tabstitchSource ↗

KEYBOARD & FORM BUGS · NVIDIA NEMOTRON ON NEBIUS · REAL BROWSER VERDICT

Repair the task.
Keep the outcome.

Tabstitch fixes keyboard accessibility and form bugs in web components, the kind that stop someone using only a keyboard from finishing a task. NVIDIA Nemotron on Nebius Token Factory writes the patch. A real browser then replays the user's task, and the patch is accepted only if the task passes, the submitted data stays the same and no guarded behaviour breaks. It was tested on real bugs that the maintainers of Bootstrap, GOV.UK Frontend and other libraries fixed by hand.

Fixed bugs, blind20/26bugs the maintainers of 12 libraries fixed by hand, repaired by Nemotron without seeing their fix Still open upstream7keyboard and form bugs still unfixed upstream, reproduced on each library's current code In a pull requestPR #1The GitHub Action caught a focus bug, rejected Super's first patch in the browser and posted Ultra's verified fixFound and reported upstream#43202A Bootstrap 6 datepicker keyboard trap no one had reported, found by Tabstitch and filed with a reduced test case. Confirmed by a Bootstrap maintainer the same day

HOW A PATCH IS ACCEPTED

The model proposes.
The browser decides.

  1. 01

    Freeze the contract

    The same keys, the data the form must submit, an independent pointer path and regression guards that already pass.

  2. 02

    Let Nemotron propose

    Nemotron 3 Super edits the library source in a disposable copy. If the browser rejects the patch, Nemotron 3 Ultra gets the exact failure and tries again.

  3. 03

    Make the browser decide

    Replay every path in a fresh browser. Reject a patch that passes the task but changes the submitted data or breaks a guarded behaviour.

ONE TASK. THE SAME KEYS. TWO OUTCOMES.

Follow the focus.

RECORDED BROWSER EXECUTION

Step through the actual browser captures.

Enter
Open the dialog

Loading the measured focus transition…

ORIGINALLoading
Original keyboard journey
FOCUS—
NEMOTRON PATCHLoading
Repaired keyboard journey
FOCUS—

Same task. Independent verdict.The model writes a proposal. Browser execution decides whether it works.

MODEL-AUTHORED SOURCE. EXECUTION-OWNED VERDICT.

A patch you can question.

Loading verdict

What Nemotron found

Loading the recorded diagnosis…

What it changed

NVIDIA Nemotron
Model time
—
Total tokens
—
This successful call
1 · not cached
ACTUAL SOURCE DIFFDownload patch ↓
Loading source diff…

What had to stay true

The repair cannot rewrite these checks.

What this pass does and does not establish

Each recording verifies only its frozen journey and recorded conditions. It does not establish a repair success rate, full WCAG conformance or real screen-reader support.

See the failed attempts, too

The historical case took three calls: one output was truncated, one invented an unapproved file and was rejected, and one passed. The synthetic checkout took one successful call. These are development attempts on two known cases, not an independent benchmark.

    Inspect all four call records ↗

    INSPECT THE RECORDING

    The evidence travels
    with the result.

    No API key needed. These files come from the recorded run. The complete proof includes browser captures, focus states, source changes and the frozen task.

    Download complete proof ↓

    Run a fresh browser journey.

    Select your static application, capture and freeze the task, then let Nemotron repair an isolated copy. Verify a human edit with added, deleted or modified files in the same flow.