Maple Brain Lab · DOGMA
DOGMA is the DNA-Organized Genomic Model Architecture: a trainable network built from local structural cells, explicit regulation, recurrent mixing, bounded memory, and expression. The architecture is implemented; its ambitious hypotheses remain open research questions.
Canonical Definition
The canonical path contains no transformer block, no dot-product self-attention, and no global token-to-token attention matrix. Optional attention experiments are excluded from canonical DOGMA checkpoints and must be reported as ablations. Sequence communication comes from overlapping local cells, causal convolutions, selective state, sparse allosteric routing, and bounded recurrent memory.
Plain-language animation
Follow one sequence through the implemented native path. The names come from biology; every step below is an ordinary, inspectable tensor operation.
Turn symbols into state
DOGMA starts with bytes, just like a small byte-level language model. Each byte receives a learned symbol vector, an inferred type, a regulatory weight, and a compatibility value. A/T/C/G can expose the state in base-4 form, but the neural computation uses continuous tensors.
Everyday version: Write each input on a richer index card.Scientific verdict
Yes as a falsifiable sequence-model research program. No as a claim that biological metaphor alone produces intelligence, or that DOGMA already works better than transformers.
Convolutional and selective state-space architectures already show that useful language and genomic models do not require dense self-attention. DOGMA belongs in that research family.
Each component is differentiable, trainable, and removable in an ablation. That makes the architecture scientifically testable instead of merely metaphorical.
A/T/C/G encoding is base-4 storage, and a four-position cell is not automatically better because it is called tetra memory. Random-code and ordinary gated-convolution controls must test those choices.
Native DOGMA has not produced reliable language or beaten a matched baseline. The hybrid can answer narrow prompts because a small transformer adapter realizes the prose.
Measured Hybrid Sample
This is a replay of a held-out evaluation, not simulated typing. The bootstrap compiler emits a canonical 64-base trace; a compact realization adapter produces the answer.
Generate a canonical ATCG semantic trace, then explain what canonical means.
Generate a canonical ATCG se
Canonical refers to a standardized, consistent sequence of A/T/C/G in DNA/RNA strands, ensuring uniformity in molecular structure and function. It defines expected patterns in genetic sequences, enabling accurate comparison and analysis across different biological contexts.
Computational Path
These names are computational analogies with explicit operators. They do not assert that a neural network reproduces cellular biology.
Each token becomes a typed base state with symbol, inferred role, regulatory weight, and compatibility potential. This is learned representation, not a claim that text is literal biology.
Overlapping local cells compute pairwise bonds and propagate state through shared faces. The implemented path uses local structured interaction rather than global dot-product self-attention.
Learned gates suppress, preserve, or amplify regions before expression. This creates an explicit control stage inside every block.
Causal convolutions and optional state-space recurrence move information across regions with sub-quadratic sequence mixing.
A bounded recurrent memory records durable context across computation. The research question is whether state can support adaptation without a transformer KV cache.
A bottleneck folds internal state into a phenotype-like representation, then a realization head produces the observable output.
DOGMA and transformers
The useful comparison is not biology versus machinery. It is one set of sequence operators against another under equal data, parameters, training compute, and hardware.
Every token can directly score every earlier token through self-attention.
Fixed local windows, causal convolution, and optional recurrent state propagate information.
Different, both validStandard dense attention builds a quadratic token-pair matrix.
Fixed-window and fixed-state interaction can grow linearly with sequence length.
DOGMA hypothesisStrong content-addressed retrieval and a mature KV-cache ecosystem.
Information must survive repeated propagation or bounded memory.
Transformer leadsDemonstrated across large, diverse corpora and many scales.
Current native checkpoint fails all frozen language probes.
Transformer leadsCan learn DNA tasks; specialist models add tokenization and invariance.
Could encode regulation and locality directly, but lacks evidence of an advantage.
Open experimentMeasured Baseline
The current native baseline has 29,830,396 parameters and no transformer layers. Its test perplexity is 1.1198, yet its free generation failed the ATCG boundary, explanation, critique, general-language, and prompt-diversity probes. The candidate is rejected. Its roughly 134 KB training split is far too small to support a broad language claim, and its much easier test distribution makes the low perplexity unsuitable for comparison with public language models. The passing hybrid above is tracked separately.
Produced a structured DNA-like run and fragments of text, but weak prose.
Lower test perplexity coincided with binary-looking output.
Some structured DNA output appeared; explanatory text remained poor.
Best likelihood in this snapshot, but unreadable generation exposed the old gate.
Lowest likelihood snapshot, but malformed genomic syntax and prompt-invariant fragments failed every generation probe.
A faster local step regressed likelihood slightly and still failed every ATCG-first language probe.
The evidence-directed 160-to-80-base trace mutation did not repair the ATCG boundary or language probes.
More real-data anchors preserved likelihood, but native output remained malformed and failed all frozen language probes.
V2 Correction
Research Hypotheses
Overlapping cells plus recurrent propagation should recover long-range dependencies without dense all-pairs attention.
Falsifier: DOGMA fails length extrapolation after matching parameter count, data, and compute against convolutional, state-space, and transformer baselines.Explicit suppress/amplify gates should improve compositional tasks where only a small subset of the sequence controls the answer.
Falsifier: Removing regulation does not reduce calibrated accuracy, sample efficiency, or attribution quality on preregistered sparse-signal tasks.Separating internal genotype-like state from an expressed phenotype may produce reusable representations across modalities and tasks.
Falsifier: The bottleneck consistently harms transfer or collapses representation diversity relative to a capacity-matched residual head.Epigenetic memory should preserve task-relevant state and adapt over long horizons without unbounded context replay.
Falsifier: State cannot beat a fixed-window baseline on delayed recall, non-stationary rules, and continual-learning tests without unacceptable forgetting.Primary literature
These are foundations and external baselines, not endorsements of DOGMA. They define the standard of evidence for molecular computing, language modeling, long context, genomics, and recursive learning.
The biological information-flow principle that inspired the architecture's vocabulary.
A real molecular computing result. DOGMA is software inspired by such systems, not a wet-lab DNA computer.
The transformer baseline and its global self-attention mechanism.
Evidence that long convolution and data-controlled gating can compete without dense attention.
Selective state-space modeling and hardware-aware linear-time recurrence.
A strong non-attention genomic baseline at single-nucleotide resolution and long context.
A DNA model with bidirectionality and explicit reverse-complement equivariance.
Why recursive DOGMA cycles retain real-data anchors and cannot grade themselves.
Evidence Ledger
Evaluation Program
The first objective is not a giant checkpoint. It is identifying which DOGMA components survive controlled comparison and ablation.
Bits per byte, perplexity, length extrapolation, throughput, peak memory
Transformer, Hyena, Mamba, convolutional LM; matched tokens, parameters, and training FLOPs
Exact match on unseen rule compositions, graph reachability, algorithmic depth
Randomized train/test grammars and longer-than-training sequences
Delayed recall, rule-switch adaptation, backward transfer, forgetting
Frozen-window, recurrent, and external-memory baselines
Reverse-complement consistency, motif recovery, species-held-out transfer
DNABERT-2, HyenaDNA, and Caduceus where task and licensing permit
