RTH-LM: A Fractal Temporal Convolutional Language Model

RTH-LM is a non-transformer language-model architecture based on a Fractal Gated Causal Temporal Convolutional Network (TCN) backbone combined with modular low-rank specialization layers.This paper presents the architecture, training methodology, deployment rationale, and constrained-compute experimental observations of RTH-LM. Rather than claiming parity with frontier-scale Transformer systems, the work investigates how modular specialization and deployment efficiency can be achieved using a persistent shared backbone and interchangeable adapter configurations.The architecture separates:a persistent shared computational substrate (project terminology: Genome),from interchangeable specialization layers (project terminology: Souls).Experimental Soul V2 runs demonstrate observable differentiation between independently trained Text, Code, and Math specialization layers while preserving the same frozen backbone.The paper also introduces Fractal Resonant Optimization (FRO), an experimental optimizer used operationally during rank-512 Soul training on A40 hardware.The project should be considered an early-stage systems and architecture feasibility study rather than a benchmark-complete language-model evaluation.Public model artifacts:https://huggingface.co/RthItalia/Rth-lm-25bhttps://huggingface.co/RthItalia/Rth-lm-code-25bhttps://huggingface.co/RthItalia/Rth-lm-math-25bRepository: https://github.com/rthgit/ZetaGrid