Reliability Audit for Mixed-Authenticity Audio-Visual Deepfake Detection

This license-safe derived-evidence package supports the Pattern Recognition manuscript entitled “Auditing Target Alignment and Marginal Contributions in Multimodal Pattern Recognition under Composite Labels.” It contains frozen aggregate reports, anonymized per-sample predictions and logits, stress-test manifests, public-baseline outputs, manuscript figures, file-level integrity hashes, and scripts for figure generation, statistical recomputation from released derived outputs, release-integrity verification, and maintainer-side sanitized packaging. Version 3.0.0 adds the quality-controlled CMU-MOSEI independent-task replication, including a 21,272-segment anonymized manifest, three-seed test logits for 4,320 segments from 648 source videos, grouped uncertainty recomputation, label-independent exclusion documentation, and refreshed integrity hashes. To respect dataset licenses and third-party redistribution restrictions, the archive does not contain original videos, extracted audio, the official aligned CMU-MOSEI HDF5, reconstructable third-party feature archives, M1–M3 training or inference software, trained checkpoints, third-party checkpoints, or restricted third-party source code. It supports verification and recomputation of the reported analyses from released derived evidence and regeneration of manuscript figures, but not end-to-end retraining from original media.