From Local Predictability to Long-Horizon Actionability

This systematic falsification study tests whether locally informative training-time signals can yield reliable long-horizon improvements in neural-network generalization. It evaluates adaptive interventions from feature-level reliability gating and counterfactual utility through pathway-gradient routing and example-level reweighting, using matched controls and predefined stage gates. Although several signals were measurable and locally predictive, no intervention produced replicated improvement over matched controls. The work therefore distinguishes local predictability and one-step causal benefit from long-horizon actionability, and documents the full experimental progression, invalidated runs, controls, and stopping criterion for a rigorous negative result.