Intelligent Tutoring Systems (ITSs) have traditionally focused on adapting instruction primarily based on student models. However, as adaptive training environments grow in complexity – with diverse intervention strategies and accounting for a multitude of student states – this singular focus can leave the system blind to the pedagogical effectiveness of its own actions. In this paper, we follow up on a modular ITS architecture [1] that integrates an independent tutor model alongside the student model. The tutor model logs and evaluates the tutor’s decision-making process in real time, enabling a more transparent and continuously refined approach to adaptive decision-making. We discuss the theoretical framing of pedagogical policy decision-making, review existing methods, detail the architecture of our tutor model (as implemented in the STATS adaptive training ecosystem [2]), and introduce a unified metric – the Action Performance Score – for evaluating intervention effectiveness. Finally, we explore implications for ITS design and propose avenues for future research.