Fig. 2: Framework performance and the observational constraint.

a, Red circles show abrupt-4 × CO₂ simulation results (‘actual’) regressed against predicted changes in q_strat, both normalized by T_g, for 27 CMIP models. The multi-model mean is indicated as a black square; the one-to-one line in solid black. Dashed lines show the least squares regression fit (black) and the 5% to 95% prediction intervals (red). The probability distributions (red curves) on the axes represent the observational estimates. The distribution on top of the x axis indicates the spread in predictions based on combining functions learned from observations with the CMIP temperature responses. The final probability distribution, defining the observational constraint, is attached to the y axis and additionally accounts for the framework prediction uncertainty. b, The observational constraint (n = 4,050) relative to CMIP model uncertainty. Circles show T_g-normalized changes in q_strat for 27 CMIP5 models (red), 34 CMIP6 models (blue) and their combination (black). The grey circles indicate the selected 27 models fulfilling the minimum variance criterion compared with SWOOSH used for the framework validation in a. The observational constraint (orange; Obs) is illustrated on the right with the horizontal black line indicating the 50th percentile (0.31 ppmv K⁻¹). The thin and thick bars denote 90% (−0.09 to 0.69 ppmv K⁻¹) and 66% (0.08 to 0.54 ppmv K⁻¹) confidence intervals, respectively. The CMIP mean (median) values are 0.67 (0.53) ppmv K⁻¹ for CMIP5, 0.55 (0.49) ppmv K⁻¹ for CMIP6, 0.60 (0.52) ppmv K⁻¹ for the combined set of CMIP5/6 and 0.63 (0.59) ppmv K⁻¹ for the 27 selected models.