Filed under model-safety.

The Six Things You Rate That Never Make It Into the Rubric

Human annotators scoring model responses for safety validation routinely observe six categories of defect the rubric cannot name — prompt-echo privacy leaks, answer substitution, unjustified compliance, trained deflection, register mismatch, and distributional patterns visible only across hundreds of items. The rubric's dimensions measure what the model produced; they cannot score what it failed to understand, should have refused, or has learned to avoid. The model ships; the gaps do not.