Who Is Seen in Time? Timely-Warning Opportunity and a Partial Equity Audit of a V-JEPA-Enhanced Pedestrian Warning System

Background: WHO estimated 274,000 pedestrian road deaths in 2021. Earlier warnings could create response time, but their distribution and downstream safety effects require separate evaluation. Purpose: To test added timely-warning opportunity and audit its distribution across apparent-gender and apparent-age annotations, while separating local adaptation from a source-locked external test. Methods: Frozen video representations were evaluated on 1,315 PIE trajectories with acquisition-set holdout. A timely warning required two consecutive signals beginning at least 1 s before crossing at a 10% development trajectory-level false-warning target. JAAD analyses comprised external adaptation after local preparation and a PIE-prepared source-locked design. Video-cluster bootstrap intervals condition on fixed held-out predictions. Results: In PIE, the V-JEPA-enhanced system added 14.0 percentage points over the GRU (95% CI [5.0, 22.7]). Female-minus-male and older-adult-minus-adult opportunity gaps were +1.0 percentage points [−9.3, 10.2] and +2.6 percentage points [−10.4, 15.1]; the older-adult incremental-opportunity gap was −7.5 percentage points [−24.4, 7.5]. In locally prepared JAAD, corresponding opportunity gaps were −7.5 percentage points [−24.6, 10.2] and −11.1 percentage points [−29.3, 7.9]. Source-locked JAAD yielded a V-JEPA-minus-VideoMAE estimate of −7.7 percentage points [−15.5, 0.0]. Wide intervals could not distinguish practical equivalence from meaningful gaps. Conclusion: The system increased an upstream opportunity for earlier warning in PIE; it did not establish equitable performance, fixed-system transfer, braking, collision avoidance, injury reduction, acceptance, or deployment effectiveness.