jev-route
Router that uses Jev's cloud decisions today, logs every decision with probabilities, and distils a local routing model from your traffic.
Overview
jev-route (mcftira) is a router you run on TypeSafe Jev's calibrated cloud decisions with no training data, while it turns your own traffic into a local routing model. Every routing decision is logged with its full probability distribution, so the log becomes a labelled dataset; once enough traffic accumulates, you distil a local model, evaluate it against the teacher on held-out decisions, and graduate to an air-gapped router. The motto is run it, log it, distil it, own it.
It routes on two axes at once, task difficulty and data sensitivity, plus a PII judgment and a task domain. The pipeline excerpts the request, runs a local gate with a deterministic layer that is never bypassed and an optional local semantic layer in shadow mode, redacts, asks a decision backend (Jev, mock, distilled, Laya or shadow), merges gate floors, escalates on uncertainty, evaluates policy into a tier and model, and logs everything with soft probabilities. Version 0.2 adds hardening, including a tier prefilter, with measured numbers. It is Apache-2.0 and not yet on PyPI.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Score | How hard is this task? | — |
| Score | How sensitive is this data? | — |
| Yes / no | Does this request contain PII? | yes · no |
| Choice | Which domain does this task belong to? | — |
How it works
- Excerpt the request and run deterministic local gates
- Redact and ask the decision backend
- Merge gate floors and escalate on uncertainty
- Map to a tier and model via policy
- Log the full distribution for later distillation
Strengths
- Path from cloud decisions to an air-gapped local router
- Routes on sensitivity as well as difficulty
- Deterministic gate that cannot be bypassed
- Multiple backends including Laya
Caveats
- Distillation quality depends on accumulated traffic
- Not yet published on PyPI
- Early release
At a glance
- Type
- Project
- Status
- beta
- Language
- Python
- Runtime
- Router library and tooling
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
\*\*Run it. Log it. Distill it. Own it.\*\*
logicrw/awesome-jev-projects
Frequently asked questions
What does Jev decide?
Task difficulty, data sensitivity, presence of PII and task domain for each request.
Can I stop using the cloud?
Yes. The logged decisions train a local distilled router you can graduate to.
Is any check model-free?
Yes. A deterministic local gate runs first and is never bypassed.
Related projects
OpenAI/Anthropic-compatible Go gateway where Jev or Laya rates topic, complexity, risk and privacy to pick the cheapest adequate LLM.
BenchmarkbetaGoSix side-by-side demos where Jev makes the decision, Claude or Kimi writes the words, and code owns control flow.
AppprototypeTypeScriptSystem 1 ingestion gate that asks Jev intent, risk and sensitivity in one call, then blocks, answers deterministically or passes to an LLM.
ProjectprototypePythonResearch code showing confidence cascades let a 0.8B prefill-only decision model approach a 4B model's accuracy at a fraction of the cost.
ProjectresearchPython0.8B open System One model with task adapters that makes millisecond decisions in front of a larger local model like Qwen3.8-27B.
ModelbetaPythonDependency-free Python wrapper that turns any LLM API into a Jev-style choose-one-option decision call for old machines.
ModelprototypePython