When nobody can write the answer

Fine-tuning has a ceiling.

To fine-tune, somebody has to write the ideal answer down. That works when there is one. Asked for the capital of France, the ideal answer is Paris, and a person can type it in a second.

Now ask the model to explain a diagnosis to a frightened patient. To turn down a request without being rude. To summarise a long argument fairly, or to judge how much detail an explanation needs.

There is no single correct reply to any of those. There are better and worse attempts, and the gap between them is something a person notices instantly and struggles to put into words.

So the answer key cannot be written. But the answers can still be judged, and those are not the same problem.