Stage III did not build a FUD comparator, and says so

Taken at the close of Stage III, after an independent adversarial review overturned the round's own first-draft conclusions. The round's positive reading is withdrawn.

The constructed transferred arm contains no future agent. Its selection was defined as the argmax of the evaluating agent's own objective under the evaluating agent's own credence, which that agent can compute at the earlier time. So the arm confers no cognition the evaluator lacks, and no object representing a distinct future agent occurs anywhere in the model. What was compared is the principal's contingent plan against the optimal later-measurable plan — the envelope that the previous phase priced and recorded as explicitly not the fully-updated comparator. Skeleton v2 §4 declared that comparator a hole and warned that careless invention is how it collapses; the round invented carelessly in the way it had been warned against.

Three consequences, all recorded rather than repaired by assumption. The dominance result carries no fairness hypothesis and is ∑ maxima ≥ ∑ anything; its original docstring described a statement that did not exist. Its real driver is future-agent infallibility, not "epistemic improvement only" — a witness with every fairness condition intact, in which a better-informed but fallible future agent makes the gap strictly negative, is now carried. And the observation that no jurisdictional term appears in the arithmetic was guaranteed by construction: the specification waived the null effect and the whole execution layer, which the previous phase recorded as the place all of protection's valuation content sits.

Verdict: not well-posed as constructed. No claim is made that fully updated deference is false, or that jurisdiction has low value; both were outside what the model could see. A FUD proof round is not to be dispatched. A successor needs two things, and they are the same two prerequisites the previous phase already named: a future agent with independent existence, so that better-informed and correct can come apart; and the execution layer reinstated with a declared null quantity, so that a jurisdiction assignment is something a valuation can price rather than a label on a selection.

What survives. Fifteen kernel-checked theorems, renamed to EnvelopeDominance to match what they prove and reusable by any successor. The fairness apparatus and three confound witnesses, each now moving exactly one variable. The reduction: the gap is the delegation deficit against the later-measurable comparator class, so the credence collapse applies to the same object rather than by analogy, and any credence-free hypothesis bounding it is it. And the confirmation that underwriting is absent from the engine.

A repeated classification error is recorded. The round's competence slot was labelled as the previous phase's credence-free hypothesis; it compares grades to a conditional expectation, so the credence occurs in it and it is a joint competence–credence hypothesis under skeleton v2 §2a. This is the same error the competence track caught for grade trust, made again. A mechanical check — does the hypothesis mention the credence? — would catch both.

The specification is kept, corrected, as a defective record rather than withdrawn: its fairness apparatus is reusable and the defect is the round's main finding. It is marked as not a binding input to any proof attempt.