Most legal/energy chatbots hallucinate law, mix jurisdictions, or give unsafe advice. This project demonstrates how to build a citation-locked, refusal-safe regulatory AI using real government documents and reproducible pipelines.
Processed regulatory corpora (cleaned TXT files) are not committed.
They are generated from public DERC PDFs using the included pipeline. This repository publishes the pipeline, not the artifacts.
See prompts/system_instructions.md for the full governance rules
used to constrain the model (citation, refusal, safety).