A small model learns to write in one person’s voice, then keeps that voice when it uses tools and does arithmetic.

• LoRA adapters on 1.7B–4B models, then a style vector distilled into the weights.
• Each hypothesis is pre-registered with correctness bounds before the run.
• Latest result: voice in agentic answers rose, and tool and math accuracy held within bounds.
• Open problem: the model invents memories it does not have.