# Paper demo UI · r7

User-requested revision, 2026-09-15. The original reference layout is preserved;
purple/lime colors follow the supplied paper figure. The XPENG logo is unchanged.

- English interface, task guides, runtime states, recovery messages and accessibility labels.
- Predictions appear directly on the main image. The original/model/reference view selector
  and reference-answer UI endpoint have been removed. Example input boxes in the Visual
  Prompt task remain input conditions, not reference answers.
- Overlays use unfilled outlines. A large late region can no longer erase smaller targets.
  Class labels are laid out separately, clamped, and omitted if they would cover small
  targets or another label. Dense classes remain available in the linked inspector.
- Geometry is not deduplicated, suppressed, merged, moved, or changed for presentation.
- The input image and prompt still form the entire model request. No ground truth is sent.
- Case counts and training-set counts are no longer presented as headline copy.
- The original release lock remains in `ui_lock.json`; `ui_paper_r7_lock.json` records
  this explicitly authorized revision. Data, source examples, and model weights are unchanged.

`validate_gradio.py` uses the new direct-overlay API result order:
image, metrics, targets, normalized JSON, raw tokens, status, desktop action, mobile action.
Historical API receipts remain unchanged. This revision is not a new accuracy evaluation.

Verification: run `python -m unittest discover -s tests -v` and `python scripts/verify_ui.py`.
Browser verification must cover desktop/mobile layout, direct overlays, category/instance
focus, zoom/reset, upload and case transitions, and error recovery. The old reference-view
browser script is historical and does not match this UI contract.