Calibrate a Coval metric

Measure whether a specific metric version matches a domain expert on its intended population. Successful metric execution and agreement on prompt-tuning examples are not independent validation.

Establish the evidence contract

Confirm organization/workspace, one metric ID, rubric, output type/polarity and decision. Read the definition and output version IDs. Inspect an existing review project and completed annotations, or a human-provided labeled export:

coval --agent metrics get <metric-id>
coval --agent review-projects get <project-id>
coval --agent review-annotations list --filter 'project_id="<project-id>" AND metric_id="<metric-id>" AND completion_status="COMPLETED"' --page-size 100