Jevcal: choose confidence thresholds from labeled data

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.

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.

An independent API-based audit of Jev that examines how confidence relates to observed correctness, including experiments with abstention. The published scripts and results are research evidence, not a guarantee for other datasets.