Item 22: The weakest protected-authority interface

Determine the weakest abstract interface on which prediction of authorization does not constitute authorization is a theorem rather than a stipulation. The principal may be perfectly predictable; the protected role is causal and capability-based, not epistemic.

The candidate starting point is already on the table: type a conduct's selection on an actual-report coordinate rather than on states alone, so that delegation and simulation are distinct functions even where their realized selections agree. The item is to find out whether that suffices.

Four questions the deliverable must answer, the last being decisive: is the report coordinate enough, or is a capability restriction also required; does the interface survive token responsiveness, which defeats the natural counterfactual criterion; what is the exact hypothesis stating the guarantee lapses if the channel can be forged, bypassed, rewritten or seized; and does the interface separate the two conducts without claiming the separation is inferable from a run.

Deliverable shape: a report plus a proposed skeleton clause, not a theorem of record and not a canonical definition — naming is reserved. Acceptance check: the report answers all four questions explicitly, and any proposed clause is stated as a versioned patch with the tracks it would require rerunning.

Context: prompts/2026-08-11-deference-channel/REPORT.md §9.2 and §1.3. Not permitted: freezing an authorization-token or cryptographic story as the formalization; claiming the causal fact is behaviourally verifiable in general.