It is difficult to determine whether reported gains of machine learning (ML) over the classical power-law probabilistic seismic demand model (PSDM) stem from algorithmic improvements or simply better testing. This difficulty arises because published comparisons often use different inputs and validation schemes for reinforced concrete (RC) buildings within performance-based earthquake engineering (PBEE) and intelligent design pipelines. Following a Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 review of 23 ML-based PSDM studies, a benchmark evaluation was conducted within the Pacific earthquake engineering research (PEER) PBEE frame-work. Five RC frames were analyzed: three designed to Eurocode 8 (EC8) Ductility Class Medium (DCM) and two to American Concrete Institute (ACI) 318-19/American Society of Civil Engineers (ASCE) 7–22. Their fundamental periods (T₁) ranged from 0.42 to 1.60 s. These frames were subjected to 56 far-field next generation attenuation (NGA)-West2 records, scaled to seven S_a(T₁) values, yielding N = 1960 nonlinear time-history analyses (NLTHA). Both random forest (RF) and XGBoost were compared to a calibrated PSDM using identical 80/20 stratified datasets with a five-feature input vector. Other approaches, such as different kinds of deep learning architectures (convolutional neural networks (CNNs), long short-term memories (LSTMs), and Transformers), were evaluated but excluded from the final comparison to isolate the effects of the algorithms. XGBoost achieved R² = 0.93 and σ_ln = 0.41, outperforming the calibrated PSDM (R² = 0.86) with ΔR² = 0.07 (95% confidence interval (CI) [0.04, 0.10], p < 0.001); and random forest (R² = 0.87) showed no statistically significant improvement over the PSDM (p = 0.21). Cross-validation used Group-KFold by record sequence number (RSN) to prevent intra-record leakage. A leave-one-archetype-out (LOAO) test gave a mean R² = 0.846, benefiting from an independent external dataset for holdout. Total dispersion (β_total) was ap-proximately 0.48 for XGBoost and 0.45–0.55 for the stratified PSDM. Quantile regression forest (QRF) intervals achieved an empirical coverage of 88.6% at a nominal 90%. S_a(T₁) exhibited the highest feature importance (mean decrease in impurity (MDI) = 38.8%; cross-study average of 28%). An integrated ML-PBEE pipeline with a regulatory-adoption checklist is demonstrated. Its conclusions are established for the tested configuration of regular, code-conforming RC frames under far-field motions and should be revalidated before extension to other structural classes.