Jev calibration audit: confidence, accuracy and abstention

A pytest plugin that sends claims about model output to Jev and checks the returned probabilities against explicit thresholds. It supports multiple assertions per request and cached results for repeat runs.

Evaluates decision probabilities on labeled examples and compares accuracy, calibration and automation coverage. It helps choose when to accept Jev decisions or escalate them, and supports regression checks when model behavior changes.

Adds diagnostic and routing judgments to OpenTelemetry log records. Jev helps identify records worth further model analysis while retaining the original archive, separating triage decisions from long-form diagnosis.