CMP.7:6 - Bias-Annotation

Training loss and benchmark performance are visible and easy to optimize. The recipient may instead need a response in a different region or after an intervention. Keep that target explicit so that improving the visible criterion does not silently replace it.

A broad function class or large pretrained model can hide a strong inductive restriction in its representation, training history and obtaining procedure. The relevant question is which continuations it favors and how that choice fits the receiving use. Human-like explanations of a model’s behavior do not by themselves supply its learning guarantee.