Scrutinizer × AdSERP

Each background image is a full-page foveated render produced by Scrutinizer, a neuroscience-based peripheral vision simulator. Scrutinizer models the human visual system's resolution falloff from fovea to periphery using a pipeline inspired by the lateral geniculate nucleus (LGN), primary visual cortex (V1), and difference-of-Gaussians (DoG) spatial frequency filtering.

Infinite visual memory mode accumulates every fixation across the entire scanpath. As each fixation is replayed, the foveal region at that position is "remembered" — remaining sharp while everything else degrades through the peripheral pipeline. The final image shows exactly what the participant could have resolved across their full search session: sharp where they looked, degraded where they didn't.

Pipeline:

1. The interesting trials script identifies prototypical search behaviors from 2,776 AdSERP trials
2. The AdSERP importer parses fixation CSVs (page-space coords from Gazepoint GP3 HD at 150Hz), mouse events (evtrack pageX/pageY), scroll timelines, and trial metadata — reconciling two coordinate systems (coordinate reference)
3. The fullpage gazeplot script walks each fixation through Scrutinizer's Electron-based WebGL pipeline with TEST_VISUAL_MEMORY=-1 (infinite accumulation), then tile-captures the full page at each scroll position and stitches them
4. The batch gazeplot mode bulk-loads all fixation positions into the visual memory buffer at once (instead of walking them one-by-one through the render loop), then tile-captures the page — reducing capture time from minutes to seconds per trial
5. The interactive overlay is generated as self-contained HTML with the gazeplot PNG as background
6. The saliency export CLI (docs) computes per-fixation saliency and Rosenholtz feature congestion using the same Oklab DoG + local variance pipeline as the real-time renderer — but headless in Node.js, no GPU required (~100ms/image)

Multi-track timeline: The timeline strip shows five parallel data streams for each fixation: Saliency (visual pop-out at gaze location), Clutter (Rosenholtz feature congestion), Pupil (dilation from 150Hz pupillometry), Scroll (forward/regression direction + depth), and Dwell (fixation duration). Bar height encodes intensity within each track; color encodes the signal value. Scrub across fixations to see how visual complexity, cognitive load, motor behavior, and attention duration co-vary in real time.

DOM-anchored fixation positioning: Fixation coordinates from the eye tracker are in screen-pixel space (1280×1024), but the SERP was displayed at 1422 CSS pixels (90% display scaling). Rendering at a different width causes content reflow, shifting elements vertically. Instead of coordinate transforms, each fixation is anchored to the DOM element it landed on: during the build step, generate-anchors.js loads each SERP in Playwright at the original window width, uses elementFromPoint to map each fixation to a CSS selector + offset, then the build script re-resolves those anchors at the render width. The fixation dot lands on the correct element regardless of layout width. This approach was inspired by DOM-based event logging (Edmonds, 2003) — the same principle that anchors interaction events to page structure rather than screen coordinates.

Data: AdSERP — Latifzadeh et al., 2,776 transactional Google SERP queries, 47 participants, Gazepoint GP3 HD eye tracker, simultaneous mouse + scroll + gaze recording. Dataset on Zenodo.