This chapter is the empirical boundary of the book. Earlier editions described a 224.7-million-parameter “DOGMA-Supreme” design and benchmark table as if they were a completed validation. That standard was not met. The material is retained in the companion design studies as a research proposal, but it is not reported here as an achieved result.

The measured baseline available on 27 July 2026 is a 29,830,396-parameter non-transformer byte-level model. It is useful because it exposes a central failure: lower held-out perplexity did not reliably produce readable free generation. That observation changed the training algorithm and evaluation contract.

14.1 Implemented Architecture

The baseline configuration is:

Each block uses an overlapping local TetraMemory representation, a DNA-compute path, a RegulationGate, and a causal transcript mixer. Optional research features include parallel-scan recurrence, bounded epigenetic memory, and sparse allosteric routing. A configuration cannot be called canonical DOGMA if it enables DNA-strand attention, self-attention, or transformer layers.

These names indicate engineering analogies. For example, RegulationGate is a learned software gate; it is not gene regulation occurring in a cell.

14.2 Data Snapshot

The earlier scheduled corpus was unexpectedly small: approximately \(44{,}439\) training bytes, \(58{,}063\) validation bytes, and \(19{,}792\) test bytes in the recorded manifest. The uneven split sizes arose from document-level construction and byte caps rather than a conventional large-corpus benchmark. Three cached reference-genome records provided real sequence anchors; generated architecture text and instruction material supplied most explanatory language.

This is enough for a software smoke test, not enough for claims about general genomic understanding. The revised corpus contract:

1.

adds versioned external teaching and distilled corpora;

2.

labels reference-genome segments separately from derived or synthetic material;

3.

reserves at least 65% of sampling mass for real anchors;

4.

preserves a higher real-anchor share when already present; and

5.

records exact train/evaluation hash overlap.

Future genomic evaluation must use species-aware or homology-aware splits. Otherwise, highly related sequences can leak across exact document boundaries.

14.3 What the Cycles Measured

Table 14.2 records four early descendants and three recent larger candidates. Lower perplexity is better. “Readable generation” is a qualitative audit of the stored sample under the earlier evaluation protocol; it is not a standardized language score.

Cycle 6 remained live under the older, smaller architecture because later candidates did not satisfy the complete promotion path. Cycle 31 is not directly comparable to those early runs: it has a larger configuration and a new ATCG-first evaluation contract. Its substantially lower likelihood numbers still did not rescue free generation. Across seeds 17 and 29, all four prompts failed the canonical trace, explanation, critique, general-language, and answer-diversity gates. Cycle 32 tested a faster local learning step and failed the same gates while slightly worsening likelihood. Cycle 33 then used that failure profile to shorten the supervised genomic trace from 160 to 80 bases. This controlled repair attempt also failed every generation gate and slightly worsened likelihood. None of the three candidates was promoted.

14.4 A Separate Hybrid Result

The native failure motivated a staged experiment rather than a false promotion. A deterministic compiler first emits a 64-base ATCG trace and a short transcript; a LoRA adapter on a 0.6B transformer then realizes natural language. The adapter was distilled from a 4B instruction model using 96 records and passed all four held-out dogma-genomic-language-v3 probes.

This result is reported as a hybrid bridge, not as native DOGMA language ability. The native 29.83M candidate remains rejected. Chapter 13, Section 13.16.4, gives the construction, observed answers, and required ablation.

14.5 Why Perplexity and Generation Disagreed

Teacher forcing evaluates \[ p_\theta (x_t\mid x_{<t}) \] on a true prefix. Free generation evaluates the model on prefixes containing its own sampled outputs. A small local error can move sampling onto a prefix that never appeared in training; subsequent errors compound. This difference is a form of exposure bias.

The byte-level vocabulary makes the failure especially visible. Every byte is a legal class, including bytes that do not form printable UTF-8 text. A model can assign good probability to held-out bytes while still sampling broken text. Genomic runs are easier because the desired alphabet has only four visible symbols.

Other plausible contributors are:

• a corpus too small and heterogeneous for stable prose;
• validation and test splits with different composition;
• a short context window;
• sampling temperature and top-\(k\) sensitivity;
• over-weighted repetitive genomic patterns; and
• an earlier gate that searched only for a canonical DNA regex.

These are hypotheses to test, not excuses. The correct response is to isolate variables and strengthen evaluation.

14.6 The v2 Correction

The next training cycles use the protocol from Chapter 12. The main corrections are:

1.

Separate capabilities. DNA syntax, readable explanation, and self-critique are scored independently.

2.

