Comment by highfrequency

> many tasks are covered in the training data; but even small perturbations within a covered class of task result in outright failure or reward hacking.

Is this true? Could you cite an example of a simple prompt that GPT 6 / Fable 5.1 consistently bungle?