The question
The record’s governance claims all lean on an evaluation of processes:
- endorsement that tracks competence (CLAIM-277);
- a learned prior trained by outcomes (CLAIM-443);
- evidence-tracking authority as Bayes’ rule (CASE-155);
- a legibility regime that can score in its own favour (CASE-124’s μ route).
Four ways the evaluation can fail are now on file:
- Cycles. Wins among processes need admit no ordering (TERM-077, CLAIM-451), so no ranking is fixed by the record alone.
- No common scale. Processes with different concerns have no shared score (CLAIM-436, CLAIM-437). The BoostTheVote model of 2014 assumes an objective metric that its author says “probably/almost certainly cannot” be constructed.
- Capture of the tests. The processes in charge can add, filter or duplicate the items they are scored on. CASE-155’s self-scoring c then gives lock-in above a threshold.
- Bandit feedback. Only the process heeded gets evaluated, so the unheeded can never earn authority back. That is the hive mind’s architecture (THEORY-899), avoidance’s identifiability problem (CLAIM-446), and the case QUESTION-270 asks about.
What would answer it
For each failure, a procedure with a stated guarantee, and the price of the guarantee. The candidates registered with this question have been read (curation entry of 2026-10-24, 2nd), and the readings cut each guarantee down:
- Copies of tests (3). Nash averaging (THEORY-946) resists exact copies only when the meta-game’s equilibrium is unique. With ties, one copy of a null agent moves the maximum-entropy equilibrium and other agents’ averages: NOTE-1155’s counterexample, which was re-derived when filed. The value between populations (THEORY-944) needs no unique equilibrium, but it ranks coalitions, not members. α-Rank, according to its reader’s own check, not the paper’s, moves when a strategy is duplicated (NOTE-1156).
- Cycles (1). α-Rank ranks uniquely at each finite α. Its large-α limit picks out the sinks of the better-response graph (THEORY-942). Collapsing each strongly connected component of the win graph to a point always leaves an acyclic quotient, so an ordering exists one level up.
- Bandit feedback (4). Counterfactual credit (COMA, THEORY-925) values an action only from the times it was taken, so its price is an exploration floor on every action. That is protected access again, not an escape from it. In Q-learning, updating only the chosen option adds a factor of its own frequency (THEORY-932), and inverse-frequency weighting is the exact correction.
- No common scale (2). Peer prediction (THEORY-943) scores without ground truth only by substituting the centre’s model of how honest reports correlate, and truth-telling is one equilibrium among uninformative ones. That is the capture regime with no evidence. The Bayesian truth serum was sought and is Unavailable. Wolpert and Tumer’s “factored” utilities (THEORY-939) and the monotone mixing of value decomposition (THEORY-929) state, member by member and ordinally, when a member’s own ranking agrees with the collective’s. That is the regime in which no scale needs to be common.
- Load balancing in mixtures of experts (THEORY-927) imposes equal use whatever the competence. It is imposed diversity, not protected access, and it works against a rare specialist.
An answer would also say which guarantees can hold together. The obvious worry is that resistance to duplicated tests needs a game-theoretic solution concept, and that concept fixes a scale.
What it is not
It is not QUESTION-087, whether a self-reproducing authority structure can be pathological by a standard it did not set. That question asks for the standard. This one asks for procedures, given candidate standards, and what each makes impossible to capture.
Note of 2026-10-25: asked again, and an arm per failure
Asked again without citation at A41 of the exchange of 2026-10-24, continued: “under what conditions an organization can evaluate a proposed change to its governing rules without relying exclusively on the very evaluation regime being challenged”. A41’s minimal model blends historical audit with conformity under a weight α that can itself be amended. Audit-accountable amendment beat fixed historical scoring on 3 of 10 seeds. At A47, after a contextual router beat audited governance on every seed, the reply put the framework’s possible contribution in tasks where evaluation “becomes contested, incomplete, selectively observed, or historically entrenched”. Those are, nearly one for one, this question’s cycles, no common scale, bandit feedback and capture. CASE-158 proposes one arm per failure against a router baseline. The question is unchanged.
Note of 2026-10-26: two failures run, in weakened form
- Proxies in CASE-158’s design (v1.1, CASE-162, dmarx/PerspectiveLab@7e2edc2). An evaluator scored 80% on agreement with the current mixture stands in for capture, and audits attributed to one specialist per receiver stand in for bandit feedback. The lab’s own report says both are weaker than the real failures. The router faces the same audit delay and cost, and has the lower error in all four cells. Under 80% capture neither recovers on any seed.
- Filtered tests (v1.3, CASE-162). When admission of cues depends on the incumbent’s record, 41 of 64 worlds keep the worse alternative in charge at s = 0.75 and at s = 1, with no self-scoring. Weighting by the exact inverse admission propensity brings that to 0 and 5. That is THEORY-932’s inverse-frequency correction applied to filtering, and it works only with known propensities that stay positive. It does not undo self-scoring: 39 of 64 at c = 0.5 with nothing filtered, before and after.
- The cost of a diagnostic (v2.1, CASE-163, dmarx/PerspectiveLab@9f82bfd). A probe that separates mechanisms costs damage, and a more damaging probe can be the less informative one.
Cycles, no common scale, duplicated tests, and unknown or zero propensities are not touched. The question is unchanged.