Live Demonstration: A Decision Is Not a Sentence

Most "AI on a website" still means a chatbot: you type a question, it writes a paragraph back. That's one shape a model can take — but it isn't the shape software needs when it has to make a call itself: approve or refuse, escalate or close, route left or right. A newer category, sometimes called "System 1" models (after the fast, intuitive mode of thinking in Daniel Kahneman's dual-process theory), skips the sentence entirely and answers a narrowly typed question instead.

Below is a small, fixed example of that idea: given a short customer review, decide whether it is positive, neutral, or negative — no other answer is allowed. Type your own text, or pick one of the three ready-made samples, press Evaluate, and see what actually happens, including how long it took and what it cost, measured on our server rather than guessed at.

The demo calls a general-purpose language model (OpenAI) with the answer format locked down to exactly those three words. Next to it, a second, independent call goes to a purpose-built typed decision model called "Jev", from a company called TypeSafe AI, reached through a third-party marketplace (OpenRouter) — the exact same fixed task, measured the exact same way: real server-side timing, a real per-call cost from the provider's own usage figures, median of three calls each. If that connection is ever unavailable, the column falls back automatically to a clearly labeled vendor estimate instead of failing the whole comparison.

Everything under the demo explains exactly what the numbers do and do not mean. A badly built measurement is worse than no measurement.

Figure 1. The same task, two ways — adapted from a diagram by the Department of Information Engineering, PEF CZU.

What this means

A typed decision model doesn't predict the next word — it evaluates every question you give it in a single pass and hands back a probability alongside its answer. The interesting part isn't that this is "better AI." It's a different shape of task: a language model normally writes a sentence; a typed decision model picks from a fixed, pre-declared set of options and attaches a number to its pick.

"It cannot hallucinate" is a claim worth being precise about. It's true only of the shape of the answer, not its content: a typed model can't invent a fourth option that wasn't in the schema — but it can just as confidently pick the wrong one of the three it was given. Early evaluations of Jev, both from the vendor and from outside testers, already show this pattern: solid on some narrow tasks, noticeably weaker on others, and how often a "confident but wrong" answer happens depends heavily on exactly how the question is phrased.

Figure 2. What "calibrated" confidence means, and why it matters — adapted from a diagram by the Department of Information Engineering, PEF CZU.

That's exactly where the probability earns its keep. A model whose confidence is well calibrated — where "90% sure" really does turn out right about nine times in ten — lets a system set a threshold: below it, don't decide automatically, hand the case to a person instead. That threshold is auditable in a way a fluent-sounding paragraph isn't, which is also why it matters for the human-oversight requirement the EU AI Act places on higher-risk automated decisions. Whether Jev's own probabilities are that well calibrated is, honestly, still an open question: the company hasn't published the methodology behind its training approach, and outside evaluations so far are mixed and depend heavily on the task.

Run a few reviews through the demo above and watch Jev's own probability move with them — that number is real, not illustrative. One caveat worth stating plainly: neither TypeSafe nor OpenRouter publish a confirmed EU processing route for this model, so treat it the same way as the language model call for data-handling purposes, and stick to a made-up review.