nucleation

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

5 min read

Properties10
StatusActive
Version2
Standingstipulated
Kindexecuted-study
Tagsself-governance, complex-systems, probabilistic-modeling
Linedistributed-agency
IllustratesCLAIM-297 — When an operator's authority is updated by how well it performs on the interpretation it helped make dominant, the coupled meaning and influence of an organization can amplify an initial interpretive advantage without any gain in fidelity to independent evidence, CLAIM-277 — Durable authority among an organization's processes need not be a primitive: trust can be the temporal regularity of endorsement that policies earn by being right often, and the same loop entrenches whatever gained influence early
Variant ofCASE-155 — Authority as a replicator population: for a distribution π of control over processes with mean qualities q, the expected Maynard-Cross update is π_a q_a / v, Bayes' rule when q is each process's likelihood; the Maynard dynamic is the Taylor dynamic slowed by 1/v, so the two share their orbits; mean quality climbs at the rate Var_π(q); and when a process's score rises with its own share, ℓ′ = ℓ + h + c(2π − 1) has an unstable threshold π* = (1 − h/c)/2, above which a process the evidence disfavours captures the organization, CASE-156 — A legibility filter as the ν of the meaning–authority–legibility model: in a safe world where 15% of raw cues are alarms, an admission rule P(admit | y, r) = σ(c + χry) with c = 0 and χr = 2 admits 88.1% of alarms and 11.9% of reassurances, so 56.6% of the admitted record is alarms with the world and the sensors unchanged; at symmetry the mean admitted evidence is χσ′(c)r to first order, so ν = η_m χ σ′(c); and r_t = λᵗr₀ + ε Σ λ^(t−1−j) a_j is an exponentially weighted history of authority, CASE-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
Cited as “grounds” byARG-042 — The record's constructions of maintained correction show that self-maintaining organizations ought to preserve their corrigibility, CLAIM-466 — Lock-in from selectively admitted evidence and lock-in from self-scoring need different remedies: weighting admitted evidence by known, positive admission probabilities undoes the first and leaves the second, so correcting for how evidence was selected is not correcting how it is scored, 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
Cited as “variant of” byCASE-186 — A proposed study: an organization in a world whose hidden regime reverses and then leaves its hypothesis family, with evidence generation, admission and scoring separately lesionable, its diagnostic machinery maintained by its own reciprocal production, from outside, or not at all, and finite fuel shared among operation, maintenance and diagnosis, to find whether a second correction after injury to the diagnostic and copying machinery separates arrangements that correct the first failure alike

Graph View

The case

Two perspectives, A and B, share authority. B is the better one throughout, and a history starts with A in charge.

  • v1.2, a four-variable model (dmarx/PerspectiveLab@7e2edc2 experiments/regime_atlas/). Its state is the log odds z of A’s authority, a record lean r, and the competences k_A and k_B, which can grow with access. Self-scoring c feeds the record back into authority, and selectivity s hides evidence that contradicts the record. An 81 × 81 grid over c and s runs two histories from identical competences. In its limit (no leak, no learning, full observation) one Euler step is CASE-155’s map, checked to 12 places by the lab’s test.
  • v1.3, an event-level population (experiments/stochastic_population/). 32 networked perspectives with credibility weights, local records, admission of cues that depends on the local record, and practice. Paired histories share a world. The outcome is a wrong incumbent: A holds at least 75% of authority at the end of 1,400 updates. 64 world seeds per controlled condition. The inverse-propensity arm is given the exact admission probabilities, as an oracle.
  • v1.1, a factorial on CASE-158’s design (experiments/capture_feedback/, EXPERIMENT_REPORT.md). It crosses an independent evaluator with one scored 0.2·truth + 0.8·distance to the current mixture, and full audits with audits attributed to one specialist per receiver, at equal query cost. Receipts arrive after the episode, for the router as for governance. 12 paired seeds.

v1.2 is not a reduction of v1.3. They share motifs, and their mechanisms differ: v1.2’s gate only attenuates an expected negative gap, while v1.3’s sampled admission can reverse its sign.

Results, as stored

v1.2 (results/atlas_summary.json):

  • With competence frozen and full observation, the positive equilibrium appears at c = 0.42468, a fold. The declared history is trapped from c = 0.4469 at s = 0, and from c = 0.2438 at s = 1.
  • With competence that grows with access, the onsets fall to 0.3900 and 0.2356. B stays the better perspective at every competence the model allows, so what entrenches is authority, not merit.
  • The onset depends on the starting history. At s = 0 it is 0.6419 from z = 1, 0.4388 from z = 2.2, and 0.4306 from z = 3.5 and from z = 5. It approaches the fold from above and does not reach it.

v1.3 (results/summary.json, results/horizon_check.json), worlds with a wrong incumbent out of 64, with no self-scoring (c = 0):

selectivity s00.50.751
naive credit0314141
exact inverse-propensity weighting0005
  • At s = 0.75 the raw cue mean is −0.193 and the mean of the admitted cues is +0.184. Selection has reversed the sign of the available evidence, as in CASE-156.
  • The 41 of 64 hold at 700, 1,400 and 2,800 updates. The corrected 5 of 64 at s = 1 is 4 of 64 at 2,800. Admission probabilities near zero are where the correction’s positivity condition fails.
  • Self-scoring alone, c = 0.5 with s = 0, leaves 39 of 64 wrong. Nothing is censored there, so the correction changes nothing. At c = 0.25, s = 0.5 it takes 41 to 0, and at c = 0.5, s = 0.5 it takes 41 to 34.

v1.1 (EXPERIMENT_REPORT.md). The lab’s verify_v10.js reproduces all 48 MSE entries and 12 amendment counts of CASE-158’s table, so that table is now reproduced, not only reported. Mean post-reversal MSE:

evaluatorauditsroutergovernance
independentfull0.022370.02546
independentselective0.032840.03477
80% capturedfull0.044800.06965
80% capturedselective0.043820.06984

The router has the lower error in all four cells. Under 80% capture neither recovers on any seed. Under selective, independent audits governance recovers on 7 of 12 seeds and the router on 4, with the worse mean error.

What it can show

  • In two stipulated models, authority scored against what it made dominant can hold the worse perspective in place, which is CLAIM-297’s amplification without fidelity, and practice that depends on access makes it easier, which is CLAIM-277’s loop.
  • That the fold and the trapping of a given history are different quantities, and the second depends on the history.
  • That selection alone can lock in, that an evaluator who knows the admission probabilities can undo it while they stay positive, and that this correction does not undo self-scoring.
  • That in CASE-158’s design with the router facing the same audit delay and cost, the router still wins.

What it cannot show

  • CASE-092 analysed. Neither model has retrieval or a meaning state.
  • A phase transition or a critical curve: these are grids and finite horizons.
  • That the two realizations are one regime. Their mechanisms disagree, which counts against reading them as one (CLAIM-286).
  • A procedure for unknown or zero propensities. The correction is an oracle.
  • Any advantage of governance over routing, or anything about real organizations.

Provenance

dmarx/PerspectiveLab@7e2edc2: experiments/regime_atlas/, experiments/stochastic_population/ and experiments/capture_feedback/, with their READMEs, stored results, and the handoffs RESEARCH_HANDOFF_V12.md and RESEARCH_HANDOFF_V13.md. The figures above are read from the stored results. When filed, the regime-atlas and stochastic-population test suites were re-run at dmarx/PerspectiveLab@11c1495 and passed; the runs themselves were not regenerated.

Backlinks