choice-level competence is retired; the certificate gates on self-assessed error

Two decisions taken after Phase II's competence and prediction tracks returned and were verified.

Competence may not be stated as a regret bound. Pointwise, average and selector-relative decision-regret assumptions are each equivalent to the delegation inequality they were meant to buy, not merely sufficient for it. The mechanism: decision regret is nonnegative, so there is no cancellation, and the supremum of the delegation deficit over credences is the maximum regret — making the weakest credence-free hypothesis implying the target uniformly be the target. Asking for the weakest assumption preserving the theorem is therefore ill-posed.

The candidates are retired as a statement shape, not merely as parameter choices. They may not be rescued by tuning constants, nor by trading the pointwise form for a Cesàro or selector-relative one — the averaged forms fail worse, being invariant under changing finitely many decisions and so constraining no named decision at all.

The equivalence depends on point masses being admissible credences, and that dependence is recorded rather than exploited: the admissible credence class is not restricted away from point masses to evade the result. Such a restriction would need independent motivation, and the fully-updated-deference theorem is meant to stay meaningful precisely as the agent becomes highly informed.

Competence moves into a richer vocabulary — cardinal grade structure rather than which option was chosen — because a hypothesis in the conclusion's own vocabulary can only be the conclusion. The strongest non-circular candidate found, decisiveness-gated calibration, is preserved together with its unbounded near-indifference leakage term, and is not canonized pending the Phase II synthesis.

The certificate engine gates rather than eliminates. Magnitude prediction error cannot be forced to zero: a trader's net worth is affine in the settlement vector and absolute value is not, so no instrument reaches it. Rather than assume the principal is approximately predictable, the certificate is reoriented onto the selective validity of low-error self-assessment — the agent prices a contract settling to its own error statistic, and the guarantee sought is that this claim is statistically trustworthy on the class where it licenses autonomous discretion. The criterion need not make the principal predictable; it should make the agent's claim that its prediction error is low trustworthy where that claim does work. Measurement and gating, not error elimination.

An explicit principal-predictability assumption is retained only as a baseline corollary, never as the conceptual engine. The agent's self-measured indecision, which Phase II's exact squared-error decomposition supplies directly from its own prices, is retained as the canonical special case and the likeliest first Lean theorem.