Pose-derived image-plane geometry for expert-rated sports performance quality: leakage-controlled development and independent evaluation on AQA-7

Background Action quality assessment (AQA) seeks to estimate how well a sports action is performed, but high-capacity video models can obscure what information drives a score. This study evaluated whether transparent pose-derived image-plane geometry provides reproducible information about expert-rated sports performance when development, model selection, and independent testing are strictly separated. Methods The AQA-7 benchmark contained 1,189 videos and 165,212 decoded frames. Official Split_4 assignments were preserved (803 development, 303 final test; 83 trampoline videos were outside the supervised split). MediaPipe Pose Landmarker 1.0.1 processed all videos on CPU in VIDEO mode; 48,344 frames (29.3%) returned at least one pose. Outcome-independent pose and geometry quality control retained 131 development samples for feature construction, 108 after the fixed 57-feature completeness lock, and 102 inferential samples (48 diving, 54 vault) for repeated nested validation. Action-only, geometry-only Ridge, and geometry-plus-action Ridge models were compared. After all development choices were frozen, score-blind final-test features and predictions were locked before the official final-test scores were linked for predictive evaluation; 33 diving/vault samples were evaluable. Results In development, geometry-only Ridge showed modest rank association (macro SRCC 0.231, 95% CI 0.039 to 0.417; Pearson r 0.227, 95% CI 0.020 to 0.409) but negative R2 (-0.574). On the independent final test, the same locked model did not replicate (SRCC -0.116, 95% CI -0.499 to 0.375; Pearson r 0.042, 95% CI -0.355 to 0.404; R2 -11.986; standardized MAE 1.322; RMSE 1.778). The action-only comparator had substantially lower independent-test error (R2 -0.076; MAE 0.478; RMSE 0.597). Development geometry validity was strongly action dependent (Cramer’s V 0.412), but the action-adjusted valid-minus-invalid expert-score difference was small (0.047 SD, 95% CI -0.148 to 0.233; permutation p = 0.874). Conclusion Frozen image-plane geometry carried limited development-set ordinal information but did not provide independent predictive value beyond action-specific baselines. Pose availability and eligibility were strongly sport dependent. The revised evidence therefore supports the pipeline as a transparent methodological audit of pose-derived AQA, not as a validated automated scoring system.