I think it's more that software is verifiable and RLVR can optimize in the correct direction. Harder to do with prose.

That is true, but I think within three years that will likely change.

That's not how training works.

Training tries to produce something that scores highly in training evaluations. With one data point, the evaluation is solely how closely the model output resembles the single input text.

Let's say you do that, and the training text is 58,100 tokens long. Let's say you ask the model to produce 58,101 tokens. Will it "reproduce [the] text verbatim"? No, it can't, because of the dissimilar requested length. Something "new" will come out.

It's also entirely possible that no matter how long you train, the model never converges on generating exactly the same output as its training data - you could end up with an average loss value of 0.001 instead of 0.0. It's not a perfectly deterministic process.

You're correct in principle, but in reality even with limited training data real-world models produce something that isn't exactly their training, especially when sampled stochastically. They're biased toward their training data, not forced to it.