Welcome to the forum, @ncoots!
The naive Bayes classifier in q2-feature-classifier should not exhibit random behavior. The only time it exhibits inconsistencies like this is when reads (or reference sequences) are in mixed orientations. The classifier determines the orientation of reads (relative to the reference database) by polling the first 100 or so seqs. So feeding the classifier a different set of sequences (e.g., unmerged vs. merged datasets) could lead to variable results if the sequences are in different orientations in one run (or both).
The easiest fix may be to just use classify-consensus-vsearch
. That method is not sensitive to read orientation so it is worth taking a look and seeing what you get.
So with all that in mind, I must ask: do you expect to see Cthulu? I only see a little bit of Cthulu there and so it warrants asking: if this pseudo-random behavior is caused by reversing the orientation of reads relative to the reference database, which classification is correct? is Cthulu really there? Maybe classify-consensus-vsearch can help you decide if Cthulu really exists or not.
The truth is out there. Good luck!
:qiime2: