YOUR SOURCE. YOUR TASK. A CHECKABLE VERDICT.
Repair the task.
Keep the outcome.
Capture the keys and choose the result that must stay true. Nemotron repairs an isolated copy; a fresh browser decides whether it works.
Pinned a11y-dialog defect. Starts in human-patch verification mode; no inference until you explicitly choose Nemotron.
REAL COMPONENT ISSUE / FORM DATA
The field looks empty. The form still sends the old country.
Clearing or mistyping a saved country in GOV.UK autocomplete can leave its hidden select unchanged (reported issue #777). This delivery form uses the pinned, unmodified component. Six plausible Nemotron patches were rejected by frozen browser gates, including one that passed an earlier contract but erased saved countries when the options were pretty-printed. The final Nemotron Ultra patch passes the keyboard task, an independent pointer path and three regression guards; three of four withheld checks pass, and the fourth is a recorded failure explained in the history.
Author-created consumer around unchanged third-party code. Tasks were strengthened four times after independent review; every step is recorded in the evidence history on each result.
REAL COMPONENT ISSUE / UNAVAILABLE DATES
Closed dates must stay unavailable.
Pikaday can select a disabled day with the keyboard while its pointer calendar blocks that day. This weekday-only booking form reconstructs reported issue #955 around the pinned, unchanged component.
Author-created consumer and weekday-navigation acceptance conditions. Loading this source makes no model call.
PUBLISHED EVIDENCE · READ ONLY
Every recorded run, including the rejected ones.
This page has no backend. It shows recorded browser evidence, raw model requests and responses, diffs and proof archives exactly as captured. Importing your own source, capturing tasks and new Nemotron inference need the local workbench described in the README.
RECORDED BROWSER EVIDENCE
Running your frozen task…
Selected attempt
RUNNINGOriginal verifier record (unchanged)
Candidate source diff · checked by git apply
Attempt history · select to inspect
Checkpoint playback displays this run’s captures. Fresh browser verification reuses the selected patch. New inference makes a new model request and rechecks frozen tasks. A pass covers only these tasks and recorded conditions.