Typed Decision Models: Jev and Laya in Agentic AI

Same task shape, different products. Both expose Noul, Choice and Score style typed questions, but their internals, training evidence and deployment choices differ and the implementations are not interchangeable. Typed decision model is a description of the interface, not the name of one architecture.

Not ground truth. TypeSafe's workflow evals measure agreement with other models rather than verified labels, and the post labels its own illustrative numbers as examples rather than measured responses. A bounded answer set is a type guarantee, not a correctness guarantee: if fraud is missing from your list, the model can pick billing confidently.

The post reports 67.8% mean agreement at roughly $0.0004 per case against 67.9% at roughly $0.0304, so near-identical agreement across a seventy-fold price gap. The captured text does not restate which figure belongs to which model, so read the comparison section before drawing a conclusion from it.

TypeSafe documents up to 255 Choice options, two to ten Score levels, a 64k combined request limit, and 32k for state plus the longest question. The design rule is one semantic judgment per question: if question B needs A's answer, put it in a later request or compute the dependency in code.

A Noul returns a probability for yes, and 0.5 means yes and no are equally likely, not a medium intensity. To represent intensity, use a Score over explicit ordered levels: a probability-weighted position, so 0.25, 0.50, and 0.25 over three levels returns 1.0.

The article says TypeSafe calls Jev a System One Model and describes a new architecture, a parallel sampler, and post-training called Reinforcement Learning for Calibrated Decisions. It has not published the neural topology, parameter count, training recipe, or loss, so the article describes documented behavior, not hidden layers.

A Choice assumes the valid outcomes form a fixed set, and the model cannot name an option outside the list you supplied. That is a type guarantee, not a correctness guarantee: if the right answer is missing, it may confidently choose the least-wrong declared option, so include other or none_of_the_above.