Antithetical Labs reports that removing all citations from research memos shifted an LLM judge's score by only +0.17, below the judge's 0.27 run-to-run noise. In an experiment with 80 model-written memos and 90 seeded defects, a typed ontology of ten relation checks caught both citation defects at 100%, while the judge missed them. The judge penalized defects that read badly (a thesis-contradicting trade cost 3.7 points) but not those making research untrustworthy.
No score is assigned. Sources and their independence are shown in the citation chain below.