Dark patterns steer consumers toward decisions they may not intend and often combine visual hierarchy with manipulative microcopy. This study presents an auditable multimodal pipeline comprising a YOLO-style branch for deceptive interface components, sparse word and character encoders for e-commerce text, and a structured consumer-risk-card layer. The empirical evaluation uses 2,356 unique ec-darkpattern strings (1,178 positive and 1,178 negative) grouped by 1,248 page identifiers to limit page-level leakage. The best binary model, a word-plus-character linear support vector machine, achieved 0.958 accuracy, 0.966 precision, 0.950 recall, 0.958 F1, and 0.988 area under the receiver operating characteristic curve. A probability ensemble produced the highest area under the curve (0.990) and a Brier score of 0.035. For positive-category classification, character n-grams reached 0.955 accuracy and 0.955 weighted F1, although macro performance remained lower for the rare forced-action and sneaking classes. The DarkLens package was also audited: its documentation declares 1,960 YOLO-format samples and five visual classes, but the available archive contains a Git-LFS pointer rather than materialized images and labels; consequently, no visual detection metrics are claimed. Deterministic templates generated schema-valid risk cards for every record, while risk-tier accuracy (0.593) and macro F1 (0.364) show that score-to-harm mapping requires human validation. The results support compact leakage-aware text detection and transparent data-readiness reporting, but not yet an end-to-end visual benchmark or validated consumer intervention.

Keywords: dark patterns, multimodal detection, interface microcopy, probability calibration, consumer-risk cards.