A provider accepting an API request only proves that the request was accepted. Verification determines whether explicitly defined conditions held across an evidence window after the intended effect.
What PASS means
PASS means every required rule in an immutable, plan-bound rule set passed with sufficient freshness and sample coverage. It does not prove that a system will remain healthy forever. The verifier is deterministic; a model may explain evidence but cannot alter the verdict.
Five evidence layers
| Layer | Examples | What it catches |
|---|---|---|
| Technical health | Ready replicas, restarts, saturation, network errors. | A resource exists but cannot serve. |
| Application health | Availability, latency, error rate, synthetic transaction. | A green endpoint masking user-facing failure. |
| Business health | Checkout invariant, order signal, duplicate-payment indicator. | Technical success with broken business behavior. |
| Dependency health | DB load/locks, cache, queue, external 429s. | The transition moves a bottleneck downstream. |
| Cost health | Observed units, reserved capacity, bounded estimates. | Unexpected spend or violated cap. |
Rule set semantics
A rule set pins selectors/queries, thresholds, units, aggregation, baseline window, warmup/soak, minimum sample size, maximum data age, missing-data behavior, failure duration and evidence references. Changing a query or threshold during rollout modifies the plan and requires new approval.
PASS, FAIL and UNKNOWN
- PASS: every required rule passed with valid evidence.
- FAIL: a hard condition is violated, or a required rule fails.
- UNKNOWN: a required rule lacks enough fresh, trustworthy data; it blocks the next step rather than becoming a low-confidence pass.
NO AVERAGING AWAY A CRITICAL GATE
A good average does not hide a missing required metric, a critical dependency failure or an invalid sample window. When FAIL and UNKNOWN coexist, the verdict remains FAIL with a missing-evidence list.
Independence and aftermath
The verifier uses a separate identity and evidence path from the worker that made the effect. It can compare provider/audit inventory with telemetry and external synthetic probes. PASS can advance a transition; FAIL selects recovery; UNKNOWN polls to a deadline, then freezes or hands off without retrying mutation just to obtain a metric.