External input · claims · trajectories · LLM outputs Connector · versioned · hashed · replay-bound Adapter · drift · search · audit mappings Governance Core — IMMUTABLE Replay Kernel · Concept Gates · Determinism Scanner · Authority Filters Output ports · report · reproducibility · citations Evolution · peripheral · branch-isolated (mutates)

1

\[ E_{t_1}(x)\equiv_{\text{byte}}E_{t_2}(x)\ \ \forall\,t_1,t_2,\qquad s(x)=H_{\textsf{SHA-256}}\!\big(\operatorname{canon}(E(x))\big),\qquad \sigma=1. \]

2

\[ c:\mathcal{O}\longrightarrow\mathcal{V},\qquad |\mathcal{V}|<\infty,\qquad \mathcal{V}\ \text{frozen at inception}. \]

3

\[ \Phi:\mathcal{T}\to\mathbb{R}^{9},\qquad C=1-\frac{\dim\Phi(\tau)}{\dim\tau}\approx 0.965,\qquad \rho=\frac{S_{\text{after}}}{S_{\text{before}}}\approx 1.06. \]

\[ \Delta C\cdot\Delta S\ \ge\ \kappa. \]

Compression C → Signal preservation S Pareto point (κ) DESi operates here beyond κ: each +C costs −S high S · low C

4

\[ G=\bigwedge_{i=1}^{6} g_i,\qquad \text{Class}=A \iff G=\top. \] \[ \text{replay collapse}\ \succ\ \text{governance failure}\ \succ\ \text{architecture instability}\ \succ\ \text{drift}. \]

g₁ g₂ g₃ g₄ g₅ g₆ AND all six gᵢ Class A any false → Class B–E priority: replay collapse > governance > architecture > drift

5

\[ \texttt{forced}\Rightarrow\texttt{keep},\quad \texttt{critical}\Rightarrow\texttt{keep},\quad \text{density}<0.35\Rightarrow\texttt{prune?};\qquad \text{preservation}=1.0. \]

6

high_dsame_fam28140.500
low_dsame_fam2620.077
high_ddiff_fam60380.633
low_ddiff_fam76100.132

Kant

Popper

Heisenberg

\[ \mathrm{CIH}=H_{\textsf{SHA-256}}\!\big(\mathrm{MH}\,\Vert\,\mathrm{PH}\,\Vert\,\mathrm{RH}\,\Vert\,\text{epoch}\big). \]

append-only · tamper-evident patch P₁#9a1… patch P₂#c4f… patch P₃#7b2… H(prev) H(prev) patch P₄main patch P₄branch · dissent BRANCH — dissent, never deleted
Kantcategories · a-priori audit gate Popperbranch = conjecture & refutation Heisenbergappend-only · no collapse Alexandria Protocol append-only · hash-chained MIVP-bound (CIH)

DESi · groundwork PES · governance layer G · stochastic generator LLM output · claims PES · deterministic evaluator E — read-only S(t) persistent state ① H_norm(p) > τ₃ yes suspend ② JSD(P‖Q) > τ₄ yes fork = BRANCH-on-dissent bounds ambiguity, not truth else ▼ ✓ admit → knowledge graph

1

\[ G:\ \text{stochastic generation}\qquad E:\ \text{deterministic evaluation},\quad E \perp G. \]

2

\[ v_t = E(x_t, S_t),\qquad S_{t+1}=U(S_t,x_t),\qquad S_t \neq \varphi(x_t). \]

3

\[ H_{\mathrm{norm}}(p) = -\frac{1}{\log n}\sum_{i=1}^{n} p_i \log p_i \in [0,1],\qquad H_{\mathrm{norm}}(p) > \tau_3 \ \Rightarrow\ \text{suspend}. \]

4

\[ \mathrm{JSD}(P\Vert Q)=\tfrac12 D_{\mathrm{KL}}(P\Vert M)+\tfrac12 D_{\mathrm{KL}}(Q\Vert M),\quad M=\tfrac12(P+Q),\qquad \mathrm{JSD}>\tau_4 \ \Rightarrow\ \text{fork}. \]

5