No artifacts match that
Try a model name, an author handle, or clear the chips above.
Still not in the open
Everything above approximates the interface (state + typed options → one-pass probabilities). None of it has these:
- TypeSafe's Jev weights. Every "openjev" / "-RLCD" Hub repo is either a stock model plus decoding code, or an independently trained head.
- The RLCD algorithm, loss or data. Reinforcement Learning for Calibrated Decisions is described only at blog-post level; no public implementation.
- A trained open model matching Jev's calibration claims. The best open scorers now reach ~90% agreement with Jev (Bespoke Nimble; snapjudge measures 81.2% for Qwen3.8-27B against Jev's own 87.7%), but they still softmax over candidate logits and the studies find confidence does not reliably flag errors.
- Archer Hume's promised open-weight version. Announced with the architecture post; training "takes time".
