JevBERT
Unofficial proof of concept serving Jev's /v1/systemone shape from a local BERT-style zero-shot classifier, with honest caveats.
Overview
JevBERT is an unofficial, independent proof of concept aiming at a Jev-compatible, instruction-conditioned decision engine built on a BERT-style encoder. At its current P0.5 stage it is a local HTTP server that accepts and returns the same JSON shape as Jev's POST /v1/systemone, and the official typesafe-sdk 0.7.0 works against it by changing only the endpoint and key. A separate layer checks structural and numeric invariants independently of model quality.
The README is explicit about what it is not: the backend is a public zero-shot NLI classifier (bge-m3-zeroshot-v2.0) plugged in temporarily, not a trained JevBERT; probabilities are uncalibrated, so thresholds used with Jev such as confidence above 0.8 must not be copied over; quality is unevaluated beyond a 52-example smoke test; and it has not been checked against the real Jev server. Compatibility is documented in three tiers, starting with what has been verified over real sockets against the SDK, such as the default model alias, request ids, error attributes and 422 detail arrays, and a numbered list of requests it refuses that Jev might accept. Documentation is in Japanese.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Other | Noul, Choice and Score requests answered by a zero-shot NLI backend in Jev's response shape | — |
How it works
- Client sends a Jev-shaped request via the official SDK
- Local server validates and maps questions to the NLI backend
- Returns Jev-shaped answers (uncalibrated)
- Invariant checks verify structure independently of quality
Strengths
- Drop-in endpoint for the official SDK
- Thorough documentation of compatibility gaps
- Runs locally
Caveats
- Not a trained model; uses a stand-in NLI classifier
- Uncalibrated probabilities
- Not verified against real Jev
At a glance
- Type
- Service
- Status
- prototype
- Language
- Python
- Runtime
- Local HTTP server
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
Unofficial PoC: a BERT-style encoder decision engine behind a typed-decision…
wh000wh000/awesome-jev-liveA local server that speaks Jev's `/v1/systemone` shape from a BERT encoder, with a numbered account of every request it refuses that Jev might accept.
MrJev/awesome-jev
Frequently asked questions
Is this a trained Jev replacement?
No, it currently uses a public zero-shot NLI classifier as a stand-in backend.
Can I reuse my Jev thresholds?
No, probabilities are uncalibrated and the README warns against porting thresholds.
Does the official SDK work with it?
Yes, typesafe-sdk 0.7.0 works by swapping endpoint and key, as tested over real sockets.
Related projects
Apache-2.0 ModernBERT decision model answering Noul, Choice, Score and multi-label questions, with a local Jev-compatible server.
ServicebetaPythonGo harness for System One models: hosted Jev, open Kev in-process via llama.cpp, and Cloudflare Clef, with typed helpers and a CLI.
ServicebetaGoTypeScript library for one-line local decisions with open models (kev-0.6b, kev-4b, open-jev), no API keys or cloud.
ProjectbetaTypeScriptSwift 6 bridge that runs System One decision models behind Apple's Foundation Models API, on device with Laya or via hosted endpoints.
ProjectbetaSwiftZero-dependency npm package that runs Laya natively or as WebAssembly in Node, Bun and browsers, drop-in for Jev's /v1/systemone.
ProjectbetaJavaScriptLocal-first desktop AI workbench for llama.cpp, vLLM and SGLang that includes a JEV typed-decision page and /v1/systemone gateway.
Servicebeta