Separating global ocular-score differences from event-boundary specificity with crossed randomization

Boundary-specific group inference requires evidence for a group contrast and exceptional alignment with annotated locations. We combine participant-label and boundary-location randomizations of one estimand with an intersection-union decision. An external ocular benchmark was trained on public Brain, Body, and Behavior Dataset sessions. Crossed participant-and-video validation used 52 participants, 251 paired participant-video units and 502 records, distinguishing attentive-first from distracted-rewatch labels (area under the curve = .877, 95% confidence interval [.817, .932]). The benchmark was frozen before application to a commercial-film sample of 104 participants in four preceding-video groups. The random-short-video group differed in full-film cohort-relative score (contrast = -.7539, 95% confidence interval [-.7737, -.7284], participant-label permutation p = .00010). At 18 annotated boundaries, the boundary-minus-pre contrast was -.04284 (95% confidence interval [-.1422, .0595]). The label test gave p = .00130, but the location test gave p = .52255; the intersection-union p value was .52255. Only 72 of 120 raw-intersubject-correlation specifications were negative, with a sign reversal at 20-second windows. Simulations recovered the crossed decision under baseline conditions but showed lower power and label-component sensitivity to smooth nonstationarity. The archive supports a global ocular-score difference conditional on the compound session-label benchmark. Boundary specificity remains unsupported, and the transported score is not an absolute attention probability.