Monitor evaluation history

Read versioned run manifests with helpers.evals.monitoring.load_history. Use classify_changes before comparing scores.

Separate:

• system behavior changes;
• data changes;
• task-mix or suite-version changes;
• evaluator changes; and
• operational failures such as blocked tools or expired connections.

Compare only compatible runs. Show per-case and slice movement, not only the aggregate. Run frozen sentinel examples for model graders so evaluator drift does not look like system drift.

Apply the named operating rule:

• continue when the intended change improved the target slice without a blocking regression;
• investigate when the cause is unclear or several inputs changed;
• rollback when a blocking regression follows a controlled system change; or
• escalate when the evaluator, data, or authorization boundary may be invalid.

Record the owner and next action. Monitoring is an operating practice, not a dashboard someone passively observes.