Auditing Audiovisual Correspondence in Generative AI Coursework with Frozen Multimodal Encoders

Computational evaluation of generative artificial intelligence coursework requires distinguishing audiovisual correspondence from creative achievement. We conducted a retrospective computational measurement audit of 49 archived student submissions containing 58 videos. Frozen CLIP and CLAP representations were queried with paired event and affect descriptions. Five fixed anchor configurations and a pretrained AudioCLIP comparator were evaluated through original-pair retrieval, temporal and coordinate controls, blur and noise transformations, teacher judgments, and selected cases. The primary configuration achieved a mean reciprocal rank of 0.179, compared with 0.093 for the designated AudioCLIP comparator; its advantage over other temporal anchor configurations remained uncertain. Temporal and anchor-index disruptions reduced discrimination, while perturbations could increase matched scores despite poorer retrieval. During revision, two teachers independently rerated all submissions after calibration on separate examples and rubric freezing. Total-score agreement was ICC(A,1) = 0.970 and ICC(A,2) = 0.985, but the primary diagnostic’s association with their mean total was near zero (Spearman ρ = 0.020). A few adjusted dimension-level associations did not retain significance in the non-silent sensitivity analysis. Selected cases exposed unverified semantic responses and nonzero outputs from silent audio. The workflow supports inspectable correspondence diagnostics; artistic grading and educational usefulness require further validation.