Projects

A formal framework that distinguishes parametric depth (δ) from schema coherence (σ) as independently necessary dimensions of agent capability. The model provides a phase-structured account of how both variables evolve during training, introduces the delegation gradient (𝒟*) and intersection activation (Ψ) mechanisms, and derives a two-component decay decomposition separating cognitive decay (λ_c) from frontier obsolescence (λ_f). The central falsifiable claim is that schema coherence is formally distinct from depth — its absence explains compositional generalization failure and OOD brittleness that depth-maximizing training regimes cannot account for. Physics-Informed Residual Learning (PIRL) serves as the motivating case study. **Status details:** - **Status**: Active writing · formal paper in progress - **Target venue**: Journal of Artificial Intelligence Research (JAIR) - **Preprint**: In preparation · arXiv cs.AI + cs.LG **Links:** - [Read article](/articles) → - [View paper entry](/papers) →

Physics-structured neural architectures for adaptive control under distribution shift. PIRL embeds a learned physics prior into the residual structure of a neural network, enabling robust transfer to novel dynamics while maintaining the interpretability properties of the physics baseline. PIRL serves as the empirical motivating case for the H-Bar Model: the architecture instantiates schema coherence (σ) as an injected physics prior rather than a learned representation, and the OOD failure modes of data-only baselines illustrate the cost of δ without σ. The variable mapping between H-Bar formalism and PIRL architecture is fully worked out and documented. **Status details:** - **Status**: Active research - **Role in H-Bar**: Motivating case study and worked example throughout the paper **Links:** - [View paper entry](/papers) →