A parks recreation assistant must quote fees correctly from the fee schedule, not only sound fluent. What evaluation addition helps?

Adding a custom metric for correct fee quotation from the fee schedule ensures a parks recreation assistant is graded on numeric and policy accuracy, not merely fluency. Parks bots that sound friendly while inventing picnic shelter prices create billing disputes and equity issues. A custom metric compares quoted fees to the authoritative schedule (including seasonal or resident/non-resident tiers) and fails answers that mismatch, giving GenAIOps a promotion gate tied to the civic source of truth.

Adding a built-in fluency metric for fee quotation from the fee schedule fails because it only measures how natural the reply sounds, not whether the quoted fee matches the schedule, so it misses the exact resident task that matters for registration. Adding a built-in coherence metric for fee quotation from the fee schedule fails because a response can flow logically and still state the wrong price, leaving the billing-dispute risk the metric is supposed to catch. Adding a custom metric for correct fee quotation from prior chat turns fails because agreement with prior turns does not confirm that the quoted picnic shelter price matches the authoritative fee schedule.

Exam caveat: Custom fee metrics must version with the fee schedule; mid-season price changes invalidate old expected values. Handle formatting variants (currency symbols, ranges) carefully so the metric measures correctness, not string identity alone.

Operational check: Implement the custom fee-quotation metric in the eval harness against the current schedule; include resident and non-resident tier cases; block promotion on fee-mismatch rates above threshold; alert when the schedule document updates without a corpus/metric refresh; and spot-check production transcripts during fee-change weeks.