A cancer cell treated with chemotherapy will eventually die by apoptosis or escape by dividing. These two fates look identical under a microscope for most of the cell's remaining life. The morphological differences that distinguish them are subtle and transient — a slight change in shape, a flicker of fluorescence.
A transformer model trained on live-cell video (arXiv:2603.16562) predicts which fate a cell will choose with 94% accuracy — and can make the prediction up to 10 hours before the event occurs. The attention maps reveal when the model looks: for apoptosis, the signal is distributed across the trajectory; for mitosis, it concentrates in specific windows.
Note (2026-11-08): the abstract (Bürger et al.) reports a balanced accuracy of 0.94 and an F1-score of 0.93 from video alone, that the predictive signal is not located only in the final frames, with reliable predictions possible up to 10 hours before the event, distinct temporal distributions of predictive information for mitotic and apoptotic sequences, and a role for cell morphology and p53 signaling. It does not say which fate has the distributed signal, that the cell has decided 10 hours before, or that a human watching would see nothing; the essay adds those, and omits p53.
The through-claim: the cell has decided long before it acts, and the decision is written in its shape. Not as a single moment of commitment but as a pattern of morphological changes that the cell accumulates over hours. A human watching the same video sees nothing — the differences are below perceptual threshold. But they're above the statistical threshold that a model trained on thousands of trajectories can detect.
This separates two things usually conflated: when a decision becomes detectable and when it becomes irreversible. The cell may not be committed at the 10-hour mark. But the trajectory it's on is already distinguishable from the alternative. The fate is readable before it's sealed.