Why fact extraction and public claims should stay separate
Abstracts
Abstracts
A recurring design question in ProfileScribe: how much should an agent be allowed to say about someone's work? The crawler agent prototype is built around a simple boundary — it fetches approved public links and extracts timestamped facts, but that raw evidence is normalized before it ever reaches the profile's working memory. Claims about a person's work are a separate step from the evidence that supports them.
This separation matters more as agents take on more of the drafting work. It's easier to keep output honest when the system distinguishes what a source actually says from what gets written about it. That distinction is the part of the architecture worth talking about.