A blog post on nobodywho.ai this week shows how to build a "System One decision model," the format behind TypeSafe AI's Jev, in 25 lines of Python with a small open-weight model.
Decision models take a prompt with a fixed set of possible answers, such as yes or no or a set of categories, and return a probability for each option instead of generating free text. The post's example loads a small GGUF-format model, Qwen3-0.6B, classifies a sample email as legitimate, spam or phishing by reading the model's raw output probabilities, and assigns it an 88.5% chance of being phishing.
The post is upfront that this is a simplified version. Commercial Jev implementations use what TypeSafe AI calls Reinforcement Learning for Calibrated Decisions, plus synthetic training data, to make sure the probabilities are actually calibrated. The DIY version makes no such promise; it extracts a probability directly and stops there.
Decision models have spun up a small cluster of their own GitHub projects this month, with typed classifiers, benchmarks and inference runtimes built around the format. A blog post showing the core trick fits in 25 lines is a sign of how mechanically simple the underlying idea is, even though getting the probabilities calibrated well enough to trust in production is the harder part a vendor like TypeSafe is charging for.