Abstract: Motivation drift, defined as the gradual decline of student engagement during learning tasks, poses a persistent challenge in sustaining effective instruction. This study evaluates the use of a Hidden Markov Model (HMM) to detect motivation drift among 120 Grade 11 Humanities and Social Sciences (HUMSS) students from three schools in Manila, Philippines. A hybrid developmental and prescriptive design guided the creation of an Intelligent Tutoring System (ITS) prototype, which logged behavioral features such as task accuracy, response time, and hint requests. The HMM was benchmarked against Logistic Regression, Random Forest, and a lightweight LSTM model. Results show that the HMM achieved an AUC of 0.869, Accuracy of 96.0%, and the best Brier Score (0.073), with low detection delay (1.4 tasks) and a 2.1% false alarm rate. Strong generalization was observed through Leave-One-Student-Out (LOSO) validation (AUC=0.704). Feature importance analysis identified Time on Task and Correctness as key predictors of motivational dynamics. While inferred states correlated strongly with self-reported motivation (r=0.957), reliance on unimodal behavioral logs limits ecological validity. Future work should integrate multimodal data (e.g., facial expressions, voice tone) and address challenges in data ethics and computational cost through lightweight feature extraction and privacy safeguards. These findings confirm that HMMs are effective for real-time modeling of motivational dynamics in ITS environments, while highlighting opportunities for multimodal extensions and advanced sequential models such as Transformers and enhanced LSTMs.