KFChow

Per-turn LLM routing: on each matching turn a TypeSafe Jev classifier call (api.typesafe.ai) receives turn-shape features plus the first 1200 chars of the last user message (shadow mode included) and labels it free/paid; high-confidence routine turns are downgraded to the provider's configured free rung, strategic mid-rung turns on Nous are escalated to the configured premium rung after a 2-token billed credit probe (Nous only, fail-closed). Fail-open, shadow-by-default, per-provider rungs from the user's lane2_config.json (leaderboard resolver script optional). Probe model id is fixed in code (z-ai/glm-5.3-flash). Disclosure — once you list models in ~/.hermes/jev/lane2_config.json, the first call of each matching turn sends message/tool counts and the first 1,200 characters of the last user message (send_excerpt, on by default) to TypeSafe Jev at api.typesafe.ai with your JEVI_API_KEY, in shadow mode as well as live; in live mode Jev's answer switches the request (and every later call of that turn, via an llm_execution wrapper) to the free, flash or premium model you configured on the same provider; with escalation set up on Nous it sends a 2-token billed ping to the Nous inference API with NOUS_API_KEY at most every 5 minutes; every turn appends a decision row (with that excerpt on routed turns) to ~/.hermes/jev/lane2-live.jsonl, which is never rotated.