Method
Feynman rests on an old result: judges who decide independently, each right more often than not, are almost always right together. Here are the arithmetic, a simulation you can run, the stages of a review, and every paper behind the approach, including the ones against it.
Method · in six rows
Asking more people only helps if they can be wrong in different ways. Copies of one AI tend to make the same mistakes, so their agreement tells you little.
Feynman’s lenses run on one model too, so they don’t fix that by themselves. What they do is put different questions to an idea, say which lens raised each objection, and check every citation. Whether that beats one strong AI asked once is the comparison we have locked in advance.
From your memory to your memory: two model calls per review, however many lenses take part. The rest is checking.
Past reviews like this one are scored in your browser. Recorded failures count for more, and only the matches travel with your idea.
In your browserThe idea is split into load-bearing claims, each typed: empirical, causal, mathematical, market…
1 model callOpenAlex is queried for the literature on each claim, with titles, DOIs, citations and abstracts.
On our serverA second pass looks for retracted work on the topic: failures that already happened.
On our serverThree lenses attack the claims, citing only the records retrieved. One verdict, ranked holes, one killer test.
1 model callAny citation that points outside the retrieved records is stripped, and the card says so.
On our serverThe review is kept in your browser. Mark what happened later: held up, failed or revised.
In your browserIn 1785 Condorcet showed that a majority of independent judges, each right more often than not, becomes almost certain as the group grows. The word that matters is independent. Copies of one AI share their mistakes, so adding more of them adds votes, not evidence. And six voices that defer to one are, in effect, one.
If every voice erred independentlyWith shared error ρvoices in the panel →
Held to our own arithmetic
Feynman’s lenses run on one model. By this arithmetic they are not independent judges: they share the model’s training, so their errors can move together. They are different questions put to one model, and each hole says which lens raised it. Whether that finds more real holes than one strong model asked once is the comparison we have pre-registered.
Illustrative models. Shared error: with probability ρ (rho) the whole panel shares one judgement; otherwise each voice errs on its own. Perspectives: the Hill–Shannon effective number of voices, which counts how many distinct voices a room really has. Neither is a measurement of any product, ours included.
Thirty-six people search a hard problem, each a row of twelve choices. On the left, everyone takes advice from one shared AI. On the right, each consults one of six different advisors. Run it and watch the left side turn into one answer.
Running 120 paired studies in your browser…
A toy model: our simplified re-implementation of the mechanism studied by O’Brien, Rubin, Wong & Fazelpour (CogSci 2025), not their model or results. The plural advice includes the shared recommendation by construction. On a smooth problem (ruggedness 0) the shared AI costs nothing; the price rises as the problem gets harder.
Every paper we rely on, when it appeared, what it found, how we use it and where the inference stops. 17 in all, 2 of them arguing against us.
X. Zhu, C. Zhang, Y. Chi, T. Stafford, N. Collier & A. Vlachos · 2026
That a Feynman review finds more real holes than one strong model asked the same question. That comparison is next, and we will publish it either way. Its design was fixed on 9 November 2026, before any run.
Not runThat the lenses reason as their sources do. Each is drawn from cited works and held to its horizon, but no one has yet measured how faithfully the model keeps to them.
Not measuredThat OpenAlex covers your field. Market, legal and very recent evidence is thin; “reasoning only” means we found nothing, not that nothing exists.
Uneven