looskis / tarski

git clone https://github.com/looskis/tarski.git

Reading diffusion language models for decisions: same OpenJev model, one pass instead of 3.65, +5 points, calibrated, no training. Plus cheap task branches on a frozen encoder.