A diagnostic testbed for selective question-answering (ANSWER / SKIP) over conflicting multi-source personal memory. Each persona has five evidence streams projected from a single latent event table with known, controlled per-source distortions (bias direction, dropout rate, granularity), allowing methods to be measured against the latent ground truth rather than against any single source. The benchmark accompanies the… See the full description on the dataset page: https://huggingface.co/datasets/ytc1997/multisource-membench.