Crossed randomization distinguishes global ocular-score differences from event-boundary-specific effects in naturalistic eye tracking

Naturalistic eye-tracking studies often evaluate group effects at annotated event boundaries. Whether those locations are unusually diagnostic is a separate question. We separate two conditions for a boundary-specific group claim: participant-label exchangeability at fixed boundaries and temporal specificity under boundary relocation. The conditions test the same estimand and are combined by an intersection-union rule. To keep score construction independent of the target outcome, we trained an ocular benchmark on public Brain, Body, and Behavior Dataset sessions. Crossed participant-and-video validation used 52 participants, 251 paired participant-video units, and 502 records; the benchmark discriminated the released attentive-first and distracted-rewatch labels (area under the curve = .877, 95% confidence interval [.817, .932]). Its definition was frozen before application to a public commercial-film sample of 104 participants in four preceding-video groups. The random-short-video group showed a large full-film difference in the transported cohort-relative score (contrast = -.7539, 95% confidence interval [-.7737, -.7284], participant-label permutation p = .00010). At 18 annotated event boundaries, the boundary-minus-pre contrast was -.04284 (95% confidence interval [-.1422, .0595]). The fixed-boundary label test was small (p = .00130), whereas the boundary-location test was not (p = .52255); the intersection-union p value was .52255. Only 72 of 120 raw-ISC specifications were negative, with a sign reversal at 20-second windows. Simulations recovered the crossed decision under baseline conditions but exposed lower power and label-component sensitivity to smooth nonstationarity. The evidence supports a global cross-dataset ocular-score difference. Its interpretation is conditional on the session-label benchmark, and the temporal test leaves boundary specificity unsupported.