The conformal prediction framework transforms any point predictor into a set predictor with formal guarantees on the coverage of the true value at a chosen level of confidence. An important component of the conformal pipeline is the nonconformity score function which assigns to each observation a measure of "strangeness" in comparison to the previously seen data points. Multiple conformal models are often compared based on their efficiency, usually measured by the average size of the predicted sets, and their informativeness, the number of predicted singletons. As shown in the literature, these two criteria are influenced by the dataset, the performance of the base model and the nonconformity score function. The joint maximisation of these criteria using a well-behaved nonconformity function is desirable. The current work presents the "Penalised Inverse Probability" (PIP) nonconformity function inspired from classical score functions (Hinge Loss and Margin Score). Using some illustrative examples and experimental results on the task of crop and weed image classification in precision agriculture, we show that PIP exhibits precisely the desired behaviour, striking a good balance between informativeness and efficiency.