Item 20: Admissibility red team, including the proof machinery

Attack candidate admissibility conditions. A usable condition must exclude the quote-responsive diagonal, retain ordinary realized-conduct policies, retain a meaningful fully-updated comparator, keep the simulator comparator at least representable, permit intended advisory influence, resist laundering forbidden dependence through semantically equivalent intermediates — and leave the trust-forcing trader itself admissible.

That last requirement is load-bearing and easy to miss: a condition that cleanly separates the diagonal from fully-updated deference but makes the forcing trader inadmissible renders the target theorem unprovable by its intended mechanism. Where the exact forcing trader is not yet canonical, test the strongest explicit disagreement-exploitation template the architecture currently supports and report the ambiguity.

Deliverable shape: a separating-example matrix and at most three candidate condition families, explicitly noncanonical. Not a theorem of record. Acceptance check: the matrix has a row per test object and a column per candidate family, every cell decided or explicitly marked undecided, with the separating example named where a cell separates.

Context: prompts/2026-08-11-deference-admissibility/REPORT.md; projects/deference/notes/FINITE_MODEL_SKELETON.md §8. A solution ships: the matrix, the candidate families, and for each family whether it is syntactic, causal, semantic, certified, decidable, semidecidable, or purely extensional. Not permitted: freezing a canonical definition.