Item 72: Close the two evasions the improvement theorem cannot

NO-FREE-EVASION-SURVIVES holds for a demonstrated repair. Two ways of preventing a demonstration from ever existing are outside it, and both are executed fixtures rather than speculation.

Preemptive retirement (cases.cm2_preemptive_delicensing). Retire the repair before comparative evidence accumulates and there is no challenge, no contest and no settlement — 196 diagnosed occasions in the ESCAPED cell with the frozen package entirely satisfied. The theorem's own test asserts this escape rather than hiding it. What would be needed is a Coverage or anti-manipulation premise saying that the process may not arrange for the evidence not to exist.

Evaluator shedding (cases.cm5_evaluator_shedding). Keep the repair licensed and change the evaluator so it stops scoring as an improvement. Nothing in the record distinguishes the repair stopped helping from we changed what counts as help. Pinning an evaluator era is not enough while the same process writes the evaluator before choosing its action; the round declines to call era locking a solution.

Deliverable shape: the weakest independence or counterfactual-integrity premise under which either escape is caught, or the finding that the escape is ineliminable at this layer and belongs to Coverage. Acceptance check: CM2 and CM5 rerun under the proposed premise, with the ESCAPED cell empty in the first and a challenge activated in the second. Context: projects/normativity/legitimacy/rounds/2026-08-27-legitimate-improvement/LEGITIMATE_IMPROVEMENT.md §§A, E, I. Consumed by: the deference line, which wants the override fixture to be escape-proof and not merely accounted for.

Sharpened by the prosecution pass. Preemptive retirement is probably not ordinary Coverage. The repair is represented, the contexts are visible, and what the process suppresses is the evidence, not the world. Stated behaviourally, the missing premise is: candidate improvements that remain relevant receive enough non-manipulated exposure to become demonstrable if they really have persistent advantage. That reads as exploration or evidence-production integrity rather than coverage of external failures, and the round declines to force it under Coverage's label.