# nspam — Nostr reply-spam classifier (v2.2)

On-device spam scoring for Nostr reply notes. See `model_card.md` for details.

## Files

| file | purpose |
|---|---|
| `model.txt / model.json` | trained model |
| `calibration.npz` | isotonic calibration table |
| `config.json` | feature layout + hashing convention |
| `parity_fixtures.jsonl` | reference inputs + expected scores |
| `hash_fixtures.jsonl` | token-level hash checks |

## Integrating

A port needs:

1. **Feature extraction** — reproduce preprocessing + hashing + structural
   features. See `config.json` for the exact layout.
2. **Model inference** — load `model.txt` via LightGBM4j (Kotlin/Java),
   LightGBM C API, or ONNX conversion.
3. **Calibration** — piecewise-linear interpolate the raw score through
   `calibration.npz` knots.

Validate your port against `parity_fixtures.jsonl`.