Live experiment · Designed and built solo
Canon
An agent can write a convincing message using an owner or launch date that stopped being true. I built a gate that checks those assumptions before the message goes out.
Flag it. Don’t silently rewrite it.
An automatic rewrite could corrupt names. I replaced it with field-level corrections and source receipts: the person redrafts, then checks again.
Independent browser demo using synthetic project data. Checks explicit facts in its record.
Try the premise checker

