decision model

What does a decision model return?

The answer is a structure, not a sentence. Jev exposes three shapes of question, and each returns probabilities that sum to one plus a confidence derived from the shape of that distribution.

• —Choice: one option among those declared by whoever asks, up to a maximum of 255.
• —Score: a rating on an ordered scale of between 2 and 10 levels, described in natural language.
• —Noul: a binary question whose answer is a single number between 0 and 1.

Does a decision model crawl my website?

No, and saying so matters because the confusion is already spreading. Jev's documentation defines its input as the content supplied by the caller — «State is the content you ask a System One model to evaluate. You pass it in the state field of an API request» — accepts text and nothing else, and describes no retrieval or browsing of its own.

A decision model is not a citation surface. Optimising a page so that Jev will cite it is a category error: the model writes no answers, links to no sources, and does not take the place of an AI search engine.

The only route by which your text reaches such a model is somebody bringing it first. The pattern repeats across its own examples and third-party services: a crawler fetches the page, turns it into clean text, and that text is pasted into the state field before any question is asked.

Nor is there a user agent to allow or block. A decision model never appears in your server logs, so no robots.txt rule affects it and none ever could.

Are they called «System One» or «decision models»?

«System One model» is the vendor's term and «decision model» is the one independent technical voices use. Simon Willison, following Maggie Appleton, wrote that the second is the better name, and this glossary records the distinction because a vendor's term and a category's term are not the same thing.

The name «System One» comes from Kahneman, who separated fast intuitive thinking from slow deliberative thinking. The image is good and needs explaining; «decision model» explains itself, which is what suits a term meant to outlast any one product.

What changes on my website?

Nothing in how you are crawled, and something in how you are judged. A decision model sits downstream of the crawler, inside pipelines that already exist, and there your page stops being read as prose and starts being evaluated with a typed question.

Price is why that has stopped being theoretical. With input at $0.042 per million tokens and output free, an audit published by Boring Tools Kit issued 4,816 judgments across 1,204 pages in under 3 minutes for less than $1.

What does not change is the work. A page that makes sense without its navigation, declares its entity and survives conversion to clean text serves the reader and the scorer alike, because in both cases the first step is taken by the same crawler.

Cheap should not be mistaken for unlimited either. Jev accepts 64,000 tokens per request and 32,000 for the state, and its own documentation warns that «accuracy falls as the state grows with content unrelated to the decision»: dumping a whole website into it makes the result worse, not better.

What is the risk in having my content judged?

Jev's documentation admits it in writing: «State is data, and jev-1.13 does not treat it as hostile by default». The same page grants that text written to steer the model — an injected instruction, a deliberately misleading framing, or a passage arguing for its own classification — can move the answer.

Primafonte documents that risk and does not teach anyone to exploit it. An index that measured how to push a verdict would stop measuring whether a website is well built and start measuring whether it is well manipulated, and only one of those can be defended to the person reading it.