Automated online proctoring has become a practical requirement in remote and blended assessment, yet many systems depend on continuous connectivity, heavy processing, or intrusive data collection. This paper presents a proof-of-concept lightweight decision-support framework for detecting potential suspicious gaze and head-pose cues using MediaPipe-derived facial and iris features with an ExtraTrees decision layer. The study uses the Columbia Gaze dataset as a controlled proxy for off-camera gaze and non-frontal head-pose cues; it does not claim that such cues directly prove academic misconduct. A compact nine-feature representation is evaluated under a unified subject-independent protocol with validation-only threshold selection. On 1,260 held-out test images, ExtraTrees achieved 79.13% accuracy, 76.42% balanced accuracy, 62.50% macro F1, and 79.69% suspicious-cue recall. Because the dataset is highly imbalanced, an always-suspicious baseline achieves higher conventional accuracy but fails completely on normal recall; therefore, balanced accuracy, macro F1, normal recall, and false-alarm rate are treated as the primary interpretation metrics. The best model still produced a 26.85% false-alarm rate, confirming that the method should be used only as a triage signal requiring temporal confirmation and human review. The study provides a reproducible low-resource baseline and identifies the main steps needed before deployment: exam-like video validation, camera calibration, threshold calibration, and human-in-the-loop review.................... Keywords: .................online proctoring; MediaPipe; ExtraTrees; gaze cue detection; head pose; lightweight machine learning; computer vision; Columbia Gaze