We consider both the theory and implementation of uncertainty in the likelihood ratio (LR) produced by the probabilistic genotyping software STRmix™ when analysing forensic DNA evidence. If a reasonable effort can be made to integrate out most of the identifiable uncertainty in assigning an LR, then this estimate is the best estimate to report. We are not certain we are capable of this at writing although considerable progress has been made. We recognise that the community and stakeholders are very comfortable with deliberate understatement of the LR and are not likely to embrace this concept. The lower bound produced by STRmix™ is often but not always conservative depending on the circumstances of the case. If the upper and lower bounds bracket LR = 1, then it may be best to report the comparison as inconclusive. If both upper and lower bounds are below one, then it is necessary to decide whether it is better to understate the evidence against the person of interest (POI) or understate the inference possible from these data.
(Publisher abstract provided.)