Game rating analysis requires locating sparse evidence in long gameplay videos, but public rating decisions provide only bag-level labels. We present a multimodal multiple instance learning (MIL) framework that represents each video as a bag of 8-s visual, audio, and text instances. Using GRAC classification decisions, we constructed a dataset of 663 continuous gameplay excerpts selected to include potentially relevant evidence while retaining ordinary gameplay. Six modalities were evaluated with five-fold cross-validation using separate training, validation, and test sets. The validation-selected pipeline achieved a mean test macro-averaged area under the receiver operating characteristic curve (macro-AUC) of $0.869 \pm 0.074$. In a descriptive comparison of ensemble configurations using all six modalities, weighted soft voting attained $0.875 \pm 0.072$. Multimodal fusion improved bag-level Content Descriptor classification. Attention-based instance relevance scores support exploratory temporal analysis, but localization was not independently validated.