Run the code locally

Run the website and reproduce the research experiments.

Rate some drawingsSee how the preference model learns. Inspect a click predictionSee where the model pointed and whether it was right.

For developers: run the site and experiments

Run the site

python scripts/build_site.py
python -m http.server 8768 --bind 127.0.0.1

Open http://127.0.0.1:8768. Reviewed page content lives in content/pages.json; the builder produces ordinary HTML. The browser preference demo works locally without a model server.

Run the preference checks

python -m venv .venv
.venv\Scripts\python -m pip install -r experiments/requirements.txt
.venv\Scripts\python experiments/preference.py
.venv\Scripts\python experiments/run_synthetic.py

Node.js is used to check the same mathematical fixtures against the browser core. The synthetic suite uses one CPU thread and records its actual results under results/synthetic/.

Run model pilots

Model pilots require separate dependencies and downloaded checkpoints. Read the experiment instructions before running them. Their scripts cap CPU parallelism and pin the model revision. The small Qwen pilot returns labels from learned heads; the coordinate pilot reads a pretrained action head.

The original feature-space Python example is retained as a distinct illustrative approach. It is not the canonical latent-logistic reference.

What to report

Keep the model/data revision, seed, split, input sizes, hardware/runtime, timing scope, quality, and failures. Generated outputs and synthetic scores are not independent evidence of human utility.

Registered protocol · Current results · Domain proposals