What problem does it solve? Voice agent evaluation metrics often produce false positives, false negatives, or inconsistent results after initial deployment. This Skill guides the iterative cycle of finding misaligned metric results, leaving structured feedback, running the Cekura labs auto-improve pipeline, and validating changes so metrics reach production quality. ## Core Features & Use Cases - Misalignment Detection: Review recent call evaluations to find false positives, false negatives, and unexpected N/A results across binary, workflow, and enum metrics. - Structured Feedback Workflow: Submit agree/disagree votes with transcript-grounded explanations via the mark_metric_vote endpoint, accumulating 6+ instances before auto-improvement. - Labs Auto-Improve Pipeline: Trigger process_feedbacks to generate improved metric prompts, then re-run evaluations on the same calls to validate fixes without regressions. - Use Case: A metric measuring one-question-per-turn keeps failing valid calls. Categorize the failure root causes, PATCH manual prompt fixes, leave feedback on remaining edge cases, run labs auto-improve, and re-evaluate a sample of calls to confirm accuracy. ## Quick Start Ask the assistant to review recent call evaluations for a specific metric, identify misaligned results, and leave feedback to start the labs improvement cycle.
