What execution can show
Actual keyboard transitions, explicit task outcomes, accessibility-tree semantics, automated WCAG rule results and the frozen functional checks. A repair is accepted only after differential replay.
BEHAVIORAL ACCESSIBILITY / EXECUTABLE EVIDENCE
Don’t trust the fix.
Replay the journey.
A button can pass a scan and still block a person. Reproduce the task. Repair the source. Replay the exact same interaction.
THE MODEL PROPOSES. EXECUTION PROVES.SEE THE FAILURE. INSPECT THE DECISION.
ACTUAL RECORDED BROWSER EVIDENCE
A passing scan is not a completed task.
Loading recorded journey evidence?
Recorded evidence ? not live inference
Focus evidence
Focus evidence
What happened to the person's path?
Observed focus before the repair
Inspect the actual source change
SCOPE IS PART OF THE RESULT
Specific evidence. Honest limits.
What still needs people
Real screen-reader experience, meaning and usefulness of content, cognitive usability and complete accessibility conformance. Reflow probes use CSS viewport equivalents, not actual browser zoom.
Where Nemotron belongs
NVIDIA Nemotron through Nebius Token Factory can diagnose focused evidence and propose bounded source edits. It cannot change the journey oracle or grant itself a passing verdict.
Zero axe violations does not mean full accessibility. AX snapshots are not real screen-reader testing. This build is a local research workbench with curated synthetic applications.