I’d bet its from the error correcting model dada2 uses. I haven’t read their paper close enough to fully understand how it works. It’s not going back to the truly raw data (the way that pyronoise error corrected), so they’re error correcting off of the fastq?