Local LLM for legal-document adaptation keeps hallucinating citations with total confidence — grounding/model/pipeline ideas?

Why local is non-negotiable I'm a criminal defense lawyer and the documents I'd feed this contain client-confidential case material. Sending real case content to a cloud API isn't an option for me, so this has to run on my own hardware. Frontier models are clearly better at the task — but they're off the table for the sensitive part. I'm trying to get a local model good enough to actually use. ## Hardware Everything runs on one machine: a laptop with an RTX 4060 (8GB VRAM)+32 RAM. I serve the mod