An On-Device, Single-Athlete Pose-Estimation Framework for Injury-Aware Performance Development in a Maturing Youth Footballer

Markerless pose estimation has reached an accuracy where smartphone cameras can, in principle, replace the optical motion-capture laboratory for many sports-biomechanics tasks. Theunsolved problem is not detection accuracy but epistemic discipline: video-derived kinetics —joint moments, ground reaction forces, mechanical power — are model estimates that currentconsumer tools either omit or present with false precision, which is dangerous when the user isa 13-year-old athlete near Peak Height Velocity whose deceleration mechanics predict anteriorcruciate ligament (ACL) injury. Existing markerless pipelines (Pose2Sim, OpenCap, Theia3D)target laboratory or multi-camera desktop settings, assume cloud or PC compute, and reportpopulation statistics rather than the within-athlete, left-versus-right trends that a coach actson. We propose Honesty-by-Construction (HbC) biomechanics, a design discipline in whichevery reported quantity is bound at construction time to an explicit epistemic class — measured, derived, or estimated — carries a propagated uncertainty band, and cannot be renderedwithout it; we instantiate HbC as IRA7-Biomech, a fully offline Android pipeline (BlazePose→ confidence-gated filtering → de Leva body-segment model → five coaching pillars) builtaround one specific inverted-winger whose left leg is the plant-and-brake limb. We show analytically and from the markerless-validation literature that commodity high-frame-rate captureresolves ground-contact time to within ±10ms, that kinematics-driven GRF preserves trendand asymmetry-validity even where absolute magnitudes carry 10–20% error, and that a singlesynchronized second phone recovers the frontal-plane knee-valgus signal that dominates ACLrisk. The result reframes measurement honesty from an apology into the central engineeringvirtue that makes phone-based biomechanics safe to put in a youth-development setting.