Conformal Prediction and Epistemic Uncertainty Decomposition for Safety-Critical Edge Vision under Dynamic Domain Shift

Deploying deep learning object detectors into safety-critical autonomous cyber-physical systems demands statistically sound uncertainty quantification. While heuristic estimators such as Monte Carlo Dropout quantify epistemic parameter dispersion, they fail to provide finite-sample coverage guarantees, particularly when vision backbones encounter out-of-distribution environmental domain shift. In this paper, we propose a unified calibration framework integrating Split Conformal Prediction with Bayesian uncertainty decomposition for edge-deployed YOLO architectures. Operating within the NeuralForgeAI inference pipeline, our approach constructs non-conformity scores over compound bounding box regression errors and class assignment confidences. We prove that for any user-specified error tolerance alpha in (0, 1), the generated prediction sets guarantee rigorous 1 - alpha marginal coverage under arbitrary data distributions. Empirical evaluations across synthetic corruptions and night-time domain shifts confirm that while uncalibrated softmax confidences collapse to 79.1% coverage at alpha=0.05, the conformalized detector preserves 95.1% empirical coverage with compact prediction set sizes (1.8 classes/box), establishing mathematically guaranteed safety envelopes on embedded edge accelerators.