Intent alignment

Attach structured intent metadata to traces (task type, allowed tools, data sensitivity). Scoring compares declared intent to observed tool calls and span attributes. We never claim access to model-internal reasoning.

In practice: tag agents and assets, declare allowed tools in policy, then review mismatches in Work.

Intent alignment is not a guarantee that the model “thought” the right thing. It is an observable mismatch signal between what you declared and what the gateway saw.