Generating trustworthy insights on the multi-dimensional impacts of urbanization processes, including urban land and infrastructure expansion, require robust evaluation methods applied to model forecasts. Significant complexities and heterogeneities underlying urbanization make the forecast evaluations challenging. For instance, novel methods applying the state-of-the-art in generative artificial intelligence to forecast fine-scale urban morphologies have been proposed. However, evaluating outputs from these models remains a challenge. Here we propose a new focus on emergent empirical regularities, as one way to evaluate forecasts of urban quantities. We demonstrate this proposition by analyzing heterogeneities in the distribution of building attributes, i.e., building floor area and volume, within and across metropolitan areas (MSAs) in the United States (US). We find a remarkable empirical regularity in the level of heterogeneity in building attributes from Oak Ridge National Laboratory’s (ORNL’s) Model America dataset when observed at multiple spatial scales defined by regular grids of varying spatial resolutions specified by cell sizes (r km) such that r ∈ [5, 5. 5, ... , 29, 29. 5 ??], consistent with Taylor’s power law (TPL) where a power law underlies the variance-mean relationship. We estimate the average TPL exponent (α) across 382 US MSAs to be 1.56 (95% confidence interval: 1.550-1.575) in the case of building volume distributions, with 80% of the MSAs with α ∈[1. 41...1. 66] and for all individual MSAs α > 1. We obtain similar results for building floor area. The stability of the exponents across MSAs represent a significant regularity in the degree of spatial aggregation of these building attributes, which we propose as an empirical constraint on multi-scale heterogeneity. Overall, our results point to the value of analyzing empirical regularities in urban quantities towards evaluating as well as constraining outputs of models forecasting urban quantities.