Self-reported evidence is not input validity: a stress test of post hoc Normal-Inverse-Gamma fusion under flagged sensor faults

Preprint (manuscript version 10). Closed-form fusion of Normal-Inverse-Gamma posteriors from several regression sources (MoNIG) weights each source by the evidence it reports. This paper stress-tests that choice on rule- or quality-control-flagged sensor faults in a building network (Intel Berkeley Lab) and an air-quality co-location campaign (SensEURCity). The evidence reported by a source did not track the error of its own output against the reference, and fusion by that evidence was less accurate than the classical scheme of an input-validity gate followed by an unweighted median. Code, plans and results: https://doi.org/10.5281/zenodo.23903050.