On the other hand if you increase the dataset to the point it no longer adds the plastic to the dish that is not because it got there trough reasoning.

Or if you use AI to start removing the autographs in the training data before givig it another go then that is not because it got there trough reasoning. If you adjust the harness or somehow add rules to the prompt to keep it within the limits of what we know people probably want then the improved result is not because it got there trough reasoning.

Unless you take a fundamentally different approach then underlying ways that caused it to add the autograph or add the plastic or what have you remain same and we know that. There is evidence for that.