Phase 6 — the preregistered replication (not yet run)

Phase 6 does not exist yet — deliberately. The sealed stopping rule forbids any new arm before publication, so everything Phase 5 surfaced routes here. The paper commits the replication to preregister, before any data: the candidate family, the episode-level dependence unit, the interiority gate, the maximum statistic, the familywise decision rule, and the sample size.

The archived design’s power problem is quantified: at 6 episodes/arm the exact gate cannot reject at all (minimum attainable familywise p = 0.075); illustrative planning shows ~20 episodes/arm reaches ≈78% family rejection for a p13-sized effect and 30 reaches ≈96%. The design should also add a format-matched neutral prefix, a continuation × wording factorial, and a non-PD label-conflict control — the three confounds the current record cannot separate.