This only logs the discarded instructions but I now think we should log all instructions so we have a full set to develop against.

I like it so far except for the "data" dir. Why not just put it in output dir so it's already configurable and exists and this file would be w/ the other output. Just name it like the generated/test/train files with a different prefix. It should be connected to those easily in a ls.

Also not sure about your comment logging everything. The good ones are already parsed and kept in generated. I'd focus on the skipped ones like you did.

I like it so far except for the "data" dir. Why not just put it in output dir so it's already configurable and exists and this file would be w/ the other output.

Yup, that would make infinity more sense, I thought that the generated files were still in the taxonomy directory

Also not sure about your comment logging everything. The good ones are already parsed and kept in generated. I'd focus on the skipped ones like you did.

What I was thinking here is that if we were to try to find a regex that doesn't discard as many instructions we'd want the raw data to test the regex against. The one in generated were parsed and don't include the data as it came from the model. Anyways I'm happy to go with this and we can add more later if it looks like we need it.