@jscholes wow! Yeah, I think with my synth work I just got lucky because I had a lot of simpler synth frameworks to go off of, like Brailab, knew enough about self-correcting it for it to not trash things up horribly, but even if this had imported every single working modern add-on I still think it would make those errors. Some of the wrong training data will still "seap in" to the model's tuning and it's just stuff you can't erase or have it ignore when processing the information, I think that's the biggest issue with the custom-GPT stuff. I've seen better results of this with direct API embedding though, we're using that at my work with some tools and it has a lot more strict framework to only process the info from the work docs we've fed it, which is interesting, but it's also done at that API level.