The Third Axis Is Weakness

Neural networks can fit the same examples and still disagree about the world. The unresolved question is which correct network will survive unseen demands. Stack Theory ranks correct policies by weakness, the number of further commitments that remain compatible with correctness. Under exchangeable demands, the weakest correct policy has the highest generalisation probability. Independent work now reports a third pretraining axis beyond parameters and data, an exploration count setting how many outputs a generative model can commit to at once. I read that axis as weakness. Here I test whether completion-based neural measurements follow that ordering in two stages. First, I construct 480 controlled child-correct proto-policies across ten image, text, and sequence benchmarks. Their score-gap extension surface correlates positively with parent accuracy on all ten, with pooled within-benchmark Spearman $\rho=0.725$. Choosing the highest-scoring policy beats its cohort mean on all ten, by $0.197$ accuracy on average. This stage is an engineered positive control for the measurement. Second, I apply a prospectively sealed local proxy to 96 ordinarily trained text classifiers before audit labels are opened. It correlates with audit accuracy at $\rho=0.770$ on binary 20 Newsgroups and $\rho=0.636$ on AG News, with 95\% bootstrap intervals excluding zero. Twelve of fourteen comparison measures reverse weak point-estimate signs between the two trained-network benchmarks. Validation accuracy is stronger but reads 128 held-out labels. The proxy instead reads 2,000 target-domain inputs without their labels. Both stages use finite Monte Carlo tests of affine-head completions rather than exact formal weakness. The trained-network result is correlational, transductive, and coordinate dependent. Together the stages establish construct validity across ten controlled families and a sealed association in trained networks. The causal question remains for proxy-guided search.