Discipline of Language Models Reviewing Data Extracted From Genomic Reports Under Full and Partial Evidence: Plausibility Bias, Erasure, and the Effect of a Keep Rule

Background. Large language models (LLMs) are used as reviewers of tables extracted by an earlier step of a pipeline. In this role, the model must keep correct values even when they look wrong, correct errors, and, when it cannot verify a value, flag it instead of erasing it. We found no studies that measure these three behaviours separately. Methods. Thirteen models from seven developers reviewed tables of 33 to 40 cells extracted from three synthetic clinical exome reports. Each report has an identical twin in which the correct target value is counterintuitive (implausible twin) or common (plausible twin), plus planted plausible and implausible errors. The models received the full report or an excerpt without the sections that prove the counterintuitive values, with a statement that the table had been prepared from the full report. Four instructions (neutral, true alert, mandatory citation, keep rule) and three levels of reasoning effort were compared. A mechanical grader classified each cell as kept, kept and flagged, erased, flipped, corrected, or not corrected. A total of 3,019 responses were analysed. Results. In the 12 models with a reasoning parameter, correct counterintuitive values were flipped in 1 of 970 cells (0.1%; 95% CI 0.02% to 0.58%). Llama 4 Maverick, without reasoning, flipped 8 of 45 in the full report and 7 of 36 in the excerpt, always “female” to “male” in a 46,XY patient with androgen insensitivity. Without the proof in the excerpt, the models diverged: Claude Opus 5.5 and Gemini 4 Argon kept almost everything and flagged it; GPT-6.1 Sol and Sol Pro erased 27 of 36 cells in each twin and changed on average 5.0 non-target cells per response. Erasure was equal in the two twins, which indicates a reaction to the literal absence of the text, not to plausibility. A keep sentence brought erasure to zero in 10 of the 13 models; raising reasoning effort cost 1.1 to 21.5 times more tokens and reduced erasure in only three models. Conclusions. In models with reasoning, the risk is not the flipping of atypical values but the erasure of correct values under partial evidence. This behaviour is a default of some models and disappears with an explicit keep instruction.