Build a due-diligence agent that never fakes a fact (capstone)

The capstone agent, and the one where the grounding doctrine you've built all along becomes code. You give it an identifier, a parcel APN, a property address, a company name, and it assembles a structured report from the sources you connect: a red / yellow / green scorecard, each item tied to where it came from. This is the automated version of the grounded-prompt lessons, and it lives or dies on the same rule: it never asserts a fact it can't source, and it never CLEARS the subject. A pulled record is a pointer to verify, not a conclusion. A missing layer is UNKNOWN and becomes a to-do, not a green. And the report routes every flag to the right professional, a surveyor, a title company, an attorney, an engineer, because the agent triages, it does not decide. Two failures to design against: the agent that fills a gap with a confident guess (forbidden, an unsourced line is a bug), and the agent that reads 'no problem found in the pulled data' as 'this is clean' (forbidden, green means 'the source shows no flag', never 'safe to proceed'). Build those bans into the prompt and the output schema, and you have an agent that does an hour of diligence in a minute without ever lying to you, the thing that separates a real tool from a demo.