Fast Temporal Posture-Only REBA Scoring from Simulated Camera Keypoints with Conformal Risk Alerts

Camera-based ergonomic screening requires continuous posture scores accompanied by explicit uncertainty. This study evaluated ErgoTCN, a causal temporal convolutional network with approximately 160,000 parameters, for estimating posture-only Rapid Entire Body Assessment (REBA) scores from 17 simulated two-dimensional camera keypoints. Wrist posture was fixed at neutral, and load, coupling and activity inputs were excluded. Reference scores were derived from 9.8 hours of three-dimensional motion capture across 113 recorded Carnegie Mellon University subject IDs. Ten virtual camera configurations and a MoveNet-like keypoint-error model generated input streams; no real pose-detector video was assessed. Model evaluation used 28 held-out subject IDs, although distinct-person separation was unverified, and validation, calibration and test sets shared the ten camera configurations. Across three training seeds, ErgoTCN achieved a mean absolute error of 0.874 REBA points and quadratic weighted kappa of 0.781, with mean absolute error 11.1% below the gradient-boosted baseline and 18.5% below the raw-keypoint Random Forest. Empirical coverage of nominal 90% conformal intervals was 0.923, with a mean width of 3.60 scores. High-risk alert recall reached 0.744, below the target of 0.900. A random frame split increased action-level accuracy from 59.1% to 76.0%. Scoring latency was 1.24 ms per frame in PyTorch and 0.307 ms in ONNX Runtime, excluding pose detection. These results characterise simulation-based posture screening, not validated workplace performance.