Deriving verdicts from stored observations…
Bright Data fetches the page. Everything after that is this system: the same four judges the archive uses, run on a page that had never been read before you pressed the button.
Two steps, and the first one is the point: a page can only be collected if somebody wrote its site down first. Everything you enter is checked here for speed and checked again on the server, where it counts.
Pick a source above. Only pages on a listed site can be collected.
The interesting half of a scraping feature is its refusals. Each of these is a property of the request's shape, not a rule somebody remembered to check.
A request carries a source id from a fixed list and an address. The id resolves to a host allowlist compiled into the server, so there is no request the browser can compose that reaches a site nobody listed.
The collector id and the API token are read from server environment and never leave it. The extraction plan travels with the source, not with the request — a caller cannot ask for different selectors.
Every target passes the same SSRF-aware validation the monitored sources do: loopback, RFC1918, link-local, cloud metadata and embedded credentials are all rejected before any fetch.
In replay mode the client refuses a billable call outright, so an unrecorded page fails with the reason rather than being fetched. Live runs are capped at 6 per visitor per 60 seconds.