2026-10-08 / Added after the held-out scores were inspected, before processing interpreter replies.
This is an exploratory engineering diagnostic, not part of the frozen predictive comparison and not a test of understanding.
For each accepted reply, feed its non-null colours in cell-array order into the selected model and generate24 colours with the already fixed return seed. Also generate a counterfactual continuation using the originating Sx-T2 field's non-null colours with the identical return seed. Retain both, the full first-step distributions, Jensen-Shannon divergence and the number of different sampled positions. Do not reject a reply if the outputs happen to match. Both conditions receive only colours; positions, nulls, typed links, regions and explanatory prose are not model inputs. Context sensitivity establishes an effect of changed numeric inputs, not recognition of a viewer's intended challenge. The public gallery labels the counterfactual condition clearly.
