Beyond Multimodal Gain: Sharp Partial Identification of Task-Relevant Latent-State Information

This record contains the preprint PDF and complete LaTeX source for “Beyond Multimodal Gain: Sharp Partial Identification of Task-Relevant Latent-State Information.” The paper develops a sharp partial-identification framework for measuring task-relevant latent-state information beyond an observed report. It provides observable information witnesses, compatibility conditions, sharp lower and upper bounds, sensitivity analysis, and cluster-aware inference, with simulations and empirical applications to NHANES, MELD, and CrossCheck. The companion reproduction package is available at:https://github.com/Jiangwei-Xue/paper-01-beyond-multimodal-gain-reproducibility The package directly recomputes the reported simulations and the NHANES, MELD, and CrossCheck analyses from bundled compact inputs. Reconstructing those inputs requires the public NHANES 2003–2006 PAXRAW, DEMO, PAQ, and 2019 linked-mortality files; MELD transcript and label files together with the published MM-Align train/test sequence features; and the public cleaned CrossCheck daily table. Neural feature extraction from the original MELD audio and video files and cleaning of raw CrossCheck sensor streams are outside the included reconstruction route.