Early access, not released

A test driver that cannot invent a selector

Reticle’s harness drives your running app and records what it drove. This one runs on @typesafeai’s Jev, a model that cannot generate a tool call at all. It can only pick one we built from the elements actually on your page.

$0.006: per full drive
342×: cheaper than a frontier driver
~340ms: per decision
0: selectors it can invent

How it drives

We write the menu. The model only points at it.

A generating driver writes its own tool calls, which is how it ends up clicking an element that was never on the page. Ours enumerates every candidate first, in code, from the snapshot.

act:e5      click the button "Create deployment"
fill:f1     fill the textbox "Project name"
act:f3      click the button "Confirm"   ← chosen, p=1.0
finish      stop driving, the app has been covered
look_again  look again without acting, this reading looks wrong

Jev returns one key. The worst it can do is point at the wrong row, which costs a step. It cannot point at a row that is not there.

What you get

We let people in a few at a time, on purpose

The first runs on somebody else’s app teach us what a benchmark on our own never will. That is the trade: you get it early, we find out where it breaks.

A driver that cannot make things up

Every candidate action is built in our code from your page's accessibility snapshot. Jev returns one key from that set. It cannot name an element that is not on the page, so it cannot hallucinate a selector.

Your model bill, near zero

A full drive across a real dashboard cost six tenths of a cent. We hold the credentials and proxy the call, so you never sign up to TypeSafe or hold their key.

A regression suite you did not write

Every drive is recorded as flows. From the second run on they replay deterministically with no model in the loop at all, which is the part that stays free forever.

A direct line while it is small

You are talking to the people writing it. What you send back decides what gets fixed first, and the first batch shapes the thing.

Get it before it is finished

Free, and one email when your turn comes up.

Jev is built by @typesafeai. The harness is ours.

Before you type your email

The harness is the loop that drives your app when no coding agent is driving it: it picks what to touch next, states the consequence it expects, and records the journey as a replayable flow. This version runs that loop on Jev instead of a frontier model.

A System One model from @typesafeai. It is not autoregressive and emits no prose. You give it a state and typed questions and it answers all of them in one pass, as a choice from a set you defined plus a probability. Their launch post explains it properly. We are not affiliated with them and they did not review this page.

No. That is why this is a list and not a download. The open-source engine is free and installable today; this driver is not merged yet. We would rather say that than ship a page you can disprove with npm i.

Never. The model picks what to try. A deterministic function in the engine decides what happened, from evidence, with no model in it. Jev cannot hallucinate a type, but it still returns a probability, and a probability is not proof.

Reticle is an SDK you embed inside your own app, so we cannot test a site we do not ship code into. On our run the cheap driver saved one replayable flow where a frontier driver saved twelve, and that run was n=1. The write-up has every caveat, including two bugs in our own driver.

We do not know yet, and it is the main thing we want to learn. Jev’s context window is 32K and our state was around 17,000 characters. A large enough DOM does not degrade the drive, it fails the call. Finding where that line sits is most of why there is a list.

The measured write-up, with the full table and every caveat, is in the post. The engine is free and open source today.