The paper described thresholds that turned out to measure the host, not the pad. Replaced with what's actually true — relative grading against a known-good pad, and confirm-before-convict — plus the story of finding it, because a miscalibrated gate that quietly rejects the good ones and passes the bad ones is the cheapest version of the exact failure this architecture exists to survive. What caught it was measuring the 47 pads I already knew were good. Which is to say: what caught the machine was a human who already had taste about the answer. That isn't an aside from the thesis, it's the thesis. Also records the surface bug — the bar was painted over the candidate and the pad still normalized taps against the full height, so I was asking people to judge a clipped instrument. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>