Many users want to reduce their screen time and time spent on social media. Yet, most existing interventions rely on static rules that may not fully account for variations in user context and receptivity. In this paper, we propose the design of SmartPause, a Just-in-Time Adaptive Intervention (JITAI) based on Reinforcement Learning (RL) that adapts intervention timing by learning directly from users’ natural compliance behavior. The system utilizes a Q-learning algorithm to learn the optimal moments for delivering vibrations as subtle digital nudges. Informed by an extension of the Predictability, Computability, and Stability (PCS) framework, we assess the algorithm's stability and potential intervention burden prior to deployment. We conducted offline simulations using historical usage logs from 26 participants (totaling 1,118 user-days) and five simulated personas, varying their responses to the intervention regarding usage time and in-session compliance. We evaluated four versions of the RL algorithm, differing in the frequency at which the model was queried and the criteria for separating different sessions. Our analysis suggests that all tested algorithms present a risk of highly burdening the users with too frequent interventions, especially for personas that are more compliant with the intervention. While personalizing querying frequencies might increase intervention opportunities, they also increase the risk of a high user burden. Further simulations should explore alternative approaches for balancing learning rates and user burden risk. This work contributes to the evolving field of using machine learning and RL to inform JITAI, and will be further evaluated in field studies using the SmartPause app.