We help use the NMPDR RAST servers and also run our own for sequence annotation. I'd have to check but if I'm remembering correctly our annotations fix errors (~%5) present in the NCBI databases and catches many more (something like ~20% of unknowns) for the organisms we work with. Problem is searching our indexes is super slow (comparatively) and not all in one place -- BLASTn/p takes some fiddling.

Could we use our own corpus for this?

Ah, I see you have some problems with species identification...are you using old NCBI datasets?

(ps. If you'd like to talk, you can email me at the address in my profile)