A cell can receive radically different fate confidence from models that fit the same data equally well. The unsettling part is that adding more plausible models does not necessarily make any one cell’s fate call more trustworthy. Single-cell measurements can reconstruct continuous paths of cellular change from snapshots, and these methods order individual cells along those paths and calculate likely fates to find points where cells commit to a lineage.

But these pipelines rely on complex chains of steps, each governed by parameters the analyst selects. That means reasonable choices can shape the answer before the biology gets a vote. So the framework evaluates candidate configurations on genes held back from fitting and admits models whose differences are no larger than the variation caused by random seeds.

The resulting certified fate margin is the minimum decision margin over all statistically admissible configurations. In plain terms, it asks how weak the fate call becomes across the models still considered acceptable. The claim is deliberately narrow: a model set can be constructed without ground-truth labels, not that the measure of multiplicity is new.

The key distinction is how the set is drawn when no observed outcome exists to define a usual error measure. Once an admissible set is constructed, downstream analysis follows standard practice. Across parameter settings that are statistically indistinguishable, reported fate confidence for individual cells varies up to 483-fold.

Within any one configuration, nearly every cell receives nearly the same value. That means reported confidence is closer to a property of the configuration than of the cell. One admitted configuration calls 2.5 percent of cells confident, while another calls 100 percent confident, and both are defensible.

But here is the surprise: multiplicity is driven more by structural diversity than by model count. Twenty-four configurations restricted to one kind of procedure expose 3.8 percent of cells, while twenty-four spanning both kinds raise that to 17.3 percent.

Most strikingly, twelve configurations of one algorithm expose 20.0 percent of cells, where twenty-four runs of another expose 3.8 percent. Different reasonable ways of modeling the same data reveal different fragile calls.

The visual gap widens steadily as equally sized sets become more diverse, while the single-method traces remain far apart rather than converging. The certified fate margin was introduced to give the analyst a per-cell reliability statistic that a single fitted model cannot.

But evaluated against simulated ground truth, it does not deliver one. So the central promise fails at the point that matters most: the extra caution does not deliver a reliable per-cell statistic when evaluated against simulation ground truth. Here is why agreement can mislead.

Every admitted configuration forces continuous inputs into discrete terminal states, so each one can divide unbranched data in a similar, arbitrary way. High agreement across the admitted set can therefore reflect a shared bias of the algorithms rather than biological ground truth.

It confirms that an assignment is determined within a model class, not that a real lineage split exists. The framework can still quantify predictive multiplicity without ground-truth labels, and measuring it is worthwhile because the magnitudes are large.

Model diversity drives ambiguity far more than configuration count. The practical lesson is to report how much plausible modeling disagreement exists: FateMultiplicity quantifies predictive multiplicity in unsupervised trajectory inference without ground-truth labels.

That ambiguity is driven far more by model diversity than by configuration count, so the number of runs alone can understate what the analysis reveals. The study finds that model disagreement is worth measuring, because it exposes fragile cell-fate calls.

But taking the weakest answer across plausible models does not improve reliability for an individual cell.