Protocol to separate aleatoric and epistemic uncertainty in instrumented push-forward distributions

Develop and validate a protocol to decompose the push-forward distribution π(q | I) of a quantity of interest q, produced by an instrumented verification-and-validation (V&V) pipeline from a single observation I, into aleatoric and epistemic components by performing repeated extractions and calibrating the perception layer.

Background

The paper defines an instrumented datum that includes a push-forward distribution π(q | I) obtained by propagating uncertainty from extracted parameters through a solver. The authors distinguish irreducible (aleatoric) from reducible (epistemic) uncertainty and note that practical separation requires repeated extraction and calibration of the perception layer.

They emphasize that, in the single-image regime, uncertainty bands are agent self-reports and require calibration against physical measurements, motivating a concrete methodological protocol to separate uncertainty types.