The design (A47)
- Each perspective holds a low-rank denoising operator, trained beforehand on its own synthetic distribution and then frozen: “the LoRA parameters θᵢ are trained beforehand and frozen while governance adapts”.
- Receivers keep trust, audited estimates of competence and spatial grants. Authorized proposals are combined, and the scheduler steps once.
- “The current episode’s audit cannot determine that episode’s influence allocation. Its results become available for future decisions.” Probes are held-out images with known injected noise, and the audit has an explicit cost “so the governance mechanism doesn’t get free information” (A47’s progress note).
- 12 paired seeds of 84 episodes. The specialist each region requires changes halfway through.
- Four arms: fixed allocation; audited governance; conformity scoring, which “measures whether contributions resemble the prevailing collective proposal”; and a contextual mixture-of-experts router.
- A counterfactual that resets only the learned grants, keeping operators, trust estimates, environment and exploration fixed.
Reported, not reproduced
Post-reversal held-out mean squared error:
| Arm | MSE |
|---|---|
| contextual mixture-of-experts router | 0.02216 |
| audited governance | 0.02521 |
| fixed allocation | 0.06476 |
| conformity scoring | 0.07259 |
Audited governance beat fixed and conformity-scored allocation on 12 of 12
seeds, and the router beat audited governance on 12 of 12. A47: “We have
evidence that learned governance adapts, but not that it outperforms
conventional learned routing.” And: “A learned governance mechanism
should not be considered explanatory simply because it changes the
allocation of influence. An ordinary contextual router can accomplish that
too.” The software (v1.0 and sd15_routed_lora.py, the latter not run on a
real model) is not held.
What the record predicts
- Every proposal is probed, which is full-information feedback. Audited governance should then recover with inertia and no lock-in, as in CASE-157’s extension (finding 1).
- The conformity arm has an endogenous metric, so it should hold the old allocation: CASE-157’s finding 2, CASE-155’s self-scoring c, and TERM-066 built in by design.
- 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.
All three of A47’s outcomes are what the record predicts. CASE-157’s extension, committed at 03:54 UTC on 2026-10-25, 42 minutes after A47 was reported, found the first two in a different model, independently.
Prior art for the fixed arm. Frozen LoRA operators whose proposals are combined in one scheduler step is LoRA Composite (LIT-1529, NOTE-1160) with spatial grants, which is MultiDiffusion’s fusion with LoRA-modified models as the views (LIT-1522, THEORY-958). Neither paper learns, audits or routes the allocation, so the governed and router arms are what this case adds. That merging LoRAs degrades as their number grows (THEORY-955) motivates combining per step. It is no evidence about governance.
Two caveats
- The router may have had more information. Governance is barred from the current episode’s audit. A47 does not say whether the router also was, what context it saw, or whether it paid the audit cost. If it saw the region and an undelayed signal, “the router beats governance 12/12” partly measures information, not mechanism.
- Competence is frozen. So the design cannot show the loop A42 called crucial, access → experience → competence → future access, which “can also preserve an early advantage” (CLAIM-277; THEORY-927’s self-reinforcing gate). A47’s claim to have “closed the capability → evaluation → authority → deployment loop” overstates this, as its own §4 concedes.
What would make it informative
Arms in which governance could beat a router, one per failure of QUESTION-271:
- Cycles: specialists whose wins against one another are intransitive.
- No common scale: specialists scored on incommensurable criteria.
- Capture: probe sets that the highest-trust specialists can add to, filter or duplicate.
- Bandit feedback: probes only of the proposals that were heeded.
Plus one arm in which specialists keep training on the opportunities they are given, for the entrenchment loop. The router faces the same audit delay and cost in every arm. Outcomes are scored against held-out truth, never against an arm’s own metric. If governance beats the router only in arms 1–4, CLAIM-278 is answered in those regimes. If it never does, the objection stands for this model class. A47 pointed the same way: “we need tasks where the evaluation process itself becomes contested, incomplete, selectively observed, or historically entrenched.”
What it cannot show
Anything about real organizations or real diffusion models: the operators are small synthetic denoisers, and no pretrained model was run.
Note of 2026-10-26: the third prediction in another model (reported)
This case’s third prediction: “A router trained on the same audits … Under full information, governance should not beat it.” T20 of the exchange “Experimental Infrastructure Review” reports a working group with delegated authority and a conventional router given “the same information and authority” at the same regret, 0.00499 (CASE-166). It is a tie, not a loss as here, and T20 does not say whether the router shares the working group’s code. Every arm is full-information, so none of this case’s arms 1–4 is present, and the result is what the prediction expects. Reported, not reproduced: T20’s “PerspectiveLab v2.11” is not held, and dmarx/PerspectiveLab holds no v2.10 or v2.11 (its latest import is c54617e, v2.9). The case is unchanged.