Christie Dam on Spencer Creek

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 0.801 and NSE 0.720 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 6 validation years the model scored KGE 0.302 to 0.898 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 FlowAi's water balance: outflow plus the storage change from the lake level and area plus lake evaporation, smoothed by a Kalman (RTS) smoother (rts-v4:n0.01:iv3.55:mb30e5). In the validation years 6 days are left out of training and scoring: 6 filled. 12 days floored at 0 are kept.