What the visual layers show

Two separate attention layers

• Element + task emphasis: deterministic geometry and typography signals combined with explicit search-term matches and a configurable paragraph adjustment. It is a review aid, not a calibrated probability.
• Research pixel overlay: a saved experimental map. Its provenance records model and input hashes. The underlying learned model and third-party training images are not distributed here. Desktop and mobile outputs are normalized separately and should not be compared as absolute attention budgets.

Captured evidence

The report separates initial HTML, rendered content, browser accessibility-tree order, actual Tab-key sequences, semantic reviewer notes and missing evidence. A browser AX tree is not a recorded screen-reader session. Policy planner presets change advice only.

• Pixel-output provenance
• Saved output arrays (NPZ) — output maps, not model weights
• AX and keyboard evidence
• Design review evidence
• Initial HTML as inert text · Rendered HTML as inert text

Research references

• Predicting Visual Attention in Graphic Design Documents — component-aware saliency and scanpaths motivate evaluation; this report does not run that model.
• DeepGaze research implementations — referenced for model comparison; DeepGaze code and weights are not bundled or executed here.
• RFC 9309 — crawl directives are separate from authorization.
• WCAG 2.2 — sampled measurements need context and exception review.

Run a new page audit. The hosted beta uses element heuristics and browser evidence; it does not run this optional research pixel model.