On the bispectrum of COBE and WMAP

The COBE-DMR 4-year maps displayed a strong non-Gaussian signal in the ``inter-scale'' components of the bispectrum: their observed values did not display the scatter expected from Gaussian maps. We re-examine this and other suggested non-Gaussian features in the light of WMAP. We find that they all disappear. Given that it was proved that COBE-DMR high noise levels and documented systematics could at most {\it dilute} the observed non-Gaussian features, we conclude that this dataset must have contained non-negligible undocumented systematic errors. It turns out that the culprit is a combination of QuadCube pixelization and data collected during the ``eclipse season''.