nucleation

CLAIM-467: Under full-information, independent audits, a learned allocation of influence among specialists is no better than a contextual router trained on the same audits, which is a mixture of experts whose experts earn responsibility by fitting better; whatever governance adds over routing has to be shown where evaluation fails, and the record has not yet found such a regime

2 min read

Properties11
StatusActive
Version3
Rolethesis
Warrantinterpretive
Tagsself-governance, learning-theory, complex-systems
Linedistributed-agency
Bears onQUESTION-271 — How can an organization evaluate its own processes so that the evaluation decides authority without being captured by it: when wins among the processes cycle, when no common scale scores them all, when the processes in charge can add, filter or duplicate the tests, and when only the processes heeded are evaluated?
GroundsCASE-158 — Audited governance against a contextual router after a reversal of expertise: specialist denoisers with frozen, separately trained operators propose for one shared latent at each step, and spatial grants are revised from independent held-out probes of every proposal, available only from the next episode, against a fixed allocation, a conformity-scored allocation and a contextual mixture-of-experts router; the question is whether governance does anything a router does not, and under full-information audits the record predicts it cannot, CASE-162 — Capture and selectively admitted evidence in two lab realizations: in a four-variable model of authority, record and competence the equilibrium fold under self-scoring is at c ≈ 0.4247 while the declared history is trapped from c ≈ 0.4469, and in a 32-member event-level population, admission of cues that depends on the record reverses the sign of the admitted evidence and leaves the worse alternative in charge in 41 of 64 worlds, which exact inverse-propensity weighting brings to 0 of 64 below full selectivity
UndercutsARG-036 — The lab's restricted models, whose predictions could have failed, support the shared-organization thesis
Corroborated byTHEORY-937 — When a mixture of experts is trained on the error of a single expert chosen stochastically by the gate, or on the negative log-likelihood of the mixture, the gradient raises the gate's logit for an expert that fits a case better than the gate-weighted average and lowers it for one that fits worse; on a four-class vowel task, mixtures of 4 or 8 linear experts learned in about half the epochs of backpropagation networks of roughly equal parameter count, at equal accuracy, and divided the task by class pair, leaving all but 2 or 3 experts with mixing proportions effectively zero
Rendered intodocs/arguments-by-conclusion.md

Graph View

The claim

CASE-158, “What the record predicts”: “A router trained on the same audits is a mixture of experts whose experts earn responsibility by fitting better (THEORY-937). Under full information, governance should not beat it.” The reason is that with every proposal audited and the audits independent, governed reallocation and a learned router are both fitting the allocation to the same signal, and the router does it directly.

The runs agree. A47’s reported results, reproduced in the lab (CASE-162, v1.1): mean post-reversal error 0.02237 for the router against 0.02546 for governance with independent, full audits. In the same table the router also had the lower mean error under selective audits (0.03284 against 0.03477) and under 80% capture (0.0448 and 0.0438 against 0.0697 and 0.0698), though under selective, independent audits governance recovered on 7 of 12 seeds and the router on 4.

So the first half, the bound under full information, is predicted and observed. The second half is what CASE-158 says would make the design informative: arms in which evaluation fails as QUESTION-271 lists, with cycles, no common scale, capture and bandit feedback. Two of those (selective and captured audits) have been run in one form, and governance did not win on mean error. CLAIM-278’s objection, that the framework adds vocabulary and not prediction, stands for this model class until one does.

What it does not say

  • That governance never helps. Cycles and incommensurable criteria are untried, and so is the entrenchment loop with competence that grows.
  • That the router saw only what governance saw. CASE-158’s caveat: A47 did not say whether its router faced the same audit delay and cost; CASE-162 reports a check with the delay matched, in which the router still won, but the case calls the design’s information-matching a condition of every future arm.
  • Anything about organizations. Specialist denoisers in a lab.

Backlinks