> For the complete documentation index, see [llms.txt](https://osintelligence-llc.gitbook.io/osintelligence/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://osintelligence-llc.gitbook.io/osintelligence/part-iv-the-evidence-what-worked/16-sovereign-domain-pruning.md).

# 16 · Sovereign Domain Pruning

Domain pruning at 35B under seal: the four-file integrity contract, with the companion validation that kept the negatives.

> **A companion paper.** This report is the architecture leg of the series' compression evidence: where *Sovereign In-Distribution Imatrix Calibration* (Chapter 15) used the sovereign corpus to produce a better quantization, this paper uses it to produce a better architecture: cutting the experts a model never uses on the sovereign workload so a 35B-class MoE fits consumer VRAM. Cited in-series by title. The main body (§1–§9) is the sealed Phase-2 result; Appendix §10 is the Phase-B companion validation, reported in full including its two honest FAIL verdicts.
>
> **Status note.** Body preserved as sealed (Phase-2 γ.2 re-seal 2026-04-25; Phase-B γ-seal 2026-04-30). The source is a late scaffold: the numerical content and discussion prose are sealed; the title, byline, and reference list were assembled at conversion from in-source material and are flagged in the audit ledger for author review.

> **What is new here.** The contribution is a destructive expert-prune sealed as a fully falsifiable result: 197 of 10,240 experts cut from a 35B-A3B mixture-of-experts against the operator's own CTI corpus, with a pre-registered mandatory long-form-reasoning gate deliberately anchored against the two published results (layer-pruning chain-of-thought collapse; super-expert removal repetition-collapse) most likely to sink it. The prune passed at a 0.9826 long-form ratio and a slightly-better-than-baseline perplexity, but the load-bearing novelty is the honesty architecture around the number: a companion validation whose own random-ablation control mechanically FAILED, revealing that at this prune fraction the verdict is robust to which experts are chosen (HumanEval byte-identical across two disjoint 197-expert sets), and whose precision-fidelity audit FAILED its rank-correlation axis while passing its limits-of-agreement axis, both reported as falsifications without rescue by redefinition.
>
> **Deepest water.** §5.3, the H6 long-form ratio that does not reproduce Wang's layer-pruning degradation signature on expert-only pruning; §10.4, the random-197-ablation control that failed the discriminative-screen hypothesis and thereby proved the prune robust to selection method; and §10.6, companion-validation-as-inferential-strategy, two ancillary FAILs that bound the primary claim without rescuing it while the Phase-2 GREEN stayed unfalsified across all three adversarial lenses.

### 1. Abstract

We sealed destructive expert-pruning of 197 of 10,240 experts (\~1.9%) at blocks {20, 21, 25} in Qwen3.5-35B-A3B against a sovereign cyber-threat-intelligence (CTI) calibration corpus, with a cryptographically pre-registered eval-bank Gate-B verdict GREEN. The sovereign held-out perplexity primary (H5) is preserved at Δppl = −0.0519 (−0.41% relative, 2.4× under the 1%-relative threshold; pruned arm slightly *better* than baseline). The long-form test-time-scaling mandatory (H6, anchored against Wang et al.'s layer-pruning chain-of-thought degradation result) lands at aggregate pass-rate ratio 0.9826 (decisive GREEN at the ≥ 0.90 floor; AMENDMENT\_4 ratified a symmetric scope reduction to 30 cells / 1,100 generations per arm after a gated-dataset authentication failure during execution). Combined Phase-2 wall-clock ≈ 68 h 20 m (γ.2-sealed; 0 errors) on consumer-class hardware: NVIDIA RTX 5070 12 GB Blackwell sm\_120 + Intel i7-14700F 20C/28T + 128 GB DDR5 + CUDA 12.8 + torch 2.10.0+cu128 + llama.cpp commit 822047a0a. Evaluation uses a pre-registered mixed-backend split: HF-bf16-in-child forward-pass for sovereign perplexity + MMLU; llama.cpp Q6\_K llama-server (production-matched) for GSM8K + HumanEval + long-form TTS. Cryptographic integrity is operationalised as a four-file γ.1 + γ.2 re-seal with a cascade-level manifest (92 files). Phase-B companion validation subsequently closed all three of its audit surfaces (Appendix §10): two ancillary-axis hypotheses failed as pre-registered and are reported as falsifications with substantive characterization; the scope-recovery hypothesis passed with margin; and the Phase-2 GREEN verdict emerged unfalsified across all three adversarial audit lenses.

