A machine verdict (pass/warn/fail) with issues grounded in a Metric — name, window, observed vs threshold, burn-rate. Deterministic, no key needed.
Grade telemetry against an error budget — must: error_rate < 1%, must: error_budget_remaining > 20%, should: p99(latency_ms) < 300ms.
analyze adds anomaly detection + an optional LLM critique. Egress happens only here; check stays offline by construction.
watch grades behaviour over a window — slow latency regression, resource leak, error spike, flatline/liveness — bounded by construction.
vitel demo grades a synthetic degrading service, prints a FAIL report (error budget burning — no LLM key), then grades the fixed version and prints PASS.
It renders — visual feedback, grounded.
It sounds — audio feedback, time-grounded.
It's healthy — metrics, SLOs, error budgets.
It's done — decide, compound verified work.
All share one contract, agentsensory — one Verdict, one Report, one Handoff.