Stage IV: the future agent is still not in the model, and the reason is the signature

Taken at the close of Stage IV, after an independent adversarial review found a conceptual collapse with no cheap repair. The round's positive reading is withdrawn.

The later agent is still derived. It was given its own credence so that it would maximise its own expectation rather than the evaluator's. But its rule differs from the evaluator's conditional argmax by exactly one argument, and remains a total function of objects known at the earlier time. In the round's own headline instance the transferred arm's realisation is constant, so the evaluator knows the realised action — the property the round's gate existed to rule out. The check meant to catch that could not fail.

Jurisdiction does no mathematical work. Setting the principal's credence to the later agent's, with the full-signal interface, makes the delegated arm identical to the transferred arm at every one of 32,805 instances tested. The transferred arm is a coordinate in the delegated arm's parameter space, and the jurisdiction assignment occurs in no formula.

The dominance result is the previous round's theorem with the arms swapped. Stage III put the evaluator's argmax on the transferred side and the transferred side trivially won; Stage IV puts it on the delegated side and the delegated side trivially wins. Its scan is padding: 19,468 of 26,244 instances contain no fallible later agent at all.

The controlling finding, and the reason both rounds failed. Two authorisation regimes that induce the same realisation map are the same object in a signature whose only outputs are such maps priced by one measure. This is a type-level obstruction, not a modelling slip: a jurisdiction assignment is exactly what that signature cannot express, and no additional parameter recovers it. The authorisation relation has to enter the type.

Consequences. No FUD proof round is to be dispatched, and no further comparator round of this shape: two attempts have now failed at the same place from opposite directions. The claimed-gate harness is deleted rather than repaired; diagnose_collapse.py replaces it and every check in it records a defect. FUTURE_AGENT_SPEC.md is kept as a corrected, collapsed record and is not a binding input. Three further round claims — the advice-loss story, the interior requirement, and the fairness accounting — were checked and are false or overstated, and are corrected in place.

A repeated harness failure mode is recorded. Stage III shipped four checks that could not fail; Stage IV shipped ten, including a literal True and an or True. A mechanical lint flagging any check whose condition is a constant or a type test would have caught both.