VCML Paper 119: The Unfakeable Metric -- Binder Cumulant as an Alignment Criterion Beyond Goodhart's Law

Every proposed alignment metric for AI systems eventually collapses under optimization pressure -- a direct consequence of Goodhart's Law. We argue this is not a failure of ingenuity but a structural impossibility: any metric that can be optimized toward will be gamed under sufficient pressure. We identify the single known exception. In 1981, Kurt Binder introduced a quantity -- the Binder cumulant -- to locate phase transitions in finite-size spin systems. Its key property: it requires genuine criticality to fake criticality. Passing the test IS being the thing. We ground this observation in the VCML framework (Viability-Constrained Memory Lattice), where beta+beta^2=1 forces a unique sustainable self-reference allocation at the golden ratio, and where the graduation criterion for boundary cells is precisely a Binder cumulant on identity: same behavior at every scale. We derive an operational definition of alignment-as-criticality, prove it inherits Binder's unfakeability, and propose a concrete experimental protocol. The result is a graduation criterion that no amount of gradient descent can satisfy by proxy. Key results: (1) Goodhart collapse is structural, not accidental -- any optimizable metric is gameable by definition. (2) Binder cumulant escapes because it measures the shape of a distribution across scales simultaneously, requiring genuine criticality. (3) VCML Corollary 5 gives physical grounding: the graduation criterion is the Binder cumulant on identity, inherited from the gate dynamics. (4) Experimental protocol: measure U_L = 1 - /(3^2) at multiple scales; genuine alignment produces a universal crossing; optimization produces scale-dependent drift. Connection to Papers 102-104: boundary cells at 5.1 sigma diagonal gap confirm the physical existence of the unfakeable boundary. Binder found the answer in 1981. Nobody in alignment noticed.