An external LLM isn't static code — a provider update or context shift can quietly degrade an agent that triaged AML alerts perfectly for months. Every cycle, score it against a golden, human-verified canary corpus. When accuracy drifts below its verified baseline, the agent degrades gracefully — autonomous → human-escrow → read-only — so bad data never touches the System of Record. Baseline 99%; warn at −5%, downgrade at −10%, floor 80%.
Honest scope. This is the deterministic scorer + decision; running the canary corpus on a schedule (the "shadow digital twin") is the caller's job. The downgrade is monotonic — a bigger accuracy drop never yields a more permissive mode — and an empty/unmeasurable canary fails safe to human_escrow, never silently "stable". It complements the sigmoid reputation kernel (long-run track record) and the breaker (runaway velocity). Source: flowdesk-drift-kernel.js.