Jev, a System One model, targets quick yes-or-no decisions that LLMs handle slowly

A System One model for small decisions

Amii argues that large language models such as Claude or ChatGPT are often over-used for minor, repeated choices. Its comparison: using an LLM for every small decision is like "hiring Einstein to change your light bulb." Typical examples are asking an agent whether a message is spam or whether an email needs a reply.

Jev is designed for these short yes-or-no style questions. Amii describes it as a System One model, meaning one built for fast, automatic, essentially intuitive decisions within a larger workflow, citing TypeSafe AI (2026). Its role is not to imitate human-like reasoning but to return quick decisions, each with a probability and confidence value.

How Jev compares with LLMs

Both receive the same context, but Jev outputs structured values while an LLM outputs generated text. Amii's comparison table lists the following:

• Response time: 70ms-500ms for Jev, 3-329 seconds for LLMs.
• Input tokens: $0.042 for Jev, $0.20-$10 for LLMs.
• Sampling: Jev samples in parallel and generates all outputs in a single query, while LLMs generate one token at a time.
• Training approach: Reinforcement Learning for Calibrated Decisions (RLCD) for Jev, Reinforcement Learning with Human Feedback (RLHF) for LLMs.

Jev supports three decision primitives. A Choice selects from known options, such as a team classification. A Score places the input on an ordered scale, such as low, medium or high. The third, listed in the article as Noul, evaluates a yes-or-no condition and returns the probability that it is true.

Calibrated probabilities and the hallucination claim

The outputs can feed directly into software that classifies information, prioritizes work or selects the next action. Because Jev returns probabilities, an application can act automatically when the evidence is strong and ask a person or another model for input when the answer is uncertain.

Amii explains that probabilities are meaningful only if calibrated: across many predictions assigned 80%, the predicted outcome should occur roughly 80% of the time. TypeSafe AI has trained Jev toward this goal, but Amii stresses that calibration does not guarantee any individual decision is correct.

In an email-triage demo, Jev gave a sample message a 78% probability of being a meeting, 14% request and 8% spam. It put 80% on handling it this week and a 70% probability on needing a reply.

One of TypeSafe's strongest claims, per Amii, is that Jev cannot hallucinate. The reasoning is that its outputs are defined in advance, so it cannot invent a fifth email category. Amii compares this to a multiple-choice question: a student cannot invent option E but can still pick the wrong answer. Jev could, for instance, classify a meeting request as spam while producing a valid output.

Where it could be used

Amii says Jev is most relevant when a task involves repeatedly evaluating information against clearly defined options or criteria. It names four areas:

• Customer support: classify requests and route them to the right team.
• Project management: categorize feedback, score issue urgency and flag messages needing follow-up.
• Education and training: classify learner questions by topic or flag responses for instructor review.
• Software development: decide whether an agent should search documents, query a database, call a tool, or pass a request to a more capable reasoning model.

For teams running automated workflows, the article's message is that routine classification and routing need not always go through a full LLM. The speed and price figures are those given in Amii's comparison table, and the article itself cautions that valid outputs are not necessarily correct ones.