Maple Brain Lab · Evo Trainer
DOGMA
The DNA-Organized Genomic Model Architecture: quality-gated training, evaluation, promotion, and research serving for a new computational direction.
Architecture
Genomic organization as a computational model
DOGMA is a research architecture rather than a fine-tuned Hermon adapter. It explores genomic embeddings, regulation gates, DNA-computing layers, tetrahedral memory, and curriculum evolution as trainable computational structures. Results are published only with their evaluation scope and limitations.
Training Contract
Every descendant must produce evidence
01Build corpus
Versioned genomic-computation data and adaptive curriculum state.
02Mutate once and train
One bounded variable changes in an isolated descendant with recorded lineage.
03Evaluate
Held-out likelihood plus DNA, readable-explanation, and self-critique probes under fixed seeds.
04Archive, promote, or reject
A niche elite may be retained for research; only a global improvement reaches the live service.
Latest Evidence
Rejected by evidence gate
Parameters: 29,830,396 · Validation perplexity: 2.4262 · Gate ceiling: 220 · Updated: 2026-07-31 07:42:45 UTC
Raw rejected generations stay out of the public status feed. A sanitized DNA sample appears only when a candidate passes lineage, data, likelihood, multi-probe, and global-improvement gates.
Archive is not deployment. Quality-diversity search may retain a stepping stone that is strongest in one behavior cell. The live research pointer changes only after every hard gate passes.