litjev
⚠️ Why not a model
Wraps off-the-shelf Qwen checkpoints as Jev-schema decisions — probabilities are uncalibrated by default.
Technical specs
- Base LLMs
- Qwen
- Decision types
- choice, score, boolean
- Features
- local-serve, typesafe-compatible
- License
- MIT
Links
litjev is a lightweight method that wraps off-the-shelf Qwen checkpoints to serve TypeSafe-compatible decisions locally.
How it works
litjev provides a POST /v1/systemone endpoint that:
- Takes Jev-schema requests (
choice,score,boolean) - Runs inference on a local Qwen checkpoint
- Returns decisions in TypeSafe-compatible format
Why it’s not a model
litjev does not train or modify any weights. It’s the same class as SemIf and simple-jev:
- Uses frozen Qwen checkpoints
- Extracts decisions from logits at inference time
- Probabilities are uncalibrated by default
History
litjev was briefly listed on the model quadrant but was demoted on 19 Sep 2026 to align with the classification of similar runtime/method projects (SemIf, simple-jev). Projects that only wrap existing LM logits without training decision weights belong in the Runtimes section, not the Models catalog.