kyegomez’s Open Jev is an unofficial PyTorch sketch of a decision model inspired by TypeSafe’s public descriptions. It uses random weights, so its outputs are demonstrations of data flow, not useful judgments.
One state, separate questions
- A state encoder turns nested input into a representation that can be cached.
- The model folds questions into the batch dimension, then lets each question’s learned slots attend to the shared state.
- Separate heads return a yes/no probability, a choice among supplied options, or a distribution over ordered score levels.
- A separate confidence head and a calibration-oriented loss illustrate possible training mechanisms; neither establishes calibrated uncertainty here.
What the sketch can and cannot show
The choice head selects from the declared options. That constrains the shape of its answer, but it does not make the choice correct. The repository also uses a hash tokenizer and lists trained checkpoints, calibration tests and scaling benchmarks as future work.
Its architecture and training recipe are hypotheses, not disclosures of Jev’s actual internals. This repository is distinct from the separately published Open-Jev model family.
Verification: inspected the README and model implementation at commit 93843ef288f3d35fd91263377c40b6f917d77da0. No training, inference, benchmark or security audit was performed.