Duarte O. Carmo at NobodyWho shows Jev's core shape in 25 lines of Python by reading Qwen3-0.6B logits and softmaxing over the choice tokens

Duarte O. Carmo at NobodyWho published a post on 22 September 2026 that shows the core behaviour of TypeSafe's Jev model, the "System One decision model" the AI trade has been talking about all week, running in 25 lines of Python on a laptop. The post reached 555 Hacker News points in a day. The trick is a small one, and the point of the post is that most teams talking about Jev do not know how small it is.

The 25 lines, in one sentence

Load Qwen3-0.6B as a GGUF file with llama-cpp-python. Send a prompt that lists the choices. Ask the model for the logits at the next token position, without generating anything. Pick out the logits belonging to the specific tokens that spell each choice. Softmax those logits into a probability distribution. That distribution is the answer.

The worked example in the post classifies an email as Phishing with probability 0.885, Spam 0.084, and Legitimate 0.031. No text was generated. The whole call is one forward pass of a 600-million-parameter model.

What the post skipped, and what that costs

Carmo is direct about what a paid Jev API adds on top. TypeSafe trains its model with reinforcement learning for calibrated decisions, so the probabilities it returns are meant to match measured frequencies on real data. The 25-line version has whatever calibration Qwen3-0.6B happened to learn during general pretraining, which is not the same thing. Carmo links to open implementations he considers more complete: OpenJev, openjev-sglang, and OpenJev on DiffusionGemma. He is not claiming the 25 lines replace a production Jev grader, and he closes with a "parody blog post" note that says so.

That caveat is doing real work. On classification tasks where a decision only needs to be roughly right, the difference between a calibrated 0.72 and an uncalibrated 0.72 is small. On a task where a downstream branch fires at 0.8 and does not fire at 0.7, the calibration is the whole point.

Why a team would run this even so

The audience the post found is a team paying for classification API calls that add up. The 25 lines are an executable baseline: same input, same choices, a probability distribution out. A team on a Jev bill this week can run its own inputs through both and compare the two distributions. On the tasks where the paid version wins by less than the API cost, the local model is the answer. On the tasks where it wins by more, the calibration is worth what it costs, and now the team has that broken out per prompt template.

For any team already routing between a small model and a paid grader, the useful assignment for the week is to add the 25 lines as a shadow path, log both distributions against the ground truth for a few days, and let the numbers say which prompts need the paid model and which ones are paying for a rounding error.

Source

Duarte O. Carmo, NobodyWho, Jev in 25 lines of Python, 22 September 2026.