A model that only decides.

Jev is a decision-focused model from TypeSafe AI. Instead of generating open-ended text, it works over constrained decision spaces such as Boolean, Choice and Score outputs. This experiment asks where that type of intelligence maps naturally to classical AI problems, and where it does not.

How the exhibits use it

• Every decision is a single Choice question: a structured state, a short neutral question, and the legal actions as labels, each with a one-line description. A few features also ask Jev to judge a state: yes/no factor questions behind the Restaurant rationale, and a 0–4 closeness score in the Missionaries engine mode.
• Jev returns the chosen label and a probability for every label. The site shows those numbers as returned and never invents a distribution. If only one action is legal, no call is made.
• Illegal actions are never offered. Movement, validation, scoring, belief updates and search are ordinary, deterministic code.
• Jev has no memory between calls. Each decision sees only the state it is given.
• Question wording was written once and not tuned to make Jev succeed. The extra context levels in Missionaries and Tic-Tac-Toe were added after Jev failed with the state alone, and every level is shown separately. Where Jev does poorly, the site shows it.
• Jev is not perfectly deterministic: repeating the same call can move its probabilities by a few points, and near-50/50 choices can flip. Treat small differences as noise.
• Browsers never send Jev any text. They send game state; the server builds the question from the game's own code, so the site cannot be used to ask Jev anything else.

Inspecting it yourself

Every decision panel has an Inspect drawer showing exactly what Jev received: state, question, and each option with its description, as returned by the server. Add ?debug=true to any exhibit URL for raw responses, latency and a downloadable log of every call made in your browser.