Before you dive in
What you’ll find in the original
- Filter candidates deterministically for capability, health, data class, context, deadline, and memory before a model ranks them; then validate the returned option and eligibility again.
- Load Laya through its published runtime so the custom decision head is present, reject over-budget inputs instead of truncating them, and provide abstention plus a permitted fallback.
- Freeze thresholds on a development set before replaying held-out cases, and budget for host-load variance: the author's p95 rose from 143.54 ms to 466.69 ms with the same decisions.
Worth knowing
This is one deployment on an M3 Ultra. The author did not compare Laya and Jev on the same 40 cases; the reported 37/40 result is against a deterministic router, and the cases are correlated.