# Reproduction instructions

Requirements: Node.js 18 or later for command-line research scripts; Python plus Playwright and local Chromium only for optional browser integration tests.

```bash
node build_artifacts.js
node test_engine.js
python test_browser.py
```

`build_artifacts.js` reads source inputs from **the same folder**. It retrains the deterministic original baseline models on `train.json`, evaluates on `iid_test.json` and `challenge.json`, writes `evaluation_snapshots.json`, and produces `benchmark_manifest.json` with SHA-256 digests of the raw inputs and prediction artifact. No Internet connection or libraries are needed for the Node tasks.

After rebuilding, run `test_engine.js`. Browser testing uses Playwright `set_content`, and injects a fetch mock with the actual local JSON files because some testing sandboxes block loopback access. It uses bundled source data, not simulated model scores. Adapt the `executable_path` to your local Chromium installation if needed.

Expected metrics:

| Method | Controlled correct / n | Challenge correct / n |
|---|---:|---:|
| NB | 130/130 | 40/54 |
| Word TF-IDF | 128/130 | 39/54 |
| Character TF-IDF | 124/130 | 38/54 |
| Hybrid | 127/130 | 42/54 |

Source data: Academic Rhetorical Intent Mini v1.1, CC BY 4.0. No model weights are served or trained by the Hugging Face Static application on initial load.