The claim
A4’s table, which the claim adopts:
| Finding | What it establishes |
|---|---|
| ”Different organizational mechanisms can produce identical ordinary behavior” | Non-identifiability from passive observation |
| ”Those mechanisms predict different responses to independently specified interventions” | That the organizational distinctions matter |
| ”The proposed framework predicts these differences better than established alternatives” | A distinctive theoretical contribution |
- Where the record stands. CLAIM-414 (Active) meets the first grade for two constructions, and its intervention meets the second within them. No entry meets the third. A3: “the current work has earned the status of a potentially useful conceptual synthesis, not yet an independent explanatory framework.”
- Why only the third. CLAIM-417’s defeater is that established accounts “explain the same transitions as well with fewer assumptions”. Showing that organizations differ under intervention does not touch it, since the established accounts may predict the same differences. A4: “beating a deliberately weak baseline is not evidence for a unifying theory.” The baselines that count are the mature theories already covering each phenomenon.
- The measure. A4 proposed one: the drop in predictive risk over a prespecified distribution of interventions when the coupling is represented, at comparable complexity and information budgets. The claim takes the form, not the formula.
- No success in the framework’s own terms. A4: “Suppose a system is described as possessing a perspective because its processes are mutually influential, and the experiment then demonstrates that its processes are mutually influential. Nothing substantive has been learned.” It is CLAIM-342’s rule for a self-confirming loop applied to the record’s own programme: the test must be scored by evidence whose scoring does not pass through the framework.
What it does not say
- Not that the framework fails. It says what would show it succeeds. A3’s assessment, “Achieved”, “Plausible”, “Not yet shown”, is the exchange’s, and the claim does not adopt its verdicts.
- Not that the conceptual work is worthless. A3 counted “preventing invalid inferences” as real, if “primarily negative or diagnostic”. The claim is about the further, positive contribution.
- Not that simulation can settle it. A5: “a simulation cannot independently validate the ontology it presupposes.” A simulation can meet the third grade only for the class of systems it models.
Note of 2026-10-26: where the lab stands on the grades
The Perspective Population Lab’s later versions meet the first two grades within synthetic constructions and do not reach the third. Matched ordinary behaviour with different responses to specified interventions is CASE-161 in a 32-member population and CASE-164 exactly. CASE-163 predicts withheld interventions from probes, but its forecaster is given the four candidate mechanisms and their equations, and every probe runs on a counterfactual reset.
A comparator pilot (dmarx/PerspectiveLab@dce302b experiments/comparator_pilot/) set rivals beside it under unequal information. A generic state-space model with the same transition kernels ties exactly, by construction. Boosted trees fitted only to the probe trajectories fail (MSE about 0.2 against 0.002). A supervised model trained on the candidates’ target outcomes beats the structural forecaster at nominal parameters (0.0013 against 0.0021). Its report says it is “NOT CLAIM-427 grade-three evidence”. The tie by construction puts this claim’s warning more sharply: renaming governing constraints as latent state changes no prediction. So the third grade has to be won by restriction, compression or transport, with the prior knowledge of mechanisms priced, and not by fit. The claim is unchanged.
The lab’s v2.3 (dmarx/PerspectiveLab@e9efcf4 experiments/matched_comparator/) removes the pilot’s unequal information. Every model is fitted on the same 208 source episodes of 96 uninterrupted updates, with the same intervention observations, and the primary comparison holds the fitted coefficients at 8 each. Tested on held-out production laws and a dependency graph absent from fitting, the structural model wins in the familiar sigmoid regime and loses on the bottleneck law with the mixed graph to a generic 20-coefficient polynomial (MSE 0.004514 against 0.001865; paired seed-cluster difference +0.0026, 95% interval [+0.0019, +0.0034]). A state-space model with the same feature basis ties exactly. The structural model’s hand-chosen basis is not priced in description length, and every law was written inside one experiment, so this is a better test of the third grade that still does not meet it. Its report says so.
The lab’s v2.5 (dmarx/PerspectiveLab@f392084 experiments/adequacy_guard/) adds a prequential check of model adequacy that switches a misled adaptive design to a fixed probe schedule. On an out-of-catalog bottleneck law it fires in all 16 test episodes and cuts the cost-adjusted loss from 0.146 to 0.022, though the fixed schedule alone does better (0.017). On an out-of-catalog cubic law it makes the loss worse (0.018 to 0.022), and on valid mechanisms it false-alarms above its nominal rate. A conventional Bayesian controller given the same kernels, likelihood, detector and costs takes the same actions and makes the same predictions in every episode. So what helps is ordinary robust model checking, not anything specific to the organizational description, and the third grade is still unmet.
