Du Nord

Physical hydrological model (Raven), corrected by a machine learning model (per-gauge XGBoost) trained on the residuals. Where the correction stops before the last day, it fades into Raven over two days. Current validation KGE 0.880 and NSE 0.778 at a 7 day horizon, measured on the record used to calibrate and select this model, so it is not independent evidence: every live gauge clears the gate because clearing it is what makes a gauge live. These are the figures the gauge must keep clearing to stay live, re-checked at each daily revalidation on the current observation record, not a fixed score from the day it was promoted (last re-checked 2026-10-10). Over its 5 validation years the model scored KGE 0.715 to 0.903 at 7 days, each year on its own. Operational skill is worse than these figures, because the forecast starts from a modelled river state that is systematically off at issue time. The natural inflow series is the provider's, as published, not smoothed.