Decorrelated reviewers still share the field's prior, so read their findings by the claim's stance

Type: kb/types/note.md · Tags: foundations, methodology

Decorrelated review — a fresh critic, a clean-context judge, a second model — removes the errors that came from the author's own reasoning trace. It does not remove the errors that come from the field. Judges drawn from one training distribution, or one discipline, carry the same consensus prior, and a verdict is evidence weighed against that prior. On the component of a claim the prior determines, N such judges are not N independent draws. Error correction works with above-chance oracles and decorrelated checks; a shared prior is a correlated error term that decorrelation of the proposer leaves untouched.

The consequence is not uniform. Where a claim uses a source on its consensus reading, the judges agree with the author and the review is cheap. Where a claim contests the consensus reading, the judges converge on the consensus and propose the reversal. Review pressure therefore concentrates on the most original claims, and a vote cannot tell "original" from "wrong" there — both look the same to a judge whose prior is the field's.

Record the stance, then read the findings by it

Because reviewers cannot reconstruct which claims the author means to defend — that is purpose, held upstream and not derivable from the artifacts, since intent-framed delegation is a control regime — the stance has to be written down before the review runs. It is the author's key for reading the verdicts, not an instruction to the judges: the judges see the artifact and nothing else, so their findings stay decorrelated from the author's purpose, and the stance is applied only when the findings come back. Two stances suffice:

  • Divergence. The author contests the consensus reading and will pay for the claim. Pushback is predicted. A reversal is accepted only when it supplies a new object-level argument — a premise the author's argument lacked, a counterexample it did not survive. A reversal that restates the consensus conclusion is the predicted event, and it is not evidence.
  • Support. The source or claim is used on its consensus reading because of what it supplies. Reviewer narrowing is accepted by default: nothing the author defends lives there, so the narrowing costs no content the author cares about.

Triage then runs stance-first: at a support claim, accept a finding unless it is factually wrong about the source; at a divergence claim, ask whether the finding adds a premise. This is a rule for allocating the check, not for deference. Divergence claims are not more likely to be true because they are original; they are the claims for which plural judgment is uninformative, so the remaining checks — an object-level refuter stated in advance, and later exposure to evidence the author did not produce — have to carry the weight.

Seen from the repair side, this is why narrowing bought to survive review is paid for in content: review pressure pushes hardest exactly where narrowing costs most. Stance makes the two cases separable. Accepting narrowing on a support claim is ordinary repair; accepting it on a divergence claim is the loss that note describes.

Review as reconnaissance: engage or deflect

The stance list is written before review, but it is not complete before review. Decorrelated judges also serve as reconnaissance: their findings map where pushback currently lands, and the map includes battles the author had not declared — a support claim that turns out to carry weight, or a divergence the author did not know the work contained. So the findings are read twice. First by stance, as above. Then, for each finding that lands on something the stance list did not settle, by a choice:

  • Engage when the claim is load-bearing — the purpose, or the claims and citers that depend on it, cannot survive its loss. Engaging means paying: an object-level argument, a stated refuter, and exposure to evidence the author did not produce. The claim joins the divergence list.
  • Deflect when it is not. Deflection is legitimate and has several moves: narrow the claim to the subject the argument actually supports, concede the consensus reading where nothing downstream depends on the contested one, reframe so the pushback lands on a claim the author is willing to defend, or defer with a named open question. Each buys the review's agreement at a price the author has decided is affordable.

What decides between the two is what depends on the claim, not how much the author likes it. The citer test from the narrowing note is the cheapest instrument: what inbound uses import from a claim is what its loss would cost. A claim nothing imports can be deflected at no cost; a claim the purpose runs through cannot be deflected at any price, and pretending otherwise is the analytic narrowing that note describes. The stance list is then updated with what the reconnaissance found, so the next review round is triaged against a map that has already absorbed this one.

The misreading: telling the judges

The first draft of this note was titled "…so declare which claims you will defend," and it was wrong in a way the note itself predicts. Read as an instruction to the judges, the stance would travel into the review: a judge told that a claim is a divergence defers to it or aims at it, and either way its verdict is now conditioned on the author's purpose. That re-correlates the judge with the author on exactly the component the technique exists to keep separate, and the review stops being a check on the divergence and becomes a negotiation about it. The temptation is real because the two uses share one act: the author writes the stance down. What differs is who reads it. The judges read the artifact and nothing else; the author reads the stance when the findings come back. The declaration is a filter on the output of review, never an input to it.

A witness

A clean-context audit of a source register in this repository, run in 2026, sorted its findings this way without being told to. It narrowed four support rows — an organizational-learning source, a philosophy-of-science source, a function-allocation taxonomy, a human-factors precedent — on grounds faithful to the ingests, and every narrowing was taken. It also proposed reversing the two rows where the series contests the consensus reading of its sources, Naur's human binding and the Bitter Lesson's weights-only extrapolation, by reading the consensus conclusion as the source's generating condition. Neither reversal added a premise; both restated the consensus. They were declined as predicted, and the register now carries the stance per row so the next audit is triaged the same way.

Guards

The stance list is itself capturable: declare every claim a divergence and the work is immune to review. Two guards keep it honest. A divergence claim must carry its object-level argument and a stated refuter in the artifact — it cannot be defended by stance alone. And the list should be short; a program with many divergences has not chosen its battles, and each undeclared or over-declared divergence is a place where review either bites where it should not or is waved away where it should bite.

Scope

  • The claim holds where the judges share a training distribution or a disciplinary consensus. Judges from genuinely different traditions decorrelate part of the prior; how much is an open question below.
  • Loading the author's theory into the judge's context reduces convergence on the proposer's blind spots and may reduce convergence on consensus, but the prior sits in weights, not in context, so this is a partial remedy at best.
  • The rule presupposes a declared program. Where the author has no stated purpose there is no stance to declare, and review pressure toward consensus operates unchecked.
  • Human peer review has the same structure; the note is stated for agent-run review because that is where the mechanism was observed and where the remedy is cheap to install.

Open Questions

  • Whether judges from different training corpora measurably decorrelate the consensus prior, or whether the corpora overlap enough that model diversity buys little at divergence points.
  • Whether a judge with the author's argument loaded reverses divergence claims less often than a bare judge, holding the finding's factual quality fixed.

Relevant Notes: