From exponential extrapolation to regime-aware foresight: A Cognitive Scaling Framework for long-run AI progress

Long-run claims about artificial intelligence progress often extrapolate historical trends while assuming continuity in the structure that generates capability. This paper develops the Cognitive Scaling Framework (CSF), a theory-grounded analytical framework for regime-aware foresight. Connecting technological-regime, path-dependence, complementarity, organizational-capability, and socio-technical transition theories, the CSF distinguishes input-implied technical potential P ( t ) from realized capability I ( t ) . Translation capacity s ( t ) mediates realization, while an effective capacity frontier K ( t ) bounds it. The framework compares conditional trajectories under declared structural configurations. Three controlled experiments combine selected external anchors with scenario assumptions: a historical-potential experiment with a zero-interaction control, matched driver-mix counterfactuals, and common-potential constraint-regime scenarios. Matched driver mixes have identical potential at year 5 and near-identical initial realized capability before diverging, demonstrating regime-conditioned equifinality. Nonlinear factor paths alone generate extrapolation error, with additional divergence under the specified time-varying interaction structure. Realized growth beyond a binding frontier requires expansion of that frontier; productivity or translation improvements alone cannot overcome an unchanged ceiling. The constraint-regime ordering persists across four potential forms and 998 of 1000 joint perturbations; historical residual signs persist across 1000 separate perturbations. The contribution is a transparent technology-management framework for examining structural non-identification in long-run AI foresight, redirecting attention from a single extrapolated slope toward complementary assets, organizational translation, infrastructure, and observable regime-transition signals.