Our new paper proposes a computational architecture of the pain system - how pain drives a diversity of responses, actions and learning. At its heart it addresses a fundamental question about what pain represents:either i) a sensory-dominant view, where pain reflects an optimal inference of perceived magnitude of a noxious event, or ii) control-dominant view, where pain reflects an optimal control signal for behavioural change? We argue for the control-dominant view, primarily on the basis of evidence from several core categories of endogenous control: modulation by decision conflict, by predictive value, and by informational value; i.e. even though Bayesian / predictive coding models can explain core instances of pain modulation, they can only be part of the solution, and a broader reinforcement learning model can accommodate pain variability more fully. This helps reframe pain as primarily and precisely tuned for learning and behavioural control. So whilst pain may be private, self-intimating, and incorrigible; it may also be precise and computationally objectifiable. This gives some insight into the broad array of brain regions needed to construct the perception of pain, and suggests a wealth of ways in abnormalities in the underlying computational architecture might predispose to chronic pain