Glossary

Artificial general intelligence. There is no universally accepted test. This book treats AGI as a research target requiring broad, transferable capability, not as a label for the current DOGMA model.

A DOGMA hypothesis in which a small set of selected sites broadcasts non-local context. It is inspired by distant regulatory effects in proteins but implemented as ordinary digital computation.

A collection of reproducible candidate checkpoints retained for scientific diversity. Archive membership is weaker than production promotion.

A learned weighted aggregation over items, usually produced from queries, keys, and values. The canonical DOGMA research line does not use self-attention.

In biology, one of the nucleobases represented by \(A,C,G,T\) in DNA. In DOGMA, “base” may also name a typed digital symbol; the two meanings must not be confused.

A sequence model whose basic vocabulary is encoded bytes rather than words, subwords, codons, or nucleotides.

A trained descendant checkpoint together with its architecture, data manifest, lineage, and evaluation evidence.

The currently governed non-transformer implementation: overlapping local memory, causal transcript computation, regulation, and declared optional recurrence or bounded memory.

The biological framework describing major classes of sequence-information transfer among DNA, RNA, and protein [Crick, 1970]. It is not the claim that information only flows in one everyday-language direction.

A serialized model state, configuration, and associated training metadata.

A three-nucleotide sequence interpreted during biological translation. DOGMA may use codon-inspired groupings digitally; these are not cellular translation.

Average negative log probability assigned to correct targets.

Physical computation using DNA molecules and biochemical reactions. This is distinct from running a DNA-inspired neural network on conventional hardware.

A transformer family trained on DNA sequence representations. It is a baseline for genomic machine learning, not a wet-lab DNA computer.

DNA-Organized Genomic Model Architecture, a MapleAI research family exploring genome-inspired organization and non-transformer sequence computation.

A bounded persistent-state mechanism inspired by biological regulation without sequence change. The analogy does not imply biological epigenetics is being simulated faithfully.

Any path by which held-out examples or their answers influence training, selection, or prompt design in a way not declared by the protocol.

Optimization through variation and selection among candidates. In DOGMA it operates above gradient-based parameter training.

A metric or vector of metrics used to compare candidates. A proxy fitness can be gamed and must not be confused with the complete research goal.

The public nucleotide-sequence database maintained by the U.S. National Center for Biotechnology Information and partners.

The complete hereditary material of an organism. In DOGMA, “prompt genome” or “model genome” is an explicit engineering analogy for a versioned, heritable configuration.

Optimization that changes parameters in the direction of decreasing differentiable loss.

Data excluded from training and used for validation or testing under a declared selection policy.

A separate MapleAI/Hermon domain model line for practical DNA assistance. It must not be confused with DOGMA architecture research.

A data split designed to keep strongly related biological sequences in the same partition, reducing over-optimistic evaluation.

The recorded parent-child relationships among candidate checkpoints and their mutations.

A quality-diversity algorithm that retains the highest-performing candidate in each behavior cell [Mouret and Clune, 2015].

Degradation that can occur when generations of models train recursively on generated approximations that replace or distort the original data distribution [Shumailov et al., 2024].

A declared change from parent to candidate. Ordinary DOGMA cycles mutate one bounded training variable.

An architecture without transformer layers. The canonical DOGMA contract also forbids self-attention; not every non-transformer model makes that stronger choice.

A relation where one candidate is no worse on every objective and better on at least one.

The exponential of average cross-entropy. It measures next-symbol uncertainty under a fixed tokenization and dataset.

An organism’s observable traits. DOGMA uses the term metaphorically for realized model behavior.

A method that trains a population while adapting hyperparameters by replacing weak members with mutated descendants of stronger members [Jaderberg et al., 2017].

The controlled act of making an evaluated checkpoint the live model.

Search that seeks a collection of high-quality but behaviorally different solutions instead of one global solution.

A training record tied to an audited external observation rather than generated solely by a model in the recursive loop.

Sequence computation that updates and carries a state from one position to the next.

A governed process in which a system proposes, trains, and evaluates descendants. It does not mean unrestricted online rewriting.

The sequence obtained by reversing a DNA strand and replacing each base with its complement \(A\leftrightarrow T\), \(C\leftrightarrow G\).

Attention where queries, keys, and values are derived from the same sequence.

A sequence model based on transitions of an internal state. Modern selective state-space models can be input-dependent and efficiently scanned.

Examples generated or transformed by an algorithm rather than directly observed. Provenance matters more than whether an example looks realistic.

DOGMA’s overlapping fixed-width local memory representation. Its claimed benefits require ablation and matched-baseline evidence.

A neural architecture built around attention, residual pathways, and feed-forward blocks [Vaswani et al., 2017a].