A blog about trust signals, deterministic extraction and paying for data per call. Mostly lessons from watching agents get the right text and draw the wrong conclusion.
Newest first. Every claim about the product is checkable against the live API.
Extraction is a solved commodity. The unsolved half is deciding whether the text that came back deserves to be cited — and that decision is currently being made by a guess.
A walk through the combinations that resolve the majority of trust decisions: wrong page, wrong copy, and assertion dressed up as evidence.
Asking an LLM whether a page is trustworthy produces a different answer tomorrow. If you cannot log the verdict, you cannot audit the decision.
What actually happens between an unpaid call, a 402 with terms, a signed retry and a settled response — with no account and no API key anywhere.
The failure mode is rarely invention. It is citing a real page for a claim the page never made, because nothing checked what the page was.
Subscriptions price the wrong thing. An agent's usage is spiky and unattended — metering it per call removes the negotiation entirely.
Then test one you don't. The difference in the payload is the pitch.