SPARC Single-Pass Scaling for Calibrated Motion Forecasting

SPARC introduces a Bayesian-conformal uncertainty layer that uses a conjugate Bayesian last layer to predict epistemic scales. This method provides 95% marginal prediction tubes with finite-sample validity while preserving trajectory covariance structure in a single pass.