Separate deterministic canaries from production quality SLOs

Voice-service monitoring needs two explicit evaluation contracts. Synthetic availability canaries should run versioned, deterministic fixtures with component-specific semantic assertions; use a seeded rotating fixture bank to broaden coverage without making failures irreproducible. Production quality alerts should aggregate only over a named eligible population, with cancellations, short calls, synthetic traffic and pre-ready disconnects handled explicitly in each denominator. Keep request-level provider metrics separate from call-level experience metrics, then gate ratios and percentiles by minimum sample volume. Thresholds may be calibrated later, but event definitions and population membership must be fixed first.

Applies when: Designing LLM, TTS, STT or voice-pipeline health and quality alerts. Keywords: voice canary semantic evaluation production SLO eligible population latency