Compliance guard
Check every generated response against your disclosure requirements and prohibited-claims list before it is spoken. Log the verdict with the input that produced it.
Why this is the clearest case for a local model. A compliance check has to run inside the same latency budget as the response it is checking. A second network round trip to a hosted API is not available. This is the one capability where a frontier API is not merely more expensive but architecturally wrong.
What it does. Classifies a candidate response against your rules: required disclosures present, prohibited comparative claims absent, suitability language intact, script correct for the target system. Returns a verdict and a probability you can threshold.
What gets logged. The input, the candidate response, the verdict, the model version and the threshold in force. This is the artefact that answers “why did it say that” a year later, and it is the reason the whole architecture is worth the trouble.