CellProof

Computer models of how cells respond to damage are only as good as the mechanisms put into them, and the literature moves faster than anyone can update them by hand. Language models can read papers, but they hallucinate, and in biology, with its vast web of interacting pathways, a wrong link is easy to state and hard to spot. In our run, edits taken straight from the AI, with no checker, broke 7 of the model's 10 known behaviours. CellProof adds the checker. Muse Spark reads papers and proposes mechanistic rules, and an exact solver tests each rule against what the model is already known to do. It can say no, and it says why: the solver rejected 14 of 30 edits as contradictions, and when a rejected rule was sent back with the counterexample, Muse Spark repaired 6 of 9. The final model breaks none of the known behaviours. The same loop can take in new papers as they appear, and could be pointed at other models that come with known behaviours to test against. A research tool for prioritising hypotheses for the lab, not medical advice.