The first step is design: stimulus-driven design. In experimental psychology, researchers typically formulate specific hypotheses and then test them. Although valuable, this strategy can constrain discovery to possibilities anticipated in advance. Stimulus-driven design allows researchers to move beyond this constraint. In ImageNet, variations across millions of images encode rich information about object categories; even when researchers do not know the relevant regularities beforehand, modeling can reveal them. Similarly, in a recent CE study of visual working memory (Huang, 2025), the underlying mechanisms were not specified in advance. Instead, modeling performance across randomly generated color patterns recovered all the classic mechanisms examined, as well as several novel ones.

The second step is data: large-scale experimentation. As in AI, this stage establishes a robust empirical foundation. For instance, Huang (2025) collected 40 million responses across 10,000 color patterns. This scale provides the basis for a comprehensive exploration of the phenomena of interest.

The third step is analysis: iterative model development, the most labor-intensive stage of the CE approach. The general strategy is to draw on both theoretical insights from the literature and patterns observed in the data to explore candidate model components and alternative ways of assembling them into a complete model. 

The fourth step is outcome: an interpretable computational model. A defining characteristic of this outcome is the deliberate balance between predictive accuracy and interpretive parsimony. Because the model is formal and computational, it offers greater precision than accounts based primarily on verbal interpretations. At the same time, it is kept relatively simple: unlike “black-box” neural networks, CE models are explicitly designed to map computations onto psychologically meaningful mechanisms. This balance helps bridge the gap between explanation and prediction. Neural networks often achieve high predictive accuracy at the expense of interpretability, whereas traditional psychological theories are often interpretable but lack comprehensive predictive power. CE models seek to combine these strengths. For example, the model in Huang (2025), containing only 57 parameters, outperforms a complex neural network with 30,796 parameters. The CE approach thus aims to deliver a relatively simple, interpretable theory with broad explanatory and predictive power.