AI assertions

AI assertions check things that are about meaning rather than exact text: that a message explains an error, that the new item is shown, that a page looks finished. They work on a page or on a locator, and like the page matchers, they must be awaited.

await expect(page).toSatisfy("the new note is shown with its title");
await expect(page.getByRole("alert")).toSatisfy("the message says the card was declined");
await expect(page).not.toSatisfy("an error message is shown");

Prefer a fixed assertion such as toHaveText when you know the exact text. Reach for an AI assertion when the wording can change but the meaning shouldn't.

toSatisfy

toSatisfy(claim, options?) asks Jev for the probability that the claim is true. It passes at minProbability or above, 0.7 by default. With .not, it passes when the probability is at most 1 - minProbability, 0.3 by default.

When it fails, the message shows the probability and what was needed:

expect(page).toSatisfy(claim)

Claim: the new note is shown with its title
Probability: 0.42, needed at least 0.70
Page: http://localhost:3000/notes "Notes"
Model: jev-1.13.0, 2 calls

toSatisfyAll

toSatisfyAll(claims, options?) checks several claims about the same page in one request. Every claim must pass. With .not, none may hold.

await expect(page).toSatisfyAll([
  "the cart shows 3 items",
  "the total is shown in dollars",
  "a Checkout button is visible",
]);

The failure message lists every claim:

expect(page).toSatisfyAll(claims)

Each claim needs a probability at least 0.70:
  ✓ 0.96  the cart shows 3 items
  ✗ 0.31  the total is shown in dollars
  ✓ 0.92  a Checkout button is visible

Page: http://localhost:3000/cart "Cart"
Model: jev-1.13.0, 1 call

toScore

toScore(question, levels, options) asks Jev to rate the page on ordered levels, lowest first. It needs atLeast, atMost, or both.

await expect(page.getByRole("alert")).toScore(
  "How clearly does the message explain what went wrong?",
  ["Not at all", "Vaguely", "Clearly", "Very clearly"],
  { atLeast: 2 },
);

• Levels are numbered from 0, so atLeast: 2 means "Clearly" or better.
• There must be from 2 to 10 levels.
• The score is the probability-weighted level, so it can land between levels, such as 2.4. atLeast and atMost include their bounds.

How an AI assertion waits

Pages change after they load, so an AI assertion keeps checking until the answer is what it wants or time runs out:

1. It captures the page, or the element, and asks Jev the claim together with "is the page still loading?" in one request.
2. If the answer passes, the assertion passes.
3. If the page looks like it's still loading, it keeps checking until the timeout. If it doesn't, it gives the page about a second to change, then fails.
4. It asks Jev again only when the page state has changed since the last request, and makes at most 8 requests.

The timeout is expect.timeout from the config, 5 seconds by default, or the timeout you pass.

Options

Option	Matchers	Default	What it does
minProbability	toSatisfy, toSatisfyAll	ai.minProbability, 0.7	The probability a claim needs.
atLeast	toScore		The lowest passing score.
atMost	toScore		The highest passing score.
timeout	All three	expect.timeout, 5000	Milliseconds to keep checking while the page changes.

Writing good claims

• One checkable fact per claim. Split "the cart has 3 items and a total" into two claims.
• Group claims about the same page into one toSatisfyAll: one request instead of several, and the failure lists every claim.
• Scope to an element by calling the matcher on a locator, so Jev reads only that part of the page.
• Be concrete. "The new note titled Groceries is at the top of the list" is easier to check than "the page looks right".
• Watch for claims near the threshold. A claim that passes at 0.72 and fails at 0.68 is too vague. Make it more specific, or check it with a fixed locator and toHaveText.

Each check is a step in the report, and the AI decisions table lists every claim with its probability and whether it passed.