Pre-registered randomized controlled trials separate prediction from confirmation, yet trials still drift between registration and publication. I ask whether scientists' engagement with the relevant literature predicts which kind of drift occurs and its impact on biomedical science. Drawing on the publications of 468,050 biomedical scientists linked to 30,084 interventional trials with a matched result paper, I compare each trial's background citations against two semantic neighborhoods: the literature the scientist could have cited given the field's collective citation pattern, and the literature they should have cited given the trial's topic. The gap between what is knowingly or unknowingly reachable but not engaged, the Citation Bias Index (CBI), varies sharply across scientists. Trials associated with larger could–should citation gaps report primary-outcome p-values disproportionately concentrated in the borderline-significant interval (0.045 < p ≤0.050), while their published reports remain closely aligned with their registered protocols. Scientists more aligned with the topical literature drift further from registration, publish more novel claims, and accumulate more downstream citations. Two scientific styles emerge: explorers, who let contribution to the field guide where a trial goes, and crusaders, who defend their original hypothesis and nudge findings past conventional thresholds. Literature-based priors thus offer a pre-results signal for distinguishing productive from defensive science.
Jiwoong Choi (Fri,) studied this question.