### 2. Introduction

#### 2.1 Motivation

The Orchestrator 9B (Qwen3.5-9B Q8\_0) is a 12 GB-VRAM compromise. The Architect (Qwen3.5-35B-A3B MoE Q6\_K, 27 GB) has 4× the parameters and demonstrably stronger reasoning, but does not fit as primary. **Pruning breaks this ceiling:** surgical removal of experts that are inactive on the sovereign CTI workload produces a model that started with 35B knowledge, was trimmed to only what matters for CTI/OSINT/dossier generation, fits in 12 GB, and is provably better at the sovereign domain than the unpruned 9B because it began with more relevant knowledge. This is the imatrix-calibration philosophy extended: the companion paper used sovereign calibration data to produce better quantization; this paper uses it to produce better architecture.

#### 2.2 Threat model and falsification bar

Wang et al. (arXiv:2510.22228, *"When Fewer Layers Break More Chains"*) establishes that layer-pruned models can lose test-time-scaling chain-of-thought quality even when perplexity preserves. Su et al. (arXiv:2507.23279, *"Unveiling Super Experts in MoE LLMs"*) establishes that pruning 3 of 6,144 experts in Qwen3-30B-A3B causes catastrophic repetition-collapse on Math-500. Phase-2 design preempts both threats: H6 (long-form TTS pass-rate ratio ≥ 0.90 GREEN / ≥ 0.80 YELLOW) is MANDATORY-falsifiable, and a pre-registered super-expert validation gate cross-referenced Su et al.'s signature against our prune set: 0 of 15 super-expert candidates appear in our 197-expert removal set across blocks {20, 21, 25} (GREEN).

#### 2.3 Contributions

1. Cryptographically pre-registered MoE expert-pruning + eval-bank Gate-B verdict on consumer 12 GB Blackwell sm\_120.
2. Three peer-review-novel substrate measurements: HF-bf16 sampling 0.35 tok/s ↔ HF-bf16 greedy 2.56 tok/s ↔ HF-bf16 forward-pass argmax 8.91 s/row ↔ Q6\_K llama-server greedy 26.14 tok/s on RTX 5070 12 GB sm\_120 + CUDA 12.8.
3. **Mixed-backend eval-bank under shared pre-registration:** HF-bf16-in-child for sovereign + MMLU forward-pass probes; llama.cpp Q6\_K llama-server for GSM8K + HumanEval + long-form-TTS generation probes; documented in the amendment deviation log and machine-readable citation envelope.
4. **Hash-first integrity discipline:** a three-file γ.1 co-seal pre-emptively executed before any aggregator-bug-fix code edit (operator directive: *"Hash first, don't contaminate"*); the γ.2 four-file re-seal preserves the three γ.1 anchors unchanged while sealing the new amendment anchor.
5. **Document-as-found scope-reduction (AMENDMENT\_4):** an execution-time gated-dataset authentication failure triggered a symmetric long-form scope reduction (30/45 cells × 1,100 generations per arm); preserved arms byte-symmetric; the H6 ratio is decisive GREEN at reduced scope, outside the CI-straddle band.
6. **Super-expert cross-reference GREEN:** the pre-registered Su et al. signature (top-5 per block by max router-gate logit) cross-referenced post-implementation against our 197-expert prune set: 0/15 hits, the four-condition conjoint screen demonstrably spared the super-expert cluster.

