HART was a machine-learning model developed with Cambridge University and used by Durham Constabulary from 2016 to support custody decisions, most notably whether a suspect was suitable for the force's "Checkpoint" diversion programme as an alternative to prosecution. It used around 30 variables, including offending history and, controversially, a postcode-derived variable correlated with the risk factors in the training data.

HART became one of the most closely studied police algorithms in the UK precisely because Durham published academic evaluations of it, in contrast to most policing AI, where independent evaluation is hard to obtain at all. That transparency is also why its weaknesses are unusually well documented: the postcode variable in particular drew sustained criticism as a proxy that could reproduce the geographic patterns of historic policing rather than correct for them.