CellCanon: A Verifiable World-State Agent for Experiment-Aware Single-Cell Perturbation Modeling

Connecting computational predictions to biological experiments requires a shared representation of predicted responses, supporting evidence, and the effects of new observations. We introduce CellCanon, an experiment-aware agent organized around a typed Cellular World State for each drug–cell condition. A support atlas supplies observed reference responses; a numerical world model constructs the predictive field, and specialized roles attach evidence-linked explanations, prioritize candidate mechanisms, and govern observation-conditioned updates. Provenance and revision authority make the state an inspectable interface between response modeling and follow-up reasoning. On the sampled held-out-drug SciPlex3 track, CellCanon improves DE AUPRC by 91.3% over the strongest evaluated learned predictor, and DE and three-state Macro-F1 by 26.7% and 50.7%, respectively, over their strongest evaluated hard-decision baselines. State reuse requires 513.6-fold fewer language-model tokens and 230.7-fold fewer logical generations than a gene-wise reasoning baseline. In retrospective cross-cell updating, reliability gating reduces the per-drug rate of performance-degrading updates by 69.3% relative to an always-execute policy. These results establish a world-state architecture connecting accurate perturbation modeling, evidence-grounded interpretation, and selective updates from new observations.