### 3. Methods

#### 3.1 Substrate

| Component      | Value                                                                                                                                        |
| -------------- | -------------------------------------------------------------------------------------------------------------------------------------------- |
| GPU            | NVIDIA RTX 5070 12 GB (Blackwell sm\_120, compute capability 12.0)                                                                           |
| CPU / RAM / OS | Intel i7-14700F 20C/28T · 128 GB DDR5 · Windows 11 Pro                                                                                       |
| CUDA / torch   | 12.8 (build target; system PATH stays on 13.1) · torch 2.10.0+cu128                                                                          |
| llama.cpp      | commit 822047a0a (b7992 + 3 patches: M-RoPE #20093 + K-shift PR #19928 + Qwen3.5-MoE handler): the patched binary of the mRoPE RCA companion |

#### 3.2 Pruning policy

* **Base model:** Qwen3.5-35B-A3B (256 experts top-8 routed + 1 shared expert × 40 layers; multimodal wrapper, text-path-only profiled).
* **Pre-registered conjoint screen:** Block-Importance p05 < 0.05 + expert frequency < 0.5% + Gini > 0.55 + max\_activation ≥ 0.4 EXCLUSION.
* **Pruned blocks:** {20, 21, 25}. **Experts removed:** 72 + 68 + 57 = 197 of 10,240 (\~1.9%).
* **Mechanism:** router rows masked to −10000.0; shared-expert pre/post hash byte-identical (preservation invariant verified in the checkpoint manifest).
* **Super-expert validation:** Su et al. signature; baseline forward-pass on 50 sovereign windows × 512 tokens; hooks on the MoE gate for blocks {20, 21, 25}; top-5 per block by max router-gate logit; **0 / 15 candidates in our prune set** (GREEN).

#### 3.3 Eval-bank (Gate-B)

| Probe                      | Backend                                                               | N                                                                                                             | Weight |
| -------------------------- | --------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------- | ------ |
| Sovereign held-out PPL     | HF-bf16 forward-pass, stride 2048                                     | 1,251 windows / 2,560,797 tokens                                                                              | 0.20   |
| MMLU 5-shot                | HF-bf16 forward-pass argmax                                           | 1,995 (57 subj × 35 rows; 70%-power @ 3-pp Δ)                                                                 | 0.10   |
| GSM8K 5-shot strict + flex | Q6\_K llama-server greedy                                             | 1,319 (full)                                                                                                  | 0.15   |
| HumanEval pass\@1          | Q6\_K llama-server greedy + sandboxed exec                            | 164 (full)                                                                                                    | 0.10   |
| Long-form TTS stratified   | Q6\_K llama-server sampled T=1.0 (Wang verbatim)                      | AMENDMENT\_4 reduced: AIME24 30 + MATH-500 50 = 80 problems × 5 budgets × 3 seeds = 30 cells / 1,100 gens/arm | 0.45   |
| MATH standalone            | DROPPED per AMENDMENT\_2 (duplicative with MATH-500 inside long-form) | 0.00 (was 0.10)                                                                                               |        |

**Backend split rationale:** HF-bf16-in-child is viable for forward-pass probes (no autoregressive decoding); the Q6\_K llama-server is viable for autoregressive generation (\~10× faster than HF-bf16 greedy on this substrate; \~75× faster than HF-bf16 sampling).

#### 3.4 Pre-registration and amendments (γ.1 + γ.2 four-file integrity)

| File                                                            | SHA-256 (leading) | Seal moment                       |
| --------------------------------------------------------------- | ----------------- | --------------------------------- |
| PRE\_REGISTRATION.md                                            | 37a90586…         | γ.1 2026-04-23 (UNCHANGED at γ.2) |
| AMENDMENT\_2 (drop MATH + reweight 0.45)                        | 7b74ac41…         | γ.1 2026-04-23 (UNCHANGED at γ.2) |
| AMENDMENT\_3 (sample caps + backend split + super-expert GREEN) | 94f675c7…         | γ.1 2026-04-23 (UNCHANGED at γ.2) |
| AMENDMENT\_4 (symmetric long-form scope reduction)              | d26d9816…         | γ.2 2026-04-25 (NEW anchor)       |

