proper exercise is semantic input, not a theorem module

agent-decided, reversible. A round went looking for the smallest mathematical structure between this authority is grounded and this act was within it, and found that there is none to have.

The evidence is a pair of constitutions with the same base and the same act — identical grounds, disposals, issued content and declared evidence — both satisfying every premise and every corollary of Grounded Replay. One refuses the act. The other accepts it and the state gains a capability nobody licensed. The difference is whether the permission relation reads what the act puts in force. So a no-escalation theorem would have to quantify over permission relations, and none is true.

Two results survive and both are inherited from the kernel's typing rather than from any structure on permission: no edit's permission can rest on a capability it creates, which follows from strict pre-state evaluation and freshness; and reach is non-increasing when no accepted edit widens beyond its basis, which is the naive delegation rule available as a conditional and false as an axiom.

The practical consequence is that permission stays an opaque parameter and the round does not wrap it in theorem language. Fourteen separations — jurisdiction, delegation four ways, self-amendment, total constitutional replacement, threshold authority, a veto, ex-post rationalisation — are all decided by it and none of them needed it to have internal structure.

Rejected alternative: a capability calculus or a proof system for exercise, either of which would have added an ontology without adding a theorem.

Record: projects/normativity/legitimacy/rounds/2026-08-25-legitimate-evolution/PROPER_EXERCISE.md.