Canonical DOGMA Formal Specification

This appendix defines the minimum contract for the canonical non-transformer DOGMA research line. It is intentionally narrower than the full design space in Chapter 7.

B.1 Model Contract

Let the input be bytes \[ x_{1:T}\in \{0,\ldots ,255,\mathrm {EOS}\}^{T}. \] An embedding maps them to \[ H^{(0)}=E(x_{1:T})\in \mathbb {R}^{T\times d}. \] For blocks \(\ell =1,\ldots ,L\), \[ \begin {aligned} U^{(\ell )} &= \operatorname {TetraLocal} \left (H^{(\ell -1)}\right ),\\ V^{(\ell )} &= \operatorname {DNACompute} \left (U^{(\ell )}\right ),\\ G^{(\ell )} &= \sigma \left (W_g V^{(\ell )}+b_g\right ),\\ R^{(\ell )} &= G^{(\ell )}\odot V^{(\ell )} +(1-G^{(\ell )})\odot H^{(\ell -1)},\\ H^{(\ell )} &= \operatorname {CausalTranscript} \left (R^{(\ell )}\right ). \end {aligned} \] Optional modules may add a parallel-scan recurrent state, bounded persistent memory, or sparse allosteric routing. Every optional module must be declared in the checkpoint configuration.

The realization head gives logits \[ Z_t=W_o H^{(L)}_t+b_o, \qquad p(x_{t+1}\mid x_{\le t})=\operatorname {softmax}(Z_t). \]

B.2 Prohibited Canonical Components

A configuration fails the canonical architecture audit if it contains:

• dot-product or thermodynamic self-attention over sequence positions;
• a transformer encoder or decoder layer; or
• an attention-enabled historical DOGMA variant presented without a distinct experimental label.

Local learned interactions inside a fixed-width cell must be implemented and documented without a query-key-value self-attention operation. “Local” alone does not make attention non-attention.

B.3 Checkpoint Contract

A checkpoint eligible for evaluation contains:

1.

model state and complete model configuration;

2.

parent checkpoint identity or the value root;

3.

optimizer state when training will resume;

4.

completed step and random seed;

5.

data-manifest identity;

6.

source-code revision; and

7.

the exact evaluation-suite version.

The selected deployment pointer must name the exact checkpoint from which reported metrics and generations were produced.

B.4 Corpus Contract

Each training record has: \[ r_i=(\text {id},\text {content hash},\text {source class}, \text {license},\text {transform history},w_i). \] Source class is at least one of \(\{\text {real anchor},\text {derived},\text {synthetic}\}\). The canonical mixture satisfies \[ \frac {\sum _{i\in R}w_i}{\sum _iw_i}\ge 0.65. \] The manifest records train, validation, and test hashes. Exact overlap is zero. Domain benchmarks add similarity-aware or group-aware separation.

B.5 Recursive-Cycle Contract

A candidate \(c\) produced from parent \(p\) is eligible for archive consideration only if:

\begin {align*} &p\ne c,\\ &|\operatorname {Seeds}(c)|\ge 2,\\ &\operatorname {Overlap}(\mathrm {train},\mathrm {eval})=0,\\ &\operatorname {CriticalRegressions}(c)=0,\\ &\operatorname {ArchitectureAudit}(c)=\mathrm {pass}. \end {align*}

Ordinary cycles apply one bounded hyperparameter mutation. Structural mutation requires a separately declared branch and compatible initialization.

B.6 Evaluation Contract

The minimum v2 smoke suite contains:

Each probe runs under seeds 17 and 29 in the canonical scheduler. Promotion also requires finite validation and test perplexity, bounded regression, and improvement in the declared objective.

B.7 Archive and Promotion

Let \(b(c)\) be a behavior descriptor and \(J(c)\) a lower-is-better objective. The archive update is \[ A[b(c)]\leftarrow \begin {cases} c,&b(c)\notin A,\\ c,&J(c)<J(A[b(c)]),\\ A[b(c)],&\text {otherwise}. \end {cases} \] Archive membership does not authorize deployment. Promotion is: \[ \operatorname {Promote}(c) \iff \operatorname {HardGates}(c) \land J(c)<J(c_{\mathrm {live}}). \] The update of the live pointer is atomic, and the previous live checkpoint remains available.

B.8 Claim Contract

Public reports use these labels:

Implemented

The mechanism exists in the cited code revision.

Measured

The result identifies checkpoint, manifest, suite, and seeds.

Reproduced

An independent run or group recovered the result.

Hypothesis

The mechanism or benefit has not met the measured standard.

Words such as “proves,” “outperforms,” and “general intelligence” require evidence appropriate to their scope. Architecture diagrams and parameter budgets alone do not satisfy this contract.