Machine-learning models for the shear strength of reinforced concrete (RC) walls are usually validated on a random split of a test database and are reported to be far more accurate than code equations. This study re-examines that claim, and asks what such models are worth for design. A database of 208 wall tests, assembled from the ACI 445B repository and from the databases of two recent empirical models, was analysed. The ACI 318-14 in-plane shear provision was re-implemented and reproduced the tabulated predictions to within 2.1%. Its model uncertainty θ = Vexp/Vpred has a mean of 1.08 and a coefficient of variation (CoV) of 0.52; the truss model of Chandra et al. (2018) and the particle-swarm equation of Baghi et al. (2019) give CoVs of 0.46 and 0.33. A random-effects decomposition shows that 67% of the variance of ln θ for ACI 318-14 lies between test programmes rather than within them. Gradient-boosting models were validated both with random ten-fold cross-validation and with leave-one-study-out validation. The random split gave RMSE = 118 kN (R2 = 0.93); leaving a whole test programme out raised it to 298 kN, worse than the code equation (250 kN). A physics-guided model that learns the logarithm of the ACI model uncertainty from dimensionless features was the only data-driven model that generalised (219 kN, CoV of θ 0.39), and SHAP analysis showed why: ACI 318-14 overestimates the contribution of horizontal reinforcement and ignores the contribution of flanges. Reliability analysis with a first-order reliability method gave β = 1.12 for ACI 318-14 at φ = 0.75 against β = 1.57 for the physics-guided model, under a gravity load model. Neither is close to the usual target, which shows that the scatter of wall shear databases, not the mean accuracy of the equation, is what limits the reliability that a shear-strength model can deliver.