In the papers we reviewed, the two modelling lineages of sleep rarely meet. Mechanistic models simulate inputs such as light or caffeine with other people’s parameters; most learned models estimate or forecast the state and take no input. We define a sleep world model: a probabilistic model of one person’s sleep in which dated events, the sleeper’s actions and assigned interventions drive a latent SleepState of physiological and mental leaves, sleep-regulation models can inform the transition as priors, and instruments and the sleeper observe the night through separate observation models. Of the 62 learned models among the papers placed on our map, 56 only estimate or forecast. Few placed papers score a model’s predicted sleep contrast against an assigned one; in a nap study given in fixed order, a two-process model with parameters fixed beforehand overestimated slow-wave activity after the nap. No paper reviewed compares, on sleep data, a learned transition with and without a mechanistic prior. Assigned emotional or cognitive manipulations shift recorded sleep onset by minutes at group level, but an influence of a mental leaf on one person’s transitions is not shown, and every dose–response relation found is a group average. We propose a capability ladder, an evidence ladder that reserves “effect” for contrasts between inputs assigned at random or otherwise independently of the state, a minimum data protocol, six benchmark tasks and a staged roadmap. Until a person-fitted model’s predicted contrast is scored against that person’s replicated, randomised contrast, its simulated branches remain scenarios.