What it is
Train and serve JEV-style decision heads on Qwen.
What it does: Distills TypeSafe Jev distributions into a LoRA + decision-head student, calibrates by question kind, and exposes a System One-compatible HTTP API.
Why it belongs in the showcase: Clear boundary label, reproducible recipe docs, and an API bridge for local/open weights.
Caveats: Unofficial; hardware-heavy; author-reported numbers.
Evidence
Public Apache-2.0 repo (3★ at intake) with Chinese/English READMEs, docs/, Hugging Face models under shibing624, and v0.1.0 notes: Qwen3.5-4B recipe, distill corpus SargeDev/jev-distill-corpus-v3, Decision Index 0.3 eval, zero-shot image decision numbers in docs/results.md. States independence from TypeSafe AI and autotrust. Serving claims System One API compatibility so existing SDKs can attach.
Caveats
- Community reproduction, not an official TypeSafe release
- Reported metrics are author runs on specific GPUs; verify docs/results.md yourself
- Training needs multi-GPU hardware and large datasets
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.