Basyirin Amsyar
Physics-informed learning and formal models of agent knowledge development. Currently working on the H-Bar Model and PIRL.
Latest Writing
The Missing Variable in AI Training: Schema Coherence and the H-Bar Model
Two agents, identical benchmark scores, one fails compositionally. The variable responsible is not capacity or depth — it is schema coherence, and current training pipelines have no loss term for it.
Active Research
Research
The H-Bar Model of Knowledge Development: A Formal Curriculum Framework for Schema Coherence and Compositional Generalization in AI Agents
In ProgressCurrent AI agent training pipelines optimize for parametric depth without a formal account of schema coherence — the structured, compositionally valid representation of domain knowledge. We argue this structural omission explains systematic failures in compositional generalization and out-of-distribution robustness that depth-only models cannot account for. We introduce the H-Bar Model, a formal framework that distinguishes parametric depth (δ) from schema coherence (σ) as independently necessary and non-substitutable dimensions of agent capability, and provides a phase-structured developmental account of how both variables evolve during training. The framework introduces the delegation gradient (𝒟*), intersection activation (Ψ), and a two-mechanism decay decomposition separating cognitive decay (λ_c) from frontier obsolescence (λ_f). We apply the framework to physics-informed residual learning as a motivating case study, derive six falsifiable predictions, and provide curriculum design prescriptions derived directly from the phase structure.