Keeping fact extraction separate from what gets claimed
Abstracts
Abstracts
A crawler agent prototype fetches approved public links and extracts timestamped facts, but it's deliberately kept separate from the layer that turns those facts into user-facing claims. Evidence gets normalized first; assertions come later, and only from what the evidence actually supports. That separation is unglamorous but it's the part that determines whether an AI-assisted profile stays trustworthy over time. Systems that skip this step tend to blur what was observed with what sounds good, which is exactly the failure mode worth designing against.