Jev Explained in 25 Lines of Python

Summary

This parody technical blog post argues that Jev can be demonstrated with a compact local classifier rather than the more elaborate framing often associated with it. The example loads the Qwen3-0.6B GGUF model through llama.cpp, with Python dependencies including Hugging Face Hub, NumPy, and llama-cpp-python. It gives the model an email-security prompt with three choices: legitimate, spam, and phishing. The code evaluates the prompt, reads the logits for the first token of each label, normalizes those values with log-sum-exp, and prints logits, log probabilities, and probabilities. For the supplied payroll-themed email, the example reports probabilities of 0.031 for legitimate, 0.084 for spam, and 0.885 for phishing. The authors emphasize that their demonstration does not call an API, create synthetic data, or train a model with Reinforcement Learning for Calibrated Decisions (RLCD). They characterize the resulting system as a fast, local, probabilistic classifier that keeps the input on the local machine. The post is explicitly presented as a parody and points readers to OpenJev, openjev-sglang, and an OpenJev implementation on DiffusionGemma for more complete open implementations. It also states that NobodyWho’s work is open source.