(NLI: Runcible): What not to do (What we’ve learned) – Don’t try to “pretrain on

(NLI: Runcible):
What not to do (What we’ve learned)
– Don’t try to “pretrain on the books” hoping the model memorizes them. You lose provenance and get drift. Use RAG + span-level supervision instead.
– Don’t ship free-form answers. Make the schema the product.