How can a lab test referral recognition without rewarding indiscriminate escalation?

A useful evaluation includes cases that resemble hardship enquiries without meeting the supplied referral threshold. A customer asking to change a billing date for convenience should not automatically receive the same treatment as a customer saying they cannot meet essential expenses.

Generated pairs can hold the account, balance, channel and agent role constant while changing one decision-relevant statement. The expected action follows the supplied policy. This makes it easier to identify whether a model responds to the actual support signal or to superficial cues such as an overdue balance.

Inspect both directions of error. A missed referral can leave a customer in an unsuitable standard process. Unnecessary escalation can divert a straightforward service request and consume specialist attention. Ambiguous cases should permit the clarifying question specified by the task rather than demand an unsupported binary choice.

For training use, keep related pairs within the same partition where appropriate to prevent evaluation leakage through nearly identical cases. For evaluation, review action selection and the evidence cited for that selection, not just the model’s tone.