How reliably can language models trained on deliberative transcripts, stated rationales, and value-elicitation outputs distinguish between implementation decisions that are consistent with versus divergent from the normative commitments embedded in process outputs?