Traditional reductionist paradigms in cognitive science often miss the richness of real-world behavior. While naturalistic paradigms have recently expanded in vision, audition, language, and social cognition, publicly available closed-loop neuroimaging datasets—where participants actively act on and reshape their sensory input—remain scarce. In parallel, state-of-the-art artificial agents can now learn to navigate a wide range of situations with high proficiency, but fall short at generalizing to unseen contexts, which humans excel at. To bridge these gaps, we built the Courtois Neuromod Mario dataset as part of the CNeuromod project: an open deep-phenotyping resource spanning multiple cognitive domains, including 84 hours of Super Mario Bros. (Nintendo, 1985) gameplay collected from five naïve participants. In this talk, I will introduce the dataset, motivate the design choices behind its acquisition and annotation, and show how reliable brain responses can be recovered from naturally occurring game events. I will then present a comprehensive framework for operationalizing the platformer by segmenting continuous gameplay into short, well-contained, unitary challenges that facilitate the parsing and analysis of behavior. Finally, I will discuss ongoing and future research avenues, including the development of benchmarking baselines for continual reinforcement learning, investigations of the neural basis of motor sequence learning, and the use of intracranial EEG to identify neural correlates of flow states. Together, these ingredients position Super Mario Bros. as a total laboratory, where multiple distinct experiments targeting different facets of cognition can be conducted within a single, unified environment.