>Coding itself is deterministic, but code quality and the product aren't.

Unless you have a definition you can articulate well enough to enforce on pre-commit hooks and at compile time.

I agree quality matters, but the writing is on the wall. It's only a matter of time before the frontier labs are training ultra-long-horizon software engineering tasks using agent teams in RL training loops that help the models essentially figure out what they're currently missing. It's going to happen. I think human judgement is needed today, but not in the limit. It's learnable. The only thing is figuring out the training loop to teach it. That's all.