Typed Structural Control of Neural Inference
Context
This entry describes speculative research that extends the formal results of the three published papers (DTS/DMM , PHG , ADM ). It is grounded in the Fidelity framework’s existing …
Research
Formal foundations for heterogeneous native compilation
SpeakEZ Technologies maintains an active research program in formal methods for heterogeneous native compilation. The work develops a compilation framework, Fidelity, and a language, Clef, that carry dimensional type annotations and deterministic memory management properties through multi-stage MLIR lowering to diverse hardware targets: x86/ARM CPUs, FPGAs, NPUs, neuromorphic processors, and (prospectively) quantum backends.
Three papers, published on arXiv in March 2026, establish the formal foundations. Each builds on the preceding work, and the trilogy is supported by a technical blog series on the Clef language site that traces the categorical structure, practical implications, and target-specific compilation strategies in detail.
The entries below summarize each paper’s contributions, describe the current state of supporting implementations in the Fidelity framework, and identify active research threads that extend the formal results toward new domains and hardware targets.
Principled Training for Domain-Specialized Models
Paper
Houston Haynes, “Adaptive Domain Models: Bayesian Evolution, Warm Rotation, and Principled Training for Geometric and Neuromorphic AI,” arXiv:2603.18104, March 2026. [abstract] …
Multi-Way Relational Structure for Spatial and Physics-Aware Compilation
Paper
Houston Haynes, “The Program Hypergraph: Multi-Way Relational Structure for Geometric Algebra, Spatial Compute, and Physics-Aware Compilation,” arXiv:2603.17627, March 2026. …
Dimensional Persistence Through Native Compilation
Paper
Houston Haynes, “Dimensional Type Systems and Deterministic Memory Management: Design-Time Semantic Preservation in Native Compilation,” arXiv:2603.16437, March 2026. [abstract] …