***Table 3.** The four-file integrity contract. Cascade-level index: manifest.sha256 ad2fd10a… (92 files).*

### 4. Phase-1 Results (Gate-A GREEN, sealed 2026-04-20)

| Metric                         | Value                                     | Threshold                | Verdict  |
| ------------------------------ | ----------------------------------------- | ------------------------ | -------- |
| Wall-clock                     | 95,550.8 s (26.54 h)                      | 108,000 s                | pass     |
| Rows OK                        | 8,358 / 8,358 (2,591,566 tokens profiled) | full                     | pass     |
| BI p05 < 0.05 blocks           | 38 / 40                                   | ≥ 1                      | pass ×38 |
| Experts < 1% activation layers | 36 / 40                                   | ≥ 1                      | pass ×36 |
| Endpoint preservation          | blocks {0, 39} non-pruneable              | per Men et al. 2024      | pass     |
| VRAM peak / host RSS / NaN-Inf | 10.20 GiB / 66.5 GiB / 0                  | < 12 GiB / < 128 GiB / 0 | pass     |

***Table 4.** Phase-1 activation profiling, all gates green.*

### 5. Phase-2 Results (Gate-B GREEN, γ.2 four-file re-seal 2026-04-25)

#### 5.1 Verdict

**GREEN.** Destructive expert-pruning of 197 of 10,240 experts (\~1.9%) across blocks {20, 21, 25} in Qwen3.5-35B-A3B preserves short-form capability AND preserves long-form reasoning at AMENDMENT\_4 reduced scope.

#### 5.2 H5 primary: sovereign held-out perplexity

| Measurement                                  | Baseline | Pruned      | Δ relative | Verdict                               |
| -------------------------------------------- | -------- | ----------- | ---------- | ------------------------------------- |
| Sovereign PPL (1,251 windows / 2.56M tokens) | 12.6418  | **12.5899** | **−0.41%** | PASS by 2.4× (pruned slightly better) |

#### 5.3 H6 mandatory: long-form test-time scaling

| Task                    | Baseline                                                                                                       | Pruned     | Ratio              | N gens/arm |
| ----------------------- | -------------------------------------------------------------------------------------------------------------- | ---------- | ------------------ | ---------- |
| AIME24                  | 0.4467                                                                                                         | 0.4222     | 0.9453             | 450        |
| MATH-500                | 0.9573                                                                                                         | 0.9573     | **1.0000**         | 750        |
| GPQA-diamond            | SKIPPED at Phase-2 (AMENDMENT\_4 symmetric; gated-auth fail): recovered at Phase-B Component C, Appendix §10.5 |            |                    |            |
| **Aggregate long-form** | **0.7020**                                                                                                     | **0.6898** | **0.9826 · GREEN** | 1,100      |

***Table 6.** H6 decisive GREEN (≥ 0.90 floor; outside the \[0.85, 0.95] CI-straddle band).*

**Per-budget aggregate (test-time-scaling signature):** parity-or-better at middle budgets (1024–4096 with cross-arm Δ ≤ +0.0044); mild degradation at 512 (under-budget; −0.0444) and 8192 (over-budget; −0.0256). **Wang's TTS-degradation prediction does not directly transfer to expert-only pruning at our 1.9% prune rate.**

#### 5.4 Secondary short-form probes

| Probe                        | Baseline | Pruned     | Δ           | Notes                                                                                        |
| ---------------------------- | -------- | ---------- | ----------- | -------------------------------------------------------------------------------------------- |
| MMLU 5-shot (1,995 cap)      | 0.8521   | 0.8396     | −0.0125     | Within the 3-pp power floor                                                                  |
| GSM8K strict (1,319 full)    | 0.8840   | **0.9014** | **+0.0174** | **Pruned BETTER** (flex consistent: +0.0121)                                                 |
| HumanEval pass\@1 (164 full) | 0.7561   | 0.7012     | −0.0549     | Largest secondary regression; marginal at N=164 (95% CI ±\~7 pp); Phase-B primary audit case |

