How Jev Scans a Website in Seconds

TypeSafe describes System One models as “a class of AI models built to make fast, structured decisions that software can use directly,” and Jev as its “flagship model and the first System One model.” Every answer arrives as one of three shapes.

• Choice is “a System One question type for selecting one option from a defined set,” and the answer “includes the selected option, a probability for each option, and confidence.” A Choice question “accepts up to 255 options, and adding options costs a few tokens each,” so a full service catalog usually fits.
• Score is “a System One question type for rating content against ordered, descriptive levels.” It takes “at least two levels; the API accepts up to 10,” so you define what each level means in words.
• Noul answers a yes or no question with a probability. The documentation puts it plainly: “A Noul answer is a single number representing the probability that the answer is yes where 0 means no and 1 means yes.” A Noul carries no separate confidence value, because the number is doing both jobs.

Requests go to a single endpoint, POST https://api.typesafe.ai/v1/systemone, carrying the text as state and the questions alongside it. TypeSafe's quick start notes you can “Mix Noul, Choice, and Score in one call and see all results at once,” which is why a page can be judged on several dimensions in a single request. If you want the general introduction first, what Jev is covers the model without the site scanning angle.

A website audit looks like one big job and behaves like thousands of small identical ones. Each page raises the same short list of questions: which service does this describe, does it name the place it serves, how specific is the contact information, does the heading match what the body actually says. Google decides indexing the same way, page by page, and says so directly in its guide to how Search works: “Indexing isn't guaranteed; not every page that Google processes will be indexed.” The unit of the problem is the page, so the unit of the tool should be the page too.

A model shaped for one small classification, answered thousands of times, matches that unit exactly. Because every answer comes back as a label or a probability, the output is something your code can count and sort the moment it lands. Say three hundred pages go through a scan. The table that comes back might tell you that forty pages never name a city and that eleven describe a service you stopped selling two years ago. That's the practical difference from asking a general model to write you an essay about your website, and it's the reason this pairs well with the findings list in what SEO and AEO audit scores measure. The scoping side of that work lives on the Indianapolis SEO and AEO page.

The limits push in the same direction. TypeSafe warns that “Accuracy falls as the state grows with content unrelated to the decision,” which is a plain argument for sending one page per request and nothing else. Feeding the whole site into one call would make every answer worse.

Here's the arithmetic behind the speed claim, with every assumption labeled. Say your site has 180 pages. Say each page carries about 900 words of visible text, and say a word averages about 1.3 tokens, which gives roughly 1,200 tokens of state per page. Say you ask four questions of each page in a single request, adding about 300 tokens of question text. That's 216,000 tokens for the pages plus 54,000 for the questions, or 270,000 tokens in total. At the documented ceiling of 250,000 tokens per second, the model side of that scan is about 1.1 seconds. At $42 per billion input tokens, the token cost is 270,000 divided by a billion, multiplied by 42, which comes to a little over one cent.

Two honest caveats sit on top of that. Fetching 180 pages over the network takes longer than the model thinking about them, so wall clock time for a real scan is dominated by crawling. And TypeSafe notes on its models page that “Rate limits are adjusting dynamically,” so treat the published throughput as the current ceiling.

TypeSafe publishes a page on where the model gets jagged, and a scan design that ignores it produces confident nonsense. Four limits matter here.

The first is arithmetic. “Jev is not a calculator,” the documentation says, adding “We strongly recommend implementing any mathematical logic in code,” and noting that counting is unreliable. So a scan asks the model whether a page names a price, and asks your code how many words the page runs to. The second is dates. The model reads a date as a string of characters, and TypeSafe warns that asking the model which of two dates comes first, how far apart two dates sit, or whether one falls inside a window is unreliable. Checking whether a copyright line is current belongs in code, every time.

The third is writing. “jev-1.13 is not trained to generate text,” and the documentation says pushing it to do so works poorly and runs very slowly. A scan gives you a sorted list of findings and stops there; turning a finding into a rewritten paragraph is a separate job, which is what pairing Jev with Claude is for. The fourth is context hygiene, the accuracy point above, plus a note worth keeping in mind when you write your questions: “It can be quite literal in its understanding.” Vague wording produces answers to a question you didn't mean to ask.

Good questions are narrow, answerable from the page alone, and phrased so a person would agree with the answer. A workable set for a local service site looks like this.

• Noul: does this page name a city or area the business serves?
• Noul: does this page show a way to contact the business?
• Choice: which service from our list does this page describe?
• Score: how specific is the local detail on this page, across four levels you define from generic to street level?
• Score: how closely does the main heading match what the body of the page delivers?

Response codes, image weight, heading order and load time stay in code, because those are measurements a crawler takes. The model handles the judgment calls that used to need a person reading every page, and it holds the same standard on page 180 as on page one. If you'd like to see what this sort of output looks like before building anything, our Jev Website Scanner takes a URL and returns a score for each page with the specific findings listed. Jev answers 44 set questions about each page, and code counts 26 technical facts exactly, like schema and tap to call links. A yes or no finding needs Jev to be at least 80 percent sure of the problem and a scored one needs at least 50 percent confidence. Weaker answers go to a separate “Worth a look” list. The first scan of a site is free for up to 100 pages, and you enter an email so we can send you the report link.

Every Choice and Score answer carries a confidence value. TypeSafe explains that the probability distribution is the real signal, and that its shape is “concentrated on one outcome means a confident answer, spread out means an uncertain one,” with confidence collapsing that shape “into a single number from 0 to 1.” The advice that matters for a scan is about thresholds: “Different actions within the same system should be gated at different levels depending on the consequences of getting it wrong.”

In practice that means a low confidence answer earns a human look, and a high confidence answer on a low risk question can go straight onto the fix list. Flagging a page as missing a city name costs little when the model is wrong. Deleting a page because the model classified it as duplicate costs a lot, so that decision wants a higher bar and a person signing off.

A scan works as a sorting step. It turns a pile of pages into a short list of pages worth your attention, at a cost low enough that rerunning it after every batch of fixes is reasonable. The value shows up when you use the same questions each time, so the numbers move for a reason you can explain. Start with the free first scan in our Jev Website Scanner to see the output shape, then decide which questions your own site needs asked of everything. Scanning the same site again is the $99 audit, and it includes one rescan after your fixes. When those answers start pointing at template level problems, the Indianapolis SEO and AEO page covers how that work gets scoped and sequenced.