly as above. To be sure the check had teeth, I also ran the deliberately mis-matched version, a five-judge decorrelated panel against a three-sample same-family panel. That one does inflate the majority-vote residual, to +0.10 with an interval clear of zero, which is the artifact the audit was right to worry about in general. It just was not the version that produced the result here. So holding panel size constant leaves the conclusion where it was: the leftover diversity effect is small, fragile, and aggregation-dependent, not a counting trick. The reanalysis is in experiments/blindspot-detection/reanalysis_count_matched.py.

What I think this means

A decorrelated judge panel is worth building, but mostly for a defensive reason rather than the romantic one. Its main value is insurance: it stops any single weak or idiosyncratic judge from quietly deciding the outcome. That is real and it matters, because in practice you often do not know which of your judges is the weak one. What the panel mostly buys you is “do not let an all-weak or single-judge setup grade your system,” not “diversity reveals truths no individual model can reach.” The second, stronger story is at best a small effect hiding under a large competence effect.

For the tool, the rule stands, with a clearer reason. Exclude the producer and use a cross-family panel, not because diversity is magic, but because it caps the damage a weak grader can do and you rarely know in advance who that is. Pick competent judges first. Decorrelate second.

What this does not say, and what bit me along the way

Six families is a small sample for a regression, and the residual effect’s significance flips between the two voting rules, so I would not bet much on the exact size of the leftover diversity effect. Everything ran on one planted-error corpus; naturally occurring errors may behave differently. Self versus other recognition (does a family miss its own family’s errors more?) was not cleanly estimable here and deserves its own study.

The process was not clean either, and that is part of the point of an honesty tool. The grader’s confidence interval was computed wrong on the first pass and a cross-model review caught it. A smoke test caught the weak judges emitting the error category instead of the actual error. The live run crashed twice on malformed model output before the parser was made forgiving. The first attempt to add statistical power by adding judges actually removed it, because a stricter “everyone else agrees” rule almost never triggers once there are six judges. Each of those was caught by checking rather than trusting, which is the same principle the panel is built on, turned back on the work itself.