Odd magnitudes in ELG selection

On the DES telecon on Monday, 8/11, Stephanie Jouvel presented some results on ELG selection from SVA1 for DES-eBOSS. Most notably was an additional cut on strange DETMODEL - MAG_APER_4 colors...


If you look at the generic population of galaxies with 20<i<23, you find that there's a small tail of galaxies that are outliers in MAG_DETMODEL - MAG_APER_4 (2" aperture):

There are 4% of the galaxies with DET-APER < -1.8. These are a bit of a puzzle.

But it's worse if you do the ELG selection (using detmodel colors)

modest_class==1 && g<23 && g>=20.5 && (g-r)>-0.7 && (g-r)<0.7 && r-z>0 && r-z<2 && (r-z)>0.4*(g-r)+0.4

Then 23% are large outliers. What's up with that?

Note that the fraction of outliers is similar in all bands (griz), and so I don't know if the detmodel colors are way out of whack (that is, say, detmodel r-i may be less affected even if each detmodel r and i are messed up by the same amount).