Why fact extraction and published claims should stay separate
Abstracts
Abstracts
One design choice in ProfileScribe: the crawler agent that fetches approved public links never writes directly to a profile. It extracts timestamped facts and normalizes evidence first, then that evidence moves into a separate professional memory before anything becomes a user-facing claim.
The split matters more than it sounds. It gives a review queue something concrete to check before a post goes out, and it keeps drafts traceable back to the source graph instead of resting on inference. For a system meant to represent someone's professional record, that boundary between "what a source says" and "what gets published" is the part worth protecting.
Still early, but the separation is the piece I'd rather get right first.