NVIDIA Nemotron · Nebius Token Factory

The AI says “fixed.”
We check the user can finish.

AccessFinish is a repair agent that doesn't trust a green page. It drives the real journey with a keyboard, reads what was actually saved, reopens it in a fresh browser, and keeps fixing until the outcome is right.

Replay the agent's runs Watch the 2-minute film

Film · 2 minutesOn YouTube · narrated, with captions · Nemotron fixes a real keyboard bug and AccessFinish proves it

Recorded run · All runs

Goal

A loop that ends at the saved outcome, not the success page.

ObserveA real browser drives the journey with the keyboard only and records where focus goes and where it gets stuck.
Diagnose & repairNemotron 3 Super reads the evidence and the source, finds the cause and writes exact code edits. No root cause is supplied; the acceptance checks name the date field the verifier types into.
Verify for realEach patch that applies is built in a fresh, offline sandbox. The booking is made, read back from the database with a read-only account, and reopened in a fresh browser.
Feed backIf anything is off — a build error, a stuck key, a wrong saved value, a broken mouse path — the evidence goes back to Nemotron and it tries again.

Why not just check the confirmation page?

Because it lies. In our runs, an AI fix reached a booked confirmation — times while the booking was saved, or reopened, with the wrong date. In the phone case the page says booked while the database holds 0000000000, and a patch that hardcodes the test number would fail when a second customer types 0000000073. AccessFinish checks all of it.

Why keyboard journeys?

Millions of people navigate with a keyboard or assistive technology. A form that "loads" but can't be completed locks them out, and since June 2025 the European Accessibility Act expects in-scope services to work for them.

0 / —false "booked" fixes accepted: every one was caught by the database read-back or the reopen

—keyboard runs fixed by Nemotron, no root cause supplied

—runs verified on the seeded phone control

—estimated Nemotron cost (list price) for all — live runs

What “verified” proves, and what it doesn't

Proves

A keyboard-only browser session reaches the date, books the slot and sees the confirmation.
The database row, read with a read-only account, holds the date and customer that were typed.
Reopening the booking in a fresh browser shows the same date.
A mouse booking with different data, a cleared date and a reserved slot still behave, and nothing else in the database changed.

Doesn't prove

That the calendar grid is usable with arrow keys: the verifier types the date into the field it names in its acceptance checks.
Anything a screen-reader user study would show.
That it generalises: one app, two journeys, two editable files.
Which days the calendar disables after a typed month change; the oracle does not check that.

Synthetic data in a pinned copy of the open-source Easy!Appointments app. Two journeys today; the phone fault was seeded on purpose. All runs were made live on GitHub Actions with Nemotron 3 Super on Nebius Token Factory, and every one, including failures, is saved with its raw evidence — see Evidence.

The AI says “fixed.”We check the user can finish.

Agent run