Representative SR model performance. Panel (a): plane spanned by the two terms in the representative model that correlate explicitly with flare probability, evaluated over the full dataset: x-axis ﹩{R}^{2}(t-3{\rm{\Delta }}t){H}_{{c}_{{\rm{total}}}}(t){H}_{{c}_{{\rm{abs}}}}({t}_{{\rm{\max }}})﹩, y-axis R(t − 2Δt)R(t). Left: calibrated probabilities from the deep learning model. Right: ratio of positive (flaring) to total samples in each bin. Panel (b): Forecasting examples for AR 13664 and AR 12673. For each AR, the horizontal strips above the time series plot summarize correct/incorrect prediction windows at a 0.5 threshold over the intervals where each model provides output (green: correct, red: incorrect). The lower plots show probability time series for the available models; purple lines denote our SR model, and black vertical segments mark observed M-class and above flares. Archived empirical models from the CCMC flare scoreboard (e.g., AMOS, AEFFort, MAG4) are included where available for each AR.