Preface

DOGMA asks a narrow but ambitious question: can abstractions from molecular information processing help build useful sequence-learning systems without transformer self-attention? The biological central dogma is not copied into software. Instead, DOGMA studies computational counterparts of persistent genomic state, selective regulation, local binding, recurrent propagation, conditional expression, and population-level selection.

The canonical candidate described here is a byte-level causal sequence model with overlapping local memory, causal convolution, state-space recurrence, bounded epigenetic memory, and sparse allosteric routing. The research program also contains design studies that have not yet entered the canonical implementation. Every chapter therefore distinguishes architecture from analogy and implementation from aspiration.

Prerequisites. Readers should understand introductory machine learning and sequence modeling. Molecular biology is helpful but not required. The companion volume, DNA Computing Foundations, provides a complete undergraduate entry point.

Scientific contract. A candidate cannot approve itself. Recursive learning proposes candidates; an external evaluator freezes tasks, baselines, seeds, retention checks, and safety gates. A disappointing measurement is published rather than rewritten as progress. This separation is the core engineering contribution of the current program.

What this book does not claim. DOGMA has not demonstrated AGI. Biological inspiration does not imply biological fidelity or computational advantage. The measured native baseline remains small and experimental, and its rejected candidates are part of the evidence.

Wenyan Qin
July 2026