The lab’s v2.6 (dmarx/PerspectiveLab@74621fb experiments/affective_dissociation/) carries the grades into affect, after CLAIM-437 and QUESTION-270. Five stipulated causes of pursuit after an outcome is devalued share one pursuit probability, 0.668188, and are separated by interventions written into the generator. That meets the first two grades within a synthetic construction, the first only for the one-step choice: over a longitudinal run two of the classes drift apart from the other three. Four interventions are withheld as targets, three of them combinations of the diagnostic arms and one a halved authority. Predicted from the catalog’s equations with the diagnostic probes and readouts, they have a mean MSE across the catalog of 0.00230, against 0.03496 from passive choice alone. A ridge regression trained on the source worlds’ target outcomes reaches 0.00646. It is given different information, and the structural predictor beats it with a bootstrap interval excluding zero in three of the five classes. The study’s same-kernel conventional row ties exactly, but it is computed by the same function as the structural row. The lab’s tests check the equivalence separately: a feature-table model of the same law gives the same probabilities, to 14 places, in every class and arm. A hybrid of cue drive and denied uptake, left out of the catalog, is assigned to a catalog class in all 40 test worlds, 27 of them with a top posterior above 0.9, and the adequacy check alarms in 9. The gain is a reward for measuring the relevant variables inside a supplied catalog. It is not a comparison at matched complexity and information, and the third grade is still unmet.
Note of 2026-10-26 (later): v2.7 to v2.9 on the grades
The lab’s v2.7 to v2.9 are at dmarx/PerspectiveLab@c54617e.
v2.7 (experiments/trained_affect/) comes nearest to the third grade’s information condition. Seven predictors of pursuit are fitted to the same 8,640 recorded choices from four synthetic worlds and scored on eight, four of them held out. Their coefficient counts run from 2 to 15, so complexity is not matched. The predictor with an uptake term has 5 coefficients. It loses in the governance world, whose generator gives that term its largest weight, to a 15-coefficient model with cross terms (MSE 0.01586 against 0.01205; paired seed-cluster difference +0.0038, 95% interval [+0.0034, +0.0041]). It also loses there to a 7-coefficient model with main effects and no interaction (0.01378). It beats the 15-coefficient model in one world, a held-out hybrid of cue drive and governance (0.00423 against 0.00511), and loses to it in the other seven, three of them held out. On the source data a BIC-like score already favours the main-effects model (7,414 bits against 7,441), before the report’s 12-bit charge for the hand-chosen term. Its same-feature conventional row is a copy of the fitted uptake predictor, so that tie is an identity. The record’s distinction is not needed to predict the model’s own synthetic interventions, and its one advantage is an inductive bias that suits one shift.
v2.8 (experiments/perspective_engine/) adds interpretive frames and runs no comparator. Its README calls it a coverage test, not a superiority test.
v2.9 (experiments/endogenous_frames/) selects a one-feature frame from a grammar of 15 rules by search on training labels, and a critic admits it on validation labels. Its conventional row runs the same selection code, so that tie is an identity too. The lab’s tests check separately that the frame, operator and governor predict exactly as the selected feature does. A 15-feature ridge regression on the same labels is not a stronger rival. It is worse than the selected frame in the five worlds whose law does not change (in the relational one 0.13442 against 0.12511). It is better only in the two worlds whose law changes after selection, and only before new feedback, where both are worse than a constant forecast of one half (0.59864 and 0.61897 against 0.25 in the reversal world). Neither version is a comparison at matched complexity, and the third grade is still unmet.
Note of 2026-10-26 (later): Northstar v2.11, reported and not reproduced
T20 of the exchange “Experimental Infrastructure Review” reports a multilevel firm, “PerspectiveLab v2.11 — Northstar” (CASE-166). Its working group with delegated authority (regret 0.00499) ties “a conventional router, same information and authority” (0.00499) exactly. T20: “That’s an important limit on what we can claim.”
As reported, the study meets neither of the first two grades. It shows no matched ordinary behaviour with different responses to interventions, except in the handoff arm, where knowledge differs (state-estimation MSE 0.00411 with records against 0.03398 without) and action “often” does not. It has no comparator at matched information that the structure could beat, and nearly all of the working group’s advantage over silos (95.9%) is fast authorization. Whether the router tie is an identity, as v2.7’s and v2.9’s were (the note above), is not said. dmarx/PerspectiveLab holds no v2.10 or v2.11 (its latest import is c54617e, v2.9), so none of it is checked. The third grade is still unmet. The claim is unchanged.
Note of 2026-10-28: moved to ARG-025
This claim is about evidential bearing: which findings would support CLAIM-417 against its defeater. Under ADR-061 that is an argument, and it is now ARG-025, from CLAIM-414, CASE-161, CASE-163, CASE-164 and CASE-166 to CLAIM-417, form evidence, Proposed. The three grades answer the form’s critical questions there, and the argument’s body carries this claim’s work and its defeat condition, which the argument scheme has no field for.