**Weighted aggregate:** baseline 0.7617 → pruned 0.7496; **weighted Δ accuracy −1.20 pp** (modest).

#### 5.5 Substrate envelope (four-point, peer-review novel)

| Regime                                         | tok/s          | Notes                                    |
| ---------------------------------------------- | -------------- | ---------------------------------------- |
| HF-bf16 sampling (T=1.0)                       | **< 0.35**     | PCIe-bound on 12 GB Blackwell            |
| HF-bf16 greedy                                 | **2.56**       | autoregressive 50-tok smoke              |
| HF-bf16 forward-pass argmax                    | **8.91 s/row** | viable for letter-argmax + perplexity    |
| Q6\_K llama-server greedy (production-matched) | **26.14**      | \~75× vs HF-sampling; \~10× vs HF-greedy |

**Wall-clock envelope:** Phase A-1 (sovereign + MMLU + GSM8K + HumanEval) 29 h 13 m (0 errors); Phase A-2 (long-form Q6\_K stratified) 39 h 02 m (0 errors); **combined ≈ 68 h 20 m (γ.2-sealed), 0 errors**, 28.5% of the operator's 240 h ceiling.

### 6. Discussion

#### 6.1 Hash-first integrity discipline as a peer-review-anchored standing rule

Phase-2's Gate-B aggregator emitted a spurious-RED verdict in its first version due to a schema-drift bug (a field-name mismatch between the aggregator and the per-arm artifacts). The bug was discovered post-Phase-A-1-landing, before Phase-A-2 had executed the pruned-arm long-form probes, that is, in a window where a code edit was about to be applied to a file whose output would inform the eventual H5/H6 verdict. The operator directive, *"Hash first, don't contaminate the experiment"*, triggered a three-file γ.1 co-seal of the pre-registration body + AMENDMENT\_2 + AMENDMENT\_3 in approximately 30 s, BEFORE any aggregator-fix code edit landed. This sequencing eliminates the post-hoc-amendment attack surface (Kargaran's ICLR rejection-ground analysis identifies pre-registration backflow as Rejection Ground #2): the edit can no longer be timestamped before the cryptographic seal, so the body of the pre-registration cannot have been retroactively shaped by the bug-fix outcome. The v1 spurious-RED artifacts are preserved as *\_v1\_buggy.*; the v2 INCONCLUSIVE artifacts (post-fix, pre-A-2-landing) are preserved as *\_v2\_pre\_A2.*; the v3 final at γ.2 four-file re-seal is the sealed public record.

#### 6.2 Document-as-found symmetric fail-soft under arms-symmetric construction

A Hugging Face Hub-side gated-authentication failure on the GPQA dataset (no token env-var; cached token absent) was caught by the long-form probe's pre-existing per-task fail-soft wrapper. Both arms inherited the same wrapper structurally; both arms therefore ran symmetrically at 30 cells × 1,100 generations (vs the pre-registered 45 cells × 1,650). Three pre-specified equalities were verified at pruned-arm landing; 3/3 hold; arm symmetry decisive. AMENDMENT\_4 ratified the scope reduction at the γ.2 four-file re-seal. The Wilson 95% CI widening factor is approximately 1.22× at the reduced N: the observed H6 ratio 0.9826 lands outside the \[0.85, 0.95] CI-straddle band, so the GREEN verdict remains decisive at the reduced scope. This deviation-handling pattern is grounded in HELM, BIG-bench, BenchBench, Bean et al., and Kargaran document-and-move-on precedent.

#### 6.3 Test-time-scaling preserved under expert-only pruning at our 1.9% prune rate

Wang et al. establishes that **layer-pruned** models can lose chain-of-thought test-time-scaling quality even when next-token perplexity is preserved. The Phase-2 H6 result (aggregate ratio 0.9826; per-budget Δ ≤ +0.0044 at the middle 1,024–4,096-token TTS budgets) does not reproduce Wang's degradation signature on this prune-class at this rate. The intervention class differs in axis: Wang prunes whole layers (depth axis); we prune experts within MoE FFN sub-layers (per-token capacity axis), preserving the full residual stream and the attention path on every block. The stronger generalisation (*expert-only pruning at small percentages preserves TTS where layer-pruning at comparable parameter savings does not*) requires the Phase-B random-expert-ablation arm as a discriminