Proposal-judgment-split
Description
Proposal-judgment-split is the structural move of assigning candidate-generation and candidate-evaluation to separate roles rather than fusing them in one actor. The distinctive claim is why the separation helps: the two roles have opposed optima. A proposer is rewarded for coverage, recall, and creativity — for not missing good options — which biases it toward over-generation and toward approving its own output. A judge is rewarded for precision and skepticism — for not admitting bad options — which biases it toward discrimination and rejection. A single actor holding both objectives compromises between them and loses the discrimination the judge is supposed to supply. Four roles compose the shape. The proposer optimizes coverage. The judge optimizes precision. The opposed optima are the structural reason the split works — coverage and precision pull in opposite directions, so they are better pursued by separate roles than averaged inside one. The independence constraint is the load-bearing condition: the judge’s verdict must not be authored by, nor optimized to please, the proposer. When the judge is the proposer wearing a second hat, the independence is nominal and the discrimination collapses — the proposer rubber-stamps its own work. The diagnostic question — is the thing judging this candidate optimizing the opposite of what produced it, and is its verdict independent? — separates a genuine split from its near-neighbors. It is not the iterative refine loop of evaluator-optimizer (which is about compounding quality over rounds, and can run with a fused self-critic); it is not reflection (which is exactly the fused, same-agent case the split warns against); and it is not generic feedback-loop (which says nothing about opposed optima). A proposal-judgment-split can be entirely single-pass — propose once, judge once, decide — with no revision at all. The shape recurs wherever a generation process needs trustworthy gatekeeping. Adversarial legal systems separate the prosecutor (who builds the case) from the judge or jury (who weighs it). Scientific publishing separates authors from independent reviewers and a deciding editor. Machine learning separates a generator from a discriminator, and a policy actor from a value critic. Software development separates the author of a change from its reviewer. In each, fusing the roles is a recognizable failure mode — the author who approves their own pull request, the agency that polices itself — and the recurring fix is to make the independence structural.Triggers
User-initiated: User describes a setup where the same actor both produces and approves work, or proposes adding an independent reviewer/judge to a generation process. Vocabulary cues: “the author shouldn’t review their own work,” “separation of powers,” “generator and discriminator,” “actor and critic,” “generate then verify.” Agent-initiated: Agent notices a generation step whose output is being graded by the same actor (or a judge optimized to agree with the producer), and the grading has stopped discriminating. Candidate inference: “is the judge independent of the proposer, and does it optimize the opposite objective? If not, the gate may be rubber-stamping.” Situation-shape signals: A producer that passes nearly everything it generates. Self-review processes with suspiciously high approval rates. Any place where coverage and precision are both demanded of one actor.Exclusions
- The iterative refine loop — generate → critique → revise → repeat to compound quality is evaluator-optimizer. Proposal-judgment-split is about who plays which role and why they must differ, not about iterating; it can be single-pass.
- Same-agent self-critique — one actor critiquing its own output is reflection, the fused case this concept warns against. The split’s claim is that fusion degrades judgment.
- Generic feedback — any feedback-loop where output influences input. Proposal-judgment-split is the division of labor between two conflicting-optima roles, not the loop.
- Role separation without opposed optima — pipeline stages or orchestrator-workers hand-offs where the parts share one objective gain modularity but not the discrimination effect. The concept requires opposed optima (recall vs precision).
Structure
Relationships
- evaluator-optimizer — the iterative-loop foil: evaluator-optimizer compounds quality over rounds; proposal-judgment-split separates roles by opposed optima. A split can be single-pass; a loop can run fused.
- reflection — the fused-case foil: reflection is one actor judging itself, exactly what the split argues degrades discrimination.
- feedback-loop — the split specializes generic feedback by fixing the topology to two conflicting-optima roles with an independence constraint.
- doctrine — institutions (separation of powers, blind peer review, code review) encode the split as standing doctrine, making independence structural rather than a matter of individual restraint.
Examples
Montesquieu, C. de Secondat, "De l'esprit des lois" (The Spirit of the Laws, 1748), Book XI, Chapter 6 · law
Montesquieu, C. de Secondat, "De l'esprit des lois" (The Spirit of the Laws, 1748), Book XI, Chapter 6 · law
Montesquieu argued that liberty requires the powers that propose-and-execute and the power that judges be held in separate hands. In the adversarial structure he influenced, the party that builds and presses a case (the prosecution) is deliberately distinct from the party that weighs it (the bench or jury). The prosecutor’s institutional incentive is to make the strongest possible case for guilt — coverage of every incriminating argument; the judge’s incentive is to admit only what meets the standard of proof — precision in rejecting the unworthy. Fusing the two roles, so that the party building the case also renders the verdict, is the paradigmatic threat to a fair trial that the separation is designed to prevent.Inference: The legal system is the institutional archetype of proposal-judgment-split, and it makes the opposed-optima claim explicit: the case-builder optimizes for a conviction, the judge for the integrity of the standard, and the structure works precisely because those objectives pull apart. The independence constraint is load-bearing — a judge who is the prosecutor wearing a second robe is not a check at all, which is why separation here is structural doctrine, not a matter of individual restraint.
Spier, R., "The history of the peer-review process" (Trends in Biotechnology, 2002, vol. 20, no. 8, pp. 357–358) · journalism-media-studies-and-communication
Spier, R., "The history of the peer-review process" (Trends in Biotechnology, 2002, vol. 20, no. 8, pp. 357–358) · journalism-media-studies-and-communication
Spier traces the editor-mediated review model to Henry Oldenburg’s editorship of the Royal Society’s Philosophical Transactions from 1665, in which the authors who proposed findings were structurally separated from the referees who judged them and from the editor who decided. The author optimizes for getting the work accepted and for the breadth of claims it can support; the reviewers and editor optimize for the soundness and significance of what enters the record. The system’s credibility rests on the judges being independent of the author — neither chosen by the author to approve, nor identical to the author.Inference: Scientific publishing is the institution that operationalizes the split outside the courtroom, and it exhibits the same independence constraint: a journal in which authors reviewed their own submissions would retain the form of review while losing its discriminating function entirely. The role-separation is single-pass in its core form (propose, judge, decide) — distinct from a revise-and-resubmit loop, which is a separable add-on, not what makes the split work.
Barto, A. G., Sutton, R. S., & Anderson, C. W., "Neuronlike adaptive elements that can solve difficult learning control problems" (IEEE Transactions on Systems, Man, and Cybernetics, 1983, vol. SMC-13, no. 5, pp. 834–846) · computer-science
Barto, A. G., Sutton, R. S., & Anderson, C. W., "Neuronlike adaptive elements that can solve difficult learning control problems" (IEEE Transactions on Systems, Man, and Cybernetics, 1983, vol. SMC-13, no. 5, pp. 834–846) · computer-science
The actor-critic architecture splits a reinforcement-learning agent into two separately-optimized components: an actor that proposes actions (a policy, optimizing for selecting actions that lead to reward) and a critic that evaluates them (a value function, optimizing for accurate estimation of how good a state or action actually is). The critic’s estimate drives the actor’s updates, but the two are distinct learners with distinct objectives — the actor is trying to act well, the critic is trying to judge accurately. The separation lets each be trained on the signal appropriate to its role rather than collapsing both into one objective.Inference: Actor-critic is the clean machine-learning instance of role-separation by opposed function. The actor’s job is generation (which action), the critic’s job is evaluation (how good) — and keeping them separate is what gives the actor a stable, independent assessment to learn from rather than a self-generated one. It illustrates that the split is about who optimizes what, orthogonal to whether the system iterates; the actor and critic are separated by role even though both update continuously.