tf.keras.losses.BinaryCrossentropy with multiple outputs gives unexpected results when called as a function

Question

When I have a model with multiple outputs and a binary cross-entropy loss (I don\'t get this issue with MSE...), I get different results when I call the loss object versus when