Candidate-scoring decision models return a probability distribution over a presented set of answers. Sudjianto and Burakov analyze three common training ingredients—soft targets from a larger teacher model, a frozen pretrained backbone, and temperature calibration—using Gaussian surrogates where finite-sample risk can be computed exactly.
Distillation is the main problem. A student trained on teacher probabilities inherits the teacher's bias, and the bias does not shrink as student data grows; agreement with teacher predictions also gives no signal. Fixed representations help only below an explicit sample-size threshold, and in a probit model the exact recalibration map is temperature scaling, but it removes all excess risk only when the teacher's coefficient has the correct direction. Otherwise a residual remains that standard calibration diagnostics do not detect.
On 231,521 Consumer Financial Protection Bureau credit card complaints, the pattern is concrete. A student trained on a logistic teacher's probabilities lands at the teacher's decision cost; a student trained on realized outcomes goes 8% below it. A 30B model read zero-shot remains worse than the teacher even after recalibration and departs from the coherence condition by 17.5 decibans. The authors conclude that pretrained representations and calibration are not enough: candidate-scoring probabilities have to be checked against realized outcomes.