Interpretation and arithmetic
A health score measures documented operational checks, not a probability of failure or an availability SLA. Within each applicable category, earned points equal category weight times the sum of severity-weighted credits divided by the sum of check weights. A 20-point resource category with one critical passing check and one warning failing check earns 20 × 3/5 = 12 points. Explicit category exclusions rescale all remaining category weights before this calculation. Ceilings apply after summing contributions.
Conservative evidence handling
An unavailable source must not improve a previously failed check. Stored failures retain their impact and become unverified. A historical daily finding remains distinct from an endpoint observation until its own check reports remediation. Acknowledgements and private incident grouping never modify evidence or score arithmetic. Backup age is separate from exporter freshness: receiving a fresh backup-status metric cannot establish that the backup itself is recent. Ages without an approved cadence are visible measurement gaps.
Attribution limits
Unassigned Cloud rules are listed as a fleet coverage gap instead of being guessed onto a server; critical firing unassigned rules still appear prominently. Dependency propagation is limited to verified essential availability relationships. A host backup failure does not imply its application endpoint is offline. Additional validated coverage can change the denominator; configuration revisions and historical snapshots make those changes auditable.