# XGBoost Arbiter

Binary classifier that predicts P(proposed_correction == ground_truth) from spellevator pipeline features.

## Latest: seed48

- **F0.5**: 0.8723 @ threshold 0.70
- **Precision**: 89.5%
- **Recall**: 79.2%
- **Config**: max_depth=6, n_rounds=500, eta=0.1

## Usage



## Files

- `model.json` — XGBoost booster
- `encoder.json` — category mappings for FeatureEncoder
- `training_log.json` — full metrics, threshold sweep, hyperparams
- `feature_importance.json` — feature gains

## Training Data

400k synthetic records from openrouter-llama70b-50k with programmatic corruption (90% corruption ratio).