The current documented runtime is local HF/PyTorch
The README describes a local server and /v1/classifier API. A generic claim that every standard inference endpoint works would overstate the reviewed implementation.
Inspect the supporting source ↗REPOSITORY ANALYSIS
simple-jev
Adapter turning open LLM endpoints into Jev-compatible classification services without training a separate classifier head.
E2 · source reviewedPublic material or source evidence was reviewed. A fixed-version source review requires checking a pinned revision; this is not a reproduction, performance endorsement, or safety guarantee.
- Stars
- 507
- Forks
- 58
- License
- Apache-2.0
- Last commit
- 2026-09-24
What Jev does here
Extracts log-probabilities of candidate tokens from model vocabulary logits, formatting them into standard Jev responses.
Lets users prototype typed candidate readouts on compatible open models using the documented local Hugging Face/PyTorch server. The reviewed interface is /v1/classifier, not a blanket guarantee of compatibility with all inference endpoints.
Pinned README and driving-demo request logic reviewed. No model server was run here. Base-model logits, application safety checks and single-eligible-action shortcuts are different evidence sources; their outputs and timings should not be merged into one model benchmark.
What task does this project fit?
Turn a compatible open language model's candidate logits into structured decisions without training a separate head.
Material analysed：
From input to action
- Validate shared state/messages and typed questions.
- Read next-token candidate scores through the local HF/PyTorch implementation.
- Construct the response in server code, with application-specific candidate constraints outside the model.
Implementation details that matter
The driving demo can skip inference for one eligible action
jev-request.js resolves singleton choices locally; simple-jev-api.js rejects an unknown selected path or a collision-imminent moving vector. Model decisions and deterministic shortcuts must be distinguished in timing reports.
Inspect the supporting source ↗When to choose it
Useful when you already serve an open model and can evaluate its candidate readout. It is not an independently calibrated substitute merely because it emits a Jev-like JSON shape.
How to inspect it yourself
Compare the documented classifier API with prepare/restore logic in the driving demo. Record decision_source and candidate IDs when interpreting a run.
Limits of this review
- No model download, server or driving simulation was executed.
- Base-model choice, prompt protocol and candidate ordering can change quality.