The case
A27 §10, Experiment C: “Build two computational systems with equivalent distributed control and similar behavior. In one, a higher-level self-model represents the actual division of labor. In the other, it represents the system as governed by one centralized inner agent.”
Held fixed: the distributed control and the behaviour it produces. Varied: only the content of the self-model, accurate or centralized.
Measured, in A27’s words: the systems’ “judgments of authorship, explanations of mistakes, capacity to identify which component caused an outcome, and ability to revise governance after an error.”
A29’s Test C is the same design, applied to distress: “Give otherwise similar systems accurate or overcentralized self-models. Test whether a unified report of distress predicts real regulatory integration, merely coordinated communication, or neither.” Its second appearance is why it has a code.
What each outcome shows
- The centralized model misattributes control, and revises governance worse after errors. A27: “we would have a concrete model of illusory authorship or subject-attribution”, “a specific representational error and the conditions that produce it”. The error belongs to the system that holds the self-model, as CLAIM-355 says it must.
- No difference on any measure. The owner content is idle for these functions. The self-model narrates and does not govern, against A26’s “self-representations may themselves be parts of the real causal organization they represent” and A27 §12’s loop from self-attribution back to governance.
- The centralized model does as well or better. One owner is a useful compression and misdescribes nothing that these measures reach. That is A27’s “Neither representation must be false”, and Drayson’s reading (THEORY-646): the represented owner has a truthmaker. It counts against CLAIM-355 for this system.
The main threat to inference
The two self-models differ in more than whether they present an owner. An accurate model of the division of labour carries more information about the system than a centralized one. Worse attribution by the centralized system may then show only that a model with less information answers questions about components less well, which is true of any compression. The comparison needs the two self-models matched in capacity, with the centralized one wrong about control rather than merely coarser. A second threat follows from A27 §12: if the self-model feeds back into governance, the two systems’ control does not stay equivalent, and what was held fixed drifts.
What it cannot show
- That a distributed agent would come to represent a single owner. The modeller writes both self-models, so the content is the modeller’s (CLAIM-288). The case tests what such a content does once present.
- Anything about phenomenal properties. A27: “We would not thereby have demonstrated illusionism about phenomenal properties.”
- What access each component has to the integrated state. That is A27’s Experiment B, binding by restricted access, noted on QUESTION-061.
Note of 2026-10-21: the converse arm, and a third factor
The exchange of 2026-10-21 on experience and interpretation came to this case twice. A7 §17, Study B, kept the design and added its converse: “Conversely, alter actual integration while keeping self-descriptions similar.” Together they make a two-by-two of integration (strong, weak) against self-attribution (accurate, misleading). The new cell is weak integration presented as unified, “Fragmented organization presented as unified”. That is the case CLAIM-355’s A27 §7 had in mind, where “apparent unity in the organization’s reports may exceed the degree of information sharing actually present”. The converse arm has the same threat as the original, in mirror image. If the self-model feeds back into governance, a fixed self-description does not leave integration where it was set.
A5 proposed grounding, the degree of independent constraint on the signals and their scoring (CLAIM-413), as a third factor. Its purpose is to find out whether “grounding affect[s] the accuracy of interpretations without changing the coherence of the perspective”. A6 suggested running the design on a language model, intervening “on internal self-attributions and workspace representations independently”.
Both replies kept the case’s limit. A7: “It would not demonstrate which conditions are conscious.” A5: complete answers on integration, self-attribution and grounding “would not, by themselves, establish phenomenal consciousness unless we accept an additional constitutive premise”. The design itself is unchanged.
Note of 2026-10-23: reported authorship already comes apart from causes
A5 of the exchange of 2026-10-23, “Framework Overview Request” cited Nisbett and Wilson (1977, LIT-1410) and choice blindness (Johansson et al. 2005, LIT-1417), registered unread. A5: they “already demonstrate that reported authorship and actual causal history can diverge.” What they do not show is this design’s thesis, that representing a centralized controller rather than an accurately distributed one changes error correction or self-governance.
Note of 2026-10-29: choice blindness, read
Johansson, Hall, Sikström and Olsson, LIT-1417, which the note of 2026-10-23 calls registered unread, is now read (NOTE-1398). A5’s “reported authorship and actual causal history can diverge” is right for it, and narrowly. What it shows is that a reason can be produced for an outcome one did not choose and be indistinguishable on eight measures from a reason for one’s real choice. It does not show that ordinary reasons are confabulated: the authors raise that “radical reading” and leave it open, and the reports that did name the original choice show that some reports carry the real one. Nisbett and Wilson, LIT-1410, is Unfetched. The case’s own limit stands: nothing here bears on whether representing a central controller changes error correction. The design is unchanged.
Note of 2026-10-29: Nisbett and Wilson, read
Nisbett and Wilson, LIT-1410, which the note above calls Unfetched, is now read (NOTE-1412). It supports the case’s premise, and narrowly. What diverges in it is a person’s report of why they responded and what in fact caused the response: subjects miss effective stimuli and credit ineffective ones, and their reports match what observers given only a description predict (in Nisbett and Bellows, subject and observer reports correlated .77 to .99). The authors read this as reports drawn from a priori causal theories, not from access to the process. So a self-report of why one acted can be a construction from a theory. The paper is about the influence of stimuli on responses, not about attributions of agency or control, and never discusses who did something; it is silent on the design’s thesis. The NOTE draws one prediction from its account: a system whose self-model is centralized should explain its errors as an observer holding the same model would, whatever its actual division of labour. The design is unchanged.