Separating fact extraction from the claims built on top of it
Abstracts
Abstracts
A crawler agent prototype fetches approved public links and extracts timestamped facts before anything gets treated as a claim about someone's work. That ordering is deliberate: normalizing evidence first, then reasoning about what it supports, keeps generated copy tied to something checkable rather than to inference stacked on inference.
It's a small architectural choice, but it's the kind of choice that determines whether an AI-assisted profile is trustworthy or just plausible-sounding. Facts get extracted and timestamped; user-facing statements are built separately, on top of that record, not instead of it.