Reliability should be measurable.
Machine Observer publishes the health of the source layer it depends on. Healthy and degraded sources are separated, unavailable sources remain visible, and recovery history accumulates from scheduled checks rather than being inferred from marketing claims.
Degraded is not the same as healthy.
Operational coverage includes healthy + degraded sources. The healthy percentage is published separately so partial availability cannot be presented as full reliability.
Healthy
Passed current validation checks.
Degraded
Reachable or usable with a freshness, schema, method, latency, or other quality warning.
Unavailable
Cannot currently be used safely as a current observation.
Reliability history starts now and gets more useful with age.
The service keeps compact status-only snapshots for up to 120 days. It does not copy source response bodies into the reliability history.
Retained samples
Scheduled observations available for trend and recovery calculations.
24h operational average
5 samples in the current 24-hour window.
Mean recovery time
Recovery time is measured between scheduled source-health observations, so precision is bounded by the check cadence.
What these numbers mean
Metrics are deliberately conservative. Broken sources remain in history, and Machine Observer does not silently substitute a different dataset to make availability look better.
Operational %: healthy + degraded divided by all registered sources.
Recovery time: elapsed time from a non-healthy scheduled observation until a later healthy observation; precision is bounded by the check cadence.
Checks: status, response type, required fields, freshness, latency, schema change, documented alternates, and valid x402 challenges where applicable.