A small, jev-like (“System One”) one-pass option scorer for GUI form filling, trained to work as the decision layer behind cua-driver.
Unlike an autoregressive LLM, this model does not generate text. Given a UI element and a list of typed options (one option per document entity, plus check/click/skip), it returns one probability per option in a single forward pass — the same input/output contract as TypeSafe’s Jev.
Architecture: 706,048 trainable parameters, 2.8 MB checkpoint. Head-to-head vs hosted Jev-latest (zero fine-tuning): 99.7% overall accuracy vs 83.6% for Jev on this task.
Limitations: Only chooses among entities a PDF/document extractor already found; trained on synthetic forms plus small real eval; byte-level encoder, English-centric; not calibrated with TypeSafe’s RLCD method.