18.16 What the experiments establish

. The evidence has three different statuses. TI/NI supplies the preregistered declaration-and-repair ladder; the sports images are qualitative external challenge cases with disclosed prompts and manual audits; and Bongard Corpus–4 is an oracle-assisted upper bound. None of these, alone or together, shows that ARTISTIC has autonomously invented and retained a new visual theory.

The evidence sequence supports four bounded claims. A diffusion-based generation-and-repair workflow can maintain and evaluate a typed visual declaration rather than relying only on pixels or a scalar similarity; this does not imply that the declaration was enforced inside every denoising step. Structural failure can be localized to typed counter-witnesses and repaired with less area than blanket projection. The declaration itself can be estimated and evaluated against information withheld from selection and repair. Finally, uncertainty can trigger another probe that increases safe relational coverage without converting disagreement into certainty.

In DIAL language, ordinary denoising is the assimilative direction: it moves within a fixed visual and conditioning theory. The proto-accommodative direction probes changes to the declaration, observer factorization, repair operator, or acquisition policy; admitted finite accommodation installs one of those changes. The TI-4 and NI-2 evidence sequences show why these directions cannot be collapsed. Source and change observers have different jobs; point and interval relations support different inferences; a failed declaration can demand a new probe rather than another optimization step.

The experiments do not yet construct a double involution algebroid or measure its abstract interchange defect. They realize a finite operational shadow of the proposed geometry: two classes of change, a mixed failure, a typed extension, and an independent admission record. Nor have they shown visual-language invention in the strongest sense. The vocabulary of people, masks, parts, colors, and visible-area relations is supplied by the experimenter. The system constructs a control structure inside that language: adaptive admission, a factored observer, an uncertainty type, or an acquisition rule.

ARTISTIC’s admitted empirical substrate defines a stronger test: withhold an object, relation, or compositional constructor and ask the system to propose the declaration type itself. Success would require communication to a fresh interpreter, transport across renderers or models, and counterfactual uses not enumerated in training examples.