Shear failure of steel-reinforced-concrete (SRC) columns and beam–column joints can precipitate rapid strength loss during earthquakes, yet current code formulas treat shear strength deterministically and mask large epistemic uncertainties. This study couples mechanics-based shear models with two data-centric approaches—(i) Bayesian inference applied to closed-form truss-and-strut formulations and (ii) probabilistic machine-learning (ML) surrogates trained on the same experimental database—to quantify and compare prediction accuracy and uncertainty. A curated set of 230 tests spanning wide ranges of shear-span ratio, axial load, and material strengths provides the calibration benchmark. Posterior sampling via Markov-chain Monte Carlo delivers median capacity curves and credibility bands for the physics models, while Gaussian-process and Bayesian-neural-network regressors supply non-parametric uncertainty envelopes. Results show that the Bayesian physics model retains physical interpretability and achieves a 30 % reduction in root-mean-square error relative to leading code equations, whereas the ML surrogate attains comparable accuracy with narrower predictive intervals in data-dense regions but extrapolates poorly outside the training domain. Fragility analysis further indicates that both probabilistic strategies expose non-uniform reliability in current design provisions, particularly for squat columns and exterior joints. The comparative framework highlights trade-offs between transparency, generalisability, and data demands, and lays the groundwork for risk-consistent, performance-based seismic design of SRC components.