Bayesian inference of the most plausible parameter values during model calibration is influenced by the method used to combine likelihood values from different observation data sets. In the traditional method of combining likelihood values (AND calibration strategy), it is inherently assumed that the model is error-free, and that different data sets are similarly informative for the inference problem. However, practically every model applied to real-world case studies suffers from model-structural errors. Forcing an imperfect model to describe all data sets simultaneously inevitably leads to a compromised solution. As a result, biased and overconfident predictions hinder responsible risk management and any other prediction-based decisions. To overcome this problem, we propose an alternative OR calibration strategy which allows the model to fit distinct data sets, individually. To demonstrate the effect of choosing between the traditional AND and the proposed OR strategy, we present a case study of calibrating a plant phenology model to observations of the maize crop grown in southwestern Germany between 2010 and 2016. We demonstrate that the OR strategy results in conservative but more reliable predictions than the AND strategy when the behaviour of the target prediction does not represent an average of all data sets. Further, an expert knowledge-based combination of AND-OR could be useful; however, selection of representative calibration data sets is not trivial. We expect our proposed strategy to improve the predictive reliability of imperfect, dynamic models in general, by a more realistic formulation of the likelihood function in the “perfect model setting” of Bayesian updating.