{"description": "Daily ML serving reliability dashboard. Probes 30+ ML endpoints (forecast, classifier, anomaly, measurement, sentinel) with HTTP availability, p50/p95 latency, schema validation, model staleness, and honest-caveats presence.", "endpoint": "/v1/atlas/serving-reliability", "honest_caveats": ["10 probes per endpoint is a small sample \u2014 p95 latency is noisy at n=10.", "Probes hit localhost upstream (forecast_api.py on [internal endpoint]); CF gateway / DNS / TLS issues are NOT detected here.", "Schema validation is a top-level-keys superset check, NOT a full JSON-schema validation.", "Model staleness is the most recent of either declared trained_at JSON field or sidecar mtime.", "MCP server endpoints (Node.js runtime) are NOT included.", "Daily snapshot \u2014 short-lived blips between cron runs are not captured."], "query_params": {"group": "Optional. One of forecast | classifier | anomaly | measurement | sentinel | atlas. Filters endpoints to that group."}, "refresh": "daily at 04:00 UTC via cron", "sidecar_last_modified": "2026-10-05T04:04:27Z", "sidecar_path": "[internal path]", "sidecar_size_bytes": 40473}