Early-exit models let a neural network stop at different depths to save computation, but conformal prediction sets calibrated at each exit lose their coverage guarantee when the exit is selected adaptively. Because the chosen depth depends on the input, the usual per-exit calibration does not account for the selection step, so the resulting sets can be too small or too large.

The proposed method avoids this by constructing a single terminal path score from prefix maxima of the scores encountered along the path. Instead of calibrating many quantiles, it calibrates one split-conformal quantile on that terminal score. This yields distribution-free coverage even under adaptive depth selection, without needing to know the selection rule in advance.

This is a practical step toward reliable uncertainty quantification for dynamic inference. The approach is simple enough to apply to existing early-exit architectures, and it directly addresses a gap between conformal prediction theory and adaptive computation in deployment.