# Reproducing the experiments

## Browser

Upload the repository files to one root folder in a Hugging Face Static Space. Open App. Models fit automatically (locally) to `train.json`. Open Overview, or Experiments → Run four baselines. Select challenge (54) or IID (130); enter a bootstrap seed. Download metrics JSON and prediction CSV.

## Node, no dependencies

```sh
node --test test_engine.js
node benchmark_cli.js
```

Both commands use only Node 18+ built-ins and the bundled dataset files. No Hugging Face token, Python packages or GPU needed.

## Browser test, optional

```sh
python -m pip install playwright
python -m playwright install chromium
python test_browser.py
```

The test injects local project files directly into the Chromium page and mocks `fetch` *only for fixture-loading* to avoid external network dependency in restricted CI. This is not a test of deployed Hugging Face hosting.

## Interpreting results

Use the fixed split names. Do not compare any archived Python-model evaluations as if they used the same code. Synthetic evaluations are not proof of generalization to real scholarly work.