Research page prototype

The page at /research/ uses the site’s existing default layout, typography, navigation, and footer. Its SCSS and plain JavaScript are loaded only by the Research layout. It adds no visualization dependencies, external API calls, or embedding computation. Homepage, Publications, and Talks sources are untouched. The only data request is for the same site’s static /assets/data/research-map.json. It uses no external runtime API. Serving JSON separately prevents the shared HTML compressor from altering abstract strings. The old duplicate Research collection document was removed to give /research/ a single source in _pages/research.html.

Data source

_data/research_map.json contains the genuine corrected analysis exported from research_analysis/results-corrected/ on branch work, commit 00a947e9886cecc6af6b8e66a36b726020c851eb. All 15 abstracts match the current website corpus, including P09’s corrected breakup-reactions abstract. The original analysis files remain unchanged on work; they are not required at website runtime. The page is ready to render from the committed JSON without Python or a model.

From the repository root, with the completed analysis results available:

python -m pip install PyYAML==6.0.2
python tools/research_map/export.py --results research_analysis/results-corrected

Required files: publications.json, pca_coordinates.csv, similarity_matrix.csv, and run_metadata.json. The exporter writes only _data/research_map.json; it never changes the analysis files. Commit that export on this feature branch to make the static page work on GitHub Pages. The full matrix is validated; neighbors use its original values with ID order as a deterministic tiebreaker. Source artifact SHA256 hashes record provenance. Titles and abstracts must match the current corpus (ignoring whitespace only). If the original analysis used an outdated abstract, the exporter rejects the mismatch rather than silently presenting it as current research.

Design and interaction

The four themes and their publication assignments are researcher-defined. The SVG uses one uniform affine scale of original PCA coordinates, with the vertical axis flipped for screen coordinates. No clustering, jitter, force layout, or category-based repositioning is applied. Filtering leaves the scale and point positions fixed. Similarities in the detail panel come from the full embedding space and include publications in other/hidden themes; selecting a neighbor reveals its theme. Displayed scores are rounded to three decimals only.

The header follows the Publications/Talks spacing, and the desktop map/details columns use a 1.9:1 ratio, stacking on mobile. Point titles appear on hover or keyboard focus rather than as overlapping permanent labels. Selected papers have a dark outline; their three closest papers have outer rings and subtle connecting lines. Theme filters still hide points and their connections, while all three original rankings and scores remain in the detail panel. No points are moved.

The detail panel initially shows an excerpt of up to 320 characters, cut at a word boundary. Read full abstract / Show less toggles the complete original text without rewriting it. Selecting another paper resets the excerpt view.

Theme filters are native toggle buttons styled as color-coded chips, with aria-pressed state and keyboard activation. The 2016–2026 slider is inclusive: it shows publications through the selected year. Show all years is checked by default; moving the slider turns it off, and checking it restores 2026. Theme and year filters intersect for the map, title list, and visible connections. If filters hide the selected paper, the detail panel clears without moving keyboard focus. Related papers retain their top-three collection-wide rankings; filtered papers are marked outside current filters. Selecting one reveals its theme and extends the year cutoff if needed. Scores and PCA coordinates never change.

The short introduction sits above the map; technical context is in About this map. Related-work copy explains that collection-relative textual rankings do not necessarily imply strong relationships, and connections use the original embedding space rather than PCA distances.

Mouse hover and keyboard focus show titles and years. Enter/Space or a tap opens the publication detail panel; all selection actions also have native buttons in the title list for mobile and assistive technology. Theme chips provide a legend and filtering. All themes may be deselected, with a clear empty state. Publication abstracts remain available without JavaScript. On narrow screens, the detail panel stacks below the map and the title list provides larger targets.

Verification

node --check assets/js/research-map.js
python -m unittest discover -s tools/research_map -p 'test_*.py'
npm install --prefix /tmp/research-map-tests jsdom@26.1.0 --no-audit --no-fund
NODE_PATH=/tmp/research-map-tests/node_modules node tools/research_map/test_filters.cjs
bundle install
bundle exec jekyll build
bundle exec jekyll serve --host 0.0.0.0 --port 4000

After exporting genuine data, check /research/ at desktop and mobile widths: select each theme, clear all themes, restore all, use Tab/Enter/Space, select a publication and its three neighbors, and inspect each link. Verify markers do not move during filtering and compare PC1/PC2 and neighbor values to the source CSVs. Review the homepage, Publications, and Talks for unchanged rendering.

JavaScript syntax, SCSS compilation, Python syntax, and all four exporter validation tests pass. All 15 coordinate pairs and 45 displayed neighbor scores match the corrected source CSVs exactly. DOM checks using the corrected data cover marker creation, abstract text, neighbor ordering, all/none filtering, unchanged positions, and keyboard selection. The saved DOM regression suite exercises all 11 year cutoffs against all four themes using the genuine committed data, inclusive boundaries, reset and empty states, selection clearing/reveal, rankings, abstract expansion, and fixed geometry. It also runs in the PR build workflow; jsdom is a test dependency only. The production build passed with Ruby 3.2 and Jekyll 3.10 locally in a container and in GitHub Actions. Chromium checks at 1280px, 390px, and 320px widths passed for all 15 original abstracts and 45 neighbor links, keyboard and pointer selection, filters, fixed geometry, no horizontal overflow, and the no-JavaScript fallback. The build retains existing warnings in unrelated Teaching, Software, CV, and collection excerpts; those pages were not edited.

Preview before merging

Open a Codespace on feature/research-map (not work or master). From the repository root:

git switch feature/research-map
git pull --ff-only origin feature/research-map
bundle install
bundle exec jekyll serve --host 0.0.0.0 --port 4000

In the Codespaces Ports panel, open port 4000 in your browser and append /research/ to the forwarded URL. This previews the branch without deploying or merging it. The analysis model does not need to be installed or run.

The PR’s Research prototype build workflow also runs a production Jekyll build and uploads a research-site-preview artifact. It never deploys the site. To view a downloaded artifact locally, unzip it, run python -m http.server 4000 inside the unzipped directory, and visit http://localhost:4000/research/.