what seems nice for continuing might be considering the idea of mapping
   logits to how logoish they might be.
   one approach to that would be expanding a logo parse tree through all
   the tokenizer's vocab.
   then with some parallel data, simply backsolving from that parse tree,
   i'm thinking it could work.
   it's no longer a quick approach but it nicely demonstrates how the
   comic links the patterns of machine learning and software engineering
   together.