JEV from Scratch

No. The repository states plainly that it is not Jev: TypeSafe has not published Jev's weights or architecture. It is the author's own reconstruction of the publicly inferred design, based on TypeSafe's launch post, Archer Hume's write-up on Jev's architecture, and the open-source Kev project.

The runs were built and tested on a laptop RTX 4070 with 8 GB of memory, using torch 2.9.1+cu130, transformers 4.57.3 and peft 0.21.2. Training runs took roughly 12 to 65 minutes each, with the 1.7B broad2 runs the longest.

Text is encoded with borrowed Qwen special tokens as delimiters, every option restarts at the same position, and a 4D additive mask lets options see the context and themselves but never each other. A pointer head scores DECIDE against each option close token, so no per-label weights are needed.

Choice questions return a probability over the given options, noul is implemented as a two-way no/yes question returning P(yes), and score questions take ordered levels and return the expected level computed as the sum of index times probability.

Metrics are generated by python -m src.report from JSON in outputs, covering accuracy, ECE, Brier and NLL per question type after temperature scaling. The broad run reports 96.3% choice accuracy with 0.022 ECE, and the plus run reports 94.0% on the harder score condition.

Yes. src/api.py and src/serve.py provide POST /v1/systemone with the same request and response shape as TypeSafe's system_one. An optional packed inference path puts one state and many questions in a single sequence with a state KV cache, matching separate rows to about 1e-6 in fp32.

It is a laptop-scale reconstruction, not an official model, and the reported numbers come from a single seed with 300 questions per condition unless noted. Dataset licences are documented in DATA_LICENSES.md, and the lab notebook records every experiment, number and mistake in order.