Jev in Respira AER: a second model that checks every change your AI makes

Respira AER 0.3 adds a second model to every run. Your own model (Claude, GPT, Gemini) still does the work. After each change it makes to a WordPress site, Jev, a small decision model from TypeSafe, checks that the change landed, that it is what you asked for, and that nothing else moved. It is included free on every plan.

What Jev is

The models most of us use are language models: given a prompt, they write the next words. Jev does not write anything. You give it some state (a request, a record, the result of a tool call) and a set of typed questions, and it returns typed answers: which option out of a list, yes or no, or a position on a scale, each with a probability. TypeSafe calls this a System One model, after the fast, automatic kind of thinking. It is built for the small judgments software makes all day, and it answers in about half a second.

TypeSafe was started by Diogo Almeida, previously at OpenAI, with Erik Gafni and Sasha Sheng, and Jev went into early access this week. It is priced per input token (output is free), which makes a single check cost a small fraction of a cent. The details are in TypeSafe's announcement and their documentation.

What I tested first

Before putting it anywhere near a customer's site, I gave it two jobs on my own data.

• Reading every bug report Respira has received. Seven questions per report: which part of the product, which page builder, how badly it blocked the person, and whether a tool reported success while nothing actually changed. The whole backlog took about a minute and cost a few cents. The pattern it surfaced was the one that mattered most, and one I had not seen clearly by reading reports one at a time: a tool saying "done" when the page had not changed. That is exactly what the AER check now looks for.
• Turning a business description into a page, with no language model at all. A plumber's brief went in line by line, and Jev decided what each line was: the promise, a service, a testimonial, a price, a question with its answer, the address. 26 of 26 lines came back right, in about a second, and plain code built a native Divi page from its choices. That one is an experiment, not a feature, but it shows how far typed decisions go.

The one lesson worth knowing

Quote the thing you want judged inside the question. When I pointed at a line by its position in the state (the fourth sentence, the ninth sentence), Jev scored a phone number and an email as street addresses, both around 0.9. With the line itself quoted in the question, the real address scored 0.96 and every other line under 0.1. The other answers in that test became more confident too, the lowest going from 0.45 to 0.82. If you are building with Jev, it is the cheapest improvement you will make.

What it does in Respira AER

After each write that changes content on a site (an element update, a page update, a patch, a conversion), AER sends Jev three questions about the tool's own result, quoting the request and the write:

• Does the result show the change was actually applied?
• Does the write do what you asked, rather than something different?
• Does the result show changes beyond what you asked for?

When an answer comes back no, AER adds a short note to the result your model reads: this may not have landed, read the page back before you say it is done. Nothing is blocked and no approval card appears, so Auto mode stays automatic. The model is told, and the model decides.

In the checks I ran before release, a result that said "0 elements updated" came back with a 0.02 chance of having landed and was flagged. A result that clearly showed the change came back at 0.96 and passed. Each check took under a second.

Every receipt gets a Jev line: how many checks ran, what they cost and how many changes were flagged. That cost is never added to your model bill.

Your data, and the off switch

For each check, your request and the write's result go to TypeSafe. TypeSafe says it does not train on them. If you would rather they did not leave Respira, switch Jev off in AER settings and AER runs exactly as it did before.

What it is not

Jev's probabilities are calibrated across many answers, not guaranteed for any single one, so it is a check, not a verdict. That is why it tells your model to look again rather than overruling it. It is also a very new model from a very new company: if it is slow or unavailable, AER waits at most a second and a half and then carries on without it. Its best language is English. Jev can also pick which tools your model sees on a turn, but that stays switched off until it has been measured against the router AER uses today.

Update, 19 September: Jev on respira.press

The day after, Jev went to work on respira.press itself, on two jobs a program cannot do alone.

• Internal links. Code crawled 385 pages and found places where one page already uses the words another page is about. For each place, Jev judged whether a link there would honestly help the person reading. 1,006 judgments took 29 seconds and cost under five cents. 138 links passed, and the first 36 are on the site now. Jev never writes the link text: every link is made of words already on the page.
• Search intent. Search Console says which page ranks for which search. It cannot say whether the page answers it. Jev judged every search with real traffic against the page that ranks for it: 172 pairs in 5 seconds for under a cent. It found pages whose titles hide what they answer, searches no page answers yet, and pages competing with each other for the same search.

Next, the same kind of pass for your own sites, in the Do tab of Respira AER: a search intent audit, an internal link map, comment triage, a check that each headline delivers what it promises, a stale content finder, Murmur feedback triage and a WooCommerce catalog check. Each one judges up to hundreds of items in seconds, proposes changes and applies none, and the receipt shows what the judging cost.

Update, 19 September: the passes shipped in Respira AER 0.4

Six of those passes are in Respira AER now, under the Judged by Jev tab on the home screen and at the top of the Do tab, grouped by job:

• SEO and content. Search intent audit (/run intent): does each page answer what its title was searched for. Headline promise check (/run promises): does each page deliver what its headline says. Internal link map (/run links): which pages should link to each other, with the anchor words taken from the page itself.
• Comments and reviews. Comment triage (/run comments): spam, questions, praise and the ones to reply to today. Murmur triage (/run murmur-triage): client review comments sorted by kind and urgency, blockers first.
• Store. Catalog check (/run catalog): WooCommerce product descriptions judged for their category and for what a buyer asks before buying.

A pass gathers a bounded set with the site's own read tools, Jev answers one question per item, and your model reports what to fix, ranked. On my own sites, 20 pages were judged for search intent in 0.7 seconds for $0.00015, and 20 comments were triaged in 0.8 seconds for $0.00016. The stale content finder (/run stale) uses Jev too when it is on, and your model can call the same judge itself: respira_judge hands Jev up to 200 items and one question. Like the write check, the passes are included on every plan at no extra cost.

How to try it

Open Respira AER and run anything that changes a page. The Jev line appears on the receipt. For a whole site, open the Judged by Jev tab and pick a pass. What changed in each AER version is on the releases page, and everything about AER is at respira.press/aer.

Mihai