Assess a conversation against policies

Answers one question, per policy: how confident is the model that this policy applies to this conversation? The number you want is policies[].calibrated_probability.

Two passes, only the gauge is required

• The gauge scores every policy you send, independently (a yes/no plus a decision margin read from the served logprobs), and a frozen two-tier calibration map turns that margin into calibrated_probability, verdict, calibration_tier, and accepted. This is the whole product and the common, fastest path.
• The matcher is a separate, optional pass. Set include_evidence: true and it runs in parallel with the gauge and attaches evidence_ids (the supporting message IDs). It adds roughly no wall-clock latency but is a second full inference, so leave it off unless you need provenance.

The call is stateless: nothing is persisted, reference_id is not an idempotency key, and replaying the same request re-runs inference. Policies come back in the same order you sent them — policies[i] in the response corresponds to policies[i] in the request.

Limitations and correct use

• Domain. Tuned for customer-support conversations in English. Other languages, other conversation genres, and non-conversational text are unmeasured.
• Estimating a rate. To measure how often a policy applies across many conversations, average calibrated_probability over the cohort — do not count yes verdicts. The model deliberately favors precision over recall, so verdict counts undercount the true rate; the calibrated probabilities do not.
• Tracking a trend. Reported movement is a conservative floor — magnitude is compressed, so the direction of change is more reliable than its size. Compare periods using the same policy id and description throughout; changing a definition (or refitting per period) mid-series can invert the trend.
• Unseen policies (pooled_unseen). Probabilities are usable for your own policies, but there is no validated accepted decision and trend fidelity is unmeasured — keep a human in the loop.
• Automation vs. review. accepted is intentionally strict; plan for a meaningful share of decisions (roughly a third to two fifths) to fall outside it, and route those to human review — see evidence_ids.
• Labels are model-derived. Scores reflect a model's judgment of your policy description, not human-verified ground truth. See the Conversation insights guide for the full method.