Determine whether fine-tuning reduces the need for prompted guidance

Determine whether fine-tuning a large language model on the four-form pattern documents reduces the amount of prompted guidance required during AI-assisted software development, while recognizing that such fine-tuning would not by itself justify retiring verification gates.

Background

The paper treats the code-generating LLM as a non-native vernacular builder whose learned defaults tend toward mainstream practices rather than the project’s local architectural tradition. The methodology therefore supplies the local tradition externally through doctrine, templates, prohibitions, and deterministic gates.

Section 8 identifies fine-tuning on the pattern documents as a possible analogue of Alexander’s re-internalization process. The authors explicitly conjecture that fine-tuning would reduce the amount of guidance needed in the prompt, but they maintain that verification would still be necessary because observed compliance cannot establish reliable internalization of the tradition.