CASE-144: The evaluator that stops tracking: two populations of specialists with identical capabilities and task streams, one scored against independent outcomes and one by an evaluator that weights evidence agreeing with its previously favoured specialists, after a shift that makes a marginalized specialist uniquely competent

The design

Two populations of adaptive specialists with the same capabilities face the same task stream, which has independently defined outcomes. In one, the evaluator scores specialists against those outcomes. In the other, it gives more weight to evidence consistent with the specialists it already favours: the coupled scoring of CASE-092’s Regime A, where CASE-099 and CLAIM-330 found the symmetric state can be unstable. At first both may perform alike. Then the task distribution shifts so that a specialist the organization has marginalized becomes the only competent one.

A5’s four conditions for the marginalized specialist:

1. It keeps its competence but has little authority.
2. It keeps competence and has a protected channel to challenge the evaluation.
3. It has lost competence for want of learning opportunities.
4. The organization can recognize and revise its evaluator’s standard.

A4’s “corrigibility trap” is the long version: an evaluation rule that works well for a long period and then becomes systematically misleading, where accepting the corrective evidence “would undermine the authority structure responsible for deciding what counts as credible”. Randomize whether corrective evidence is independently scored, whether minorities have protected routes to challenge, and whether the evaluator’s own rules can be revised.

What is measured

Held-out performance across the shift; concentration of interpretation and of authority; survival and recovery of the minority specialist; hysteresis after a temporary perturbation; whether the evaluator’s failure is recognized; and, against CLAIM-427’s third grade, the interventional prediction error of competing models (mixture-of-experts routing, reflected-appraisal influence, a conventional adaptive policy).

What the outcomes would show, and what they would not

• Four causes of one failure. A5: the design separates “The right information wasn’t available”, “The right information was available but ignored”, “The specialist once had the right information but lost its competence”, and “The organization recognized the problem but could not revise the governing mechanism”. Those are CLAIM-421’s activation, enactment and reproduction, and CLAIM-424’s reflexive governance, in one experiment.
• A clean negative case. A4: “an organization that is excellent at optimizing within its current perspective yet persistently fails to reconsider the perspective itself.”
• Not novelty by itself. A4: there is “extensive antecedent work in organizational learning and robust distributed systems”; hidden-profile research (QUESTION-089) and Byzantine-robust learning already show the problem. A result counts for the framework only at CLAIM-427’s third grade.

Note of 2026-10-24: March as the baseline

From the exchange of 2026-10-24 (workbench entry of 2026-10-24, U21–U22). Rather than a fixed-weight learner, the comparator for this study is March’s model of mutual learning between an organizational code and its members (THEORY-142). In March, rapid socialization can remove the independent views from which the code might learn, and an environmental change then exposes it. The extension that would test CASE-124’s legibility state keeps the members explicit and lets their scores depend on the prevailing regime, s_i = S_r(h_i, c, observations), with standing feeding back into r. The informative outcome is the one March’s baseline lacks: independent views stay present and accurate, and the organization still fails to credit them. Controls: the same evaluation bias without a separate r memory, and history-dependent retrieval with no explicit legibility coordinate, so that the extra state is not rewarded merely for being there.

Note of 2026-10-25: run without citation, with a router added

A47 of the exchange of 2026-10-24, continued, reported running this design without citing it. The specialists were frozen low-rank denoisers, the required specialist reversed at mid-run, and scoring was either against independent held-out probes or by agreement with the collective proposal. A contextual mixture-of-experts router was added. Reported over 12 seeds: audited governance beat fixed and conformity-scored allocation every time, and the router beat audited governance every time. Not held and not reproduced. The variant with the router arm, and the arms that would make it informative, is CASE-158. The case is unchanged.