Today a guardrail hit is only visible by comparing the model's draft with the message the guest actually received.

Request: log every guardrail hit as its own session event, with the word or rule that matched, the action taken (replaced, blocked) and the channel. Ideally also a report of hits per guardrail.

Impact: partners can see how often a guardrail fires, tune prompts accordingly, and explain to a customer why a guest received a changed message.