'Jev in 25 lines of Python' reproduces the decision-model pattern with Qwen3-0.6B logprobs and no training
NobodyWho·medium signal
NobodyWho's post (22 Sep, 191 HN points) gets calibrated-looking typed decisions by reading next-token logits over option labels from a 0.6B GGUF model in llama-cpp-python. Arcturus Labs' 'Will OpenAI eat Jev's lunch' (21 Sep, 295 HN points) makes the same technical argument and says TypeSafe's only moat is its RLCD training data. New repos Nokia's AnyJev and rizzo-flow ship the same no-training approach. This is the skeptic's side of the Jev story. What you would actually pay TypeSafe for is calibration, not the interface.