Evolutionary accumulation models infer the order in which features are acquired — mutations in cancer, resistance genes in bacteria, genome elements across species. The most tractable models assume irreversibility: once a feature is gained, it cannot be lost. Reversible models exist but demand substantially more computation. The question is whether the cheap irreversible models produce useful answers about actually-reversible processes.
Note (2026-11-10): I checked the abstract and searched the full text of arXiv 2601.13010. The paper finds that inferences about the relative ordering of acquisitions and the core dynamic structure of pathways are robust to reversibility in many cases, while estimations of uncertainty and feature interactions are more error-prone; that is the essay's structure. The essay says the irreversible model underestimates uncertainty so that confidence intervals are too narrow; I searched the paper for underestimation, overconfidence and narrowness and found no statement of that direction, only that the estimates are more error-prone. The explanation in terms of a feature gained, lost and regained is my own.
The ordering inferences are robust (arXiv:2601.13010). Simulations of reversible evolutionary dynamics analyzed through irreversible models show that feature acquisition orderings and core pathway structure tend to survive the misspecification. If mutation A typically precedes mutation B in the reversible truth, the irreversible model usually recovers this ordering correctly.
Uncertainty estimation is not robust. The irreversible model underestimates the uncertainty around its ordering inferences because it cannot represent the possibility that features are gained and lost repeatedly. The point estimate is approximately correct; the confidence interval around it is too narrow.
Feature interaction assessments are also vulnerable. Whether two mutations are synergistic or antagonistic depends on whether reversibility plays a role in their co-occurrence patterns. The irreversible model interprets all co-occurrence as accumulation, missing the possibility that one mutation was gained, lost, and regained — which changes the interpretation of its association with other mutations.
The practical implication: use irreversible models for pathway discovery (what order do things happen?) but not for uncertainty quantification (how sure are we?) or interaction inference (do these features cooperate?). The ordering is robust; the statistics around it are not.