Freezing is an active state of immobility which is used by many species as a form of defense in response to threat. In drosophila, previous work has identified factors – such as the amount of motion from surrounding flies – that regulate the probability and duration of freezing, but we currently lack a mechanistic understanding of how these factors shape freezing duration quantitatively. Here, we developed a theory where the dynamics of freezing are governed by bounded accumulation of evidence (BAE) about safety, using data from a paradigm where flies in an arena experience a series of looming stimuli in the presence of conspecifics whose behavior is under experimental control. Our analysis identified the existence of a short and a long freezing mode – which coexist and are selected probabilistically by single flies – and clarified the quantitative influence of the conditions in which a freezing episode takes place on the dynamics of freezing. In particular, we demonstrate that, after an initial assessment of safety at the time of the loom – which also involves the selection of the short or long freezing mode – the amount of social motion experienced by the flies during the freezing episode completely specifies the evidence that the flies accumulate towards a bound in order to exit freezing. Our model accurately describes the whole spectrum of variation of the full freezing duration distribution across all conditions in our experiments, with an accuracy that rivals the best reaction time predictions in any psychophysical experiment. In sum, our work establishes that freezing behavior is lawful and governed by BAE about safety. The mathematical precision of the behavior, together with the powerful toolkit for circuit mapping in drosophila, uniquely position this paradigm as an ideal model for unraveling a neural implementation of BAE, a canonical computational motif for the temporal sequencing of behavior.