Repeat generation. Every smoke probe runs under seeds 17 and 29.

3.

Bind metrics to a checkpoint. Promotion creates selected_model.pt pointing to the exact evaluated file.

4.

Use a two-level search. A behavior archive retains diverse stepping stones; a strict objective and hard gates control live promotion.

5.

Mutate one variable. A descendant changes only one bounded scalar unless it is a separately declared architecture experiment.

6.

Protect data provenance. Real anchors keep at least 65% of sampling mass, and exact split overlap must be zero.

The revised objective adds explicit penalties to held-out perplexity: \[ J=P_{\mathrm {test}}+0.25P_{\mathrm {valid}} +1.0I_D+1.5I_E+1.0I_C+1.0I_N, \] where \(I_D,I_E,I_C,I_N\in \{0,1\}\) indicate failures of DNA, explanation, critique, and general-language probes. The implementation also requires prompt-conditioned answer diversity. Every hard gate must pass for promotion.

14.7 Evaluation Program

A credible DOGMA result must graduate through four tiers.

14.7.1 Tier 1: Software Correctness

Unit tests verify shapes, causality, checkpoint loading, deterministic seeds, architecture audits, archive replacement, and data-mixture constraints. Property tests should cover reverse complement handling and finite gradients.

14.7.2 Tier 2: Controlled Sequence Tasks

Synthetic tasks isolate mechanism:

• delayed copy and selective recall;
• motif detection under position shift;
• finite-state languages;
• reverse, complement, and reverse-complement transforms;
• length extrapolation beyond the training range.

These tasks can reveal whether local memory and recurrence actually transmit information. They do not establish biological usefulness.

14.7.3 Tier 3: Genomic Benchmarks

Use frozen tasks with documented licenses and biological splits: promoter or enhancer prediction, splice-site recognition, species classification, variant-effect prediction, and long-range regulatory tasks. Report AUROC, AUPRC, calibration, confidence intervals, throughput, memory, and all preprocessing. Compare with matched-capacity convolutional, recurrent, DNABERT-2, HyenaDNA, Caduceus, and Mamba-style baselines [Ji et al., 2021b, Poli et al., 2023, Gu and Dao, 2023].

14.7.4 Tier 4: Recursive-Learning Evidence

Measure improvement across lineages rather than one checkpoint:

• promotion rate and false-promotion rate;
• retained performance on old tasks;
• archive coverage and diversity;
• improvement per accelerator-hour;
• sensitivity to mutation schedule and random seed;
• synthetic-data fraction and tail retention.

Only after these tiers should the project discuss increasingly general capability. “AGI” is not a benchmark label for a genomic sequence model.

14.8 Near-Term Experiments

The highest-value next experiments are not simply larger models.

1.

Diagnose byte generation. Compare byte vocabulary with a constrained UTF-8/text head and a four-base genomic head.

2.

Matched architecture ablation. Remove TetraMemory, recurrence, and regulation one at a time at constant parameter count.

3.

Long-range tests. Train on lengths up to 256 and evaluate at 512, 1024, and 2048 where the implementation permits.

4.

Archive study. Compare the quality-diversity archive with a single best-so-far lineage using equal training cycles.

5.

Data study. Vary real-anchor sampling among 65%, 80%, and 95%, measuring both genomic tasks and explanation quality.

6.

Baseline study. Match parameters and training bytes against a causal convolutional model, GRU, small Mamba-style model, and transformer.

14.9 Exercises

14.1.

Compute the percentage change in test perplexity from cycle 6 to cycle 14. Why is that not sufficient evidence to promote cycle 14?

14.2.

The validation perplexity is higher than test perplexity in every row. Give three data-split explanations and an experiment to distinguish them.

14.3.

Design a printable UTF-8 validity metric that does not reward empty or repetitive output.

14.4.

Write a matched-compute comparison protocol for DOGMA and a GRU.

14.5.

Explain why reverse-complement augmentation can improve invariance but can also leak information if applied before splitting.

14.6.

Propose confidence intervals for a promoter-classification experiment with organism-level grouping.

14.7.

Design an ablation that tests whether RegulationGate contributes beyond adding an equal number of ordinary parameters.

14.8.

Specify evidence that would falsify the hypothesis that TetraMemory helps length extrapolation.

14.10 Key Takeaways

Suggested Reading

• Gu and Dao [2023] for selective state-space sequence modeling.
• Poli et al. [2023] for long-range genomic sequence modeling.
• Mouret and Clune [2015] for quality-diversity archives.
• Shumailov et al. [2024] for recursive synthetic-data risks.

Part V
Toward AGI