Keras model.fit and model.predict return vastly different results on exact same data during binary classification

Question

Using a fixed random seed, I shuffle my training data and generate x_train, x_valid, y_train, y_valid and then call model.fit() with these new spli