Rather than describe it, here's one evening.

First, we gave it what we already knew.

Fifteen documents, straight from the folder: user stories in Markdown, Gherkin and a Jira-style CSV export; business rules as a PDF; a glossary; an API spec; a data dictionary. Each became facts the scan could later check. Fourteen were read into facts before anything touched the application.

Then we pointed it at the app.

No login, no sitemap, no hints. Discovery narrates what it maps as it maps it: the index page, classified as a navigation hub with 121 interactive elements; then page after page, until 27 pages and 17 flows were in the knowledge base.

It stopped, and showed its homework.

Before submitting a single form it paused for review: 47 fields, 47 valid values and 113 invalid ones it had prepared, each with a confidence score and a written reason. Three values it wasn't sure about, and said so. You can edit, remove or add any of them. Nothing runs until you say so.

It submitted, and wrote down what the app said.

Seventeen flows, sixty-three observations. Each record holds what was submitted, what the app answered, verbatim, and how sure Correctest is. The judgement stays neutral: it records; you decide.

It found eleven things our app accepted that it shouldn't have.

A postcode field took NOTAPOSTCODE!. A terms checkbox took no. An email field took nothing at all. None of these become tests that would quietly pass; each is held out as a finding until you decide what it means.

It also found our own story disagreeing with our own app. Three times.

The registration story from 5:18 PM says a mismatched password confirmation, an under-age visitor and a reserved username are each rejected. The app accepted all three registrations. Correctest put document and app side by side and asked which one is wrong.

It handed us a repository.

58 tests across 12 flows, as a plain Playwright project: page objects, specs, fixtures, a CI workflow. Every test names the observation it is grounded in. Five flows were skipped, and the coverage report says why, including a password reset it refused to drive for real.

// tests/flows/autocomplete-preferences-form.spec.ts — first test of two, as generated
import { test, expect } from "@playwright/test";
import { resolveCredentials, resolveFixedTestData, recordCreatedTestData } from "../shared/fixtures";
import { TestbedAutocompletePage } from "../pages/TestbedAutocompletePage";
import data from "../pages/TestbedAutocompletePage.data.json" with { type: "json" };

test.describe("autocomplete-preferences-form", () => {
  /**
   * Asserts the happy path: valid city + postcode saves location
   * preferences and navigates to the submission confirmation page with
   * a "Location preferences saved." banner.
   * Grounded in observation: 52f4f5ce-7952-432a-8efa-1deea36387cc
   * (confidence 0.95).
   */
  test("autocomplete preferences success — valid city and postcode save preferences", async ({ page }) => {
    const autocomplete = new TestbedAutocompletePage(page);
    const preferredCity = resolveCredentials(data.valid.preferredCity);
    const postcode = resolveCredentials(data.valid.postcode);
    await autocomplete.goto();
    await autocomplete.fillPreferredCity(preferredCity);
    await autocomplete.fillPostcode(postcode);
    await autocomplete.clickSubmit();
    await expect(page).toHaveURL(/\/testbed\/submit\/autocomplete/);
    await expect(page.locator('div.banner-success')).toBeVisible();
    …
  });
});

This scan explored as a signed-out visitor, so the project it produced carries no sign-in step at all: a suite authenticates only when the scan that produced it authenticated, and it says so rather than filling a login form with sample values.

Then we asked it to follow our stories on their own. And it caught us out.

A second scan, signed in as a Customer, walking the acceptance criteria in the documents from 5:18 PM. One story file was out of date: it fills a "full name" field the registration page no longer has. Correctest didn't force the walk through, and it didn't mark the story failed. It stopped, said what it couldn't find, and asked.