We explore local-first, evidence-grounded AI systems designed to preserve provenance, uncertainty, bounded authority, and human review. Our work studies how capable systems can remain useful without making their claims or actions more powerful than the evidence and controls that support them.