Fuel Price Anomaly Radar

Every month, a GitHub Actions job downloads the newest price survey published by ANP, applies the same feature engineering and the same model (OneClassSVM, contamination=0.01) trained and evaluated in the repository's 04_avaliacao_tuning.ipynb notebook — no retraining, it's production scoring on new data.

Caveat: states with few records (e.g. Acre, Amapá) get noisier percentages — a handful of stations can already swing the rate a lot. On top of that, raw price is still one of the model's features, so historically more expensive states (the North region, generally) can show a higher apparent rate even with nothing actually wrong going on there. Read these numbers as a lead for investigation, not a verdict.