Class-Conditional Regret Offsets for Budgeted Early Exits: A Multi-Seed Pilot on Handwritten Digits

Early-exit networks can reduce conditional inference work by replacing a subset of full-depth predictions with shallow classifier outputs. This exploratory preprint examines a calibration-only, class-conditional correction based on estimated regret when shallow and final predictions disagree. Using a two-exit PyTorch multilayer perceptron and 12 stratified splits of the 1,797-example scikit-learn handwritten-digits benchmark, we evaluate clean and synthetic-corruption scenarios under three levels of class imbalance, with confidence and margin early-exit baselines. On clean balanced data at a 60% calibration exit target, the examined route attains 96.55% mean classification accuracy and reduces estimated dense-layer multiply-accumulates by 43.72% relative to full-depth execution of the same two-head network, or 41.68% relative to a final-head-only backbone. Standard confidence routing achieves comparable results. One synthetic noise/imbalance setting shows a modest condition-specific advantage, but the explored correction does not consistently improve accuracy or class-level performance. We report all results and runnable source, including the limitations of a small dataset and operation-count rather than wall-clock evaluation. This is a non-peer-reviewed exploratory research note, not a claim of state-of-the-art performance. Generative AI assisted with research prototyping, implementation, analysis, and manuscript drafting.