phyground.github.io

Public companion site for the wmbench physics-grounded video benchmark.

• Video Gallery — browse video-generation model outputs, side-by-side per physical law.
• Leaderboard — slice rankings by evaluator × dataset × subset × scoring schema.

The site is static and deployed via GitHub Pages from the repository root. Videos are hosted on HuggingFace; this repo only carries the index/manifest, rendered HTML, and a small CSS asset.

Architecture

phyground.github.io/                 ← GitHub Pages serves this root, with Jekyll disabled
├── .nojekyll                        ← disable Jekyll on GitHub Pages
├── index.html                       ← rendered home page          (BUILT)
├── leaderboard/index.html           ← leaderboard                 (BUILT)
├── videos/index.html                ← video gallery               (BUILT)
├── about/index.html                 ← about / citation            (BUILT)
├── static/css/base.css              ← mirrored from tools/static_src
│
├── tools/                           ← build infrastructure (source of truth)
│   ├── build_snapshot.py            ← _wmbench_src/ → snapshot/
│   ├── verify_snapshot.py           ← sha256 check vs snapshot/MANIFEST.json
│   ├── build_site.py                ← Jinja2 + snapshot → static HTML
│   ├── site_config.example.json     ← stub config (offline preview without a snapshot)
│   ├── templates/                   ← Jinja2 templates
│   └── static_src/                  ← CSS/JS source, mirrored to /static
│
├── _wmbench_src/                    ← hard-copied wmbench source (frozen; no network deps)
│   ├── PROVENANCE.md                ← wmbench HEAD sha + per-file sha256 at copy time
│   ├── data/{vis_datasets.json, paperdemo/{manifest.csv, figs/*.pdf}}
│   ├── evals/{eval_registry.json, eval_types.py}
│   └── videogen/runner/MODEL_CATALOG.py
│
├── snapshot/                        ← deterministic build output, the site's only data source
│   ├── MANIFEST.json                ← sha256 over every snapshot file
│   └── index/{site_config.json, eval_registry.frozen.json,
│              paperdemo.manifest.csv, vis_datasets.frozen.json,
│              model_catalog.frozen.json, humaneval_100.json}
│
├── docs/exp-plan/public/{plan.md, humaneval_100.md}
├── requirements.txt
└── LICENSE

Quick start

git clone https://github.com/phyground/phyground.github.io.git
cd phyground.github.io
python -m venv .venv && source .venv/bin/activate    # optional
pip install -r requirements.txt

# 1. Build the data snapshot from _wmbench_src/.
python tools/build_snapshot.py

# 2. Verify the manifest.
python tools/verify_snapshot.py

# 3. Render HTML against the snapshot.
python tools/build_site.py --config snapshot/index/site_config.json

After step 3 the rendered files at the repo root (index.html, leaderboard/index.html, videos/index.html, about/index.html, static/css/base.css) are exactly what GitHub Pages will serve.

For an offline preview without a snapshot (the renderer reads templates only):

python tools/build_site.py                            # uses tools/site_config.example.json

Determinism

• tools/build_snapshot.py produces byte-identical snapshot/ output for the same _wmbench_src/ inputs (any timestamp can be pinned via --now <ISO>; the default uses UTC now() only for build_meta.built_at).
• tools/build_site.py produces byte-identical HTML for the same config + templates.
• tools/verify_snapshot.py exits non-zero on any drift between snapshot/MANIFEST.json and the actual file tree.

Refreshing wmbench data

When upstream wmbench files change:

1. Re-copy the affected files into _wmbench_src/ (preserving paths). The expected layout is documented in _wmbench_src/PROVENANCE.md.
2. Update _wmbench_src/PROVENANCE.md (HEAD sha, sha256 per file, copy date).
3. python tools/build_snapshot.py && python tools/verify_snapshot.py.
4. python tools/build_site.py --config snapshot/index/site_config.json.
5. Commit the diff (the snapshot/index/*.json files travel with git so the deployment is reproducible).

Deployment

GitHub Pages serves the repo root on the default branch:

1. Run the four-step Quick Start above on your laptop.
2. git add the rendered HTML, static/, snapshot/index/, snapshot/MANIFEST.json.
3. git commit && git push.
4. GitHub Pages picks up the change at https://phyground.github.io/.

The .nojekyll file disables GitHub's automatic Jekyll processing so the rendered HTML is served verbatim.

Plan and progress

• Implementation plan: docs/exp-plan/public/plan.md
• humaneval-100 selection spec: docs/exp-plan/public/humaneval_100.md
• _wmbench_src/ provenance: _wmbench_src/PROVENANCE.md
• RLCR loop state: .humanize/rlcr/<timestamp>/

License

Code: MIT (see LICENSE). Generated videos: per upstream model license.