# Models

## v10_honest (primary)

- Per-class binary XGBoost heads (career, relationship, family, health_death, financial, education, relocation, social, life_events, conflict_legal)
- **626 features** from Swiss Ephemeris **V8FeatureGenerator** + miss_house/dignity/arabpart
- Inference: official Python `xgboost` Boosters (train parity)
- Use for production heatmaps

## clean_v4 (experimental)

- Single binary event score
- Not the same feature contract as v10
- Marked experimental in UI/API

## Health check

`GET /api/health?deep=1` reports `featureParity` and canary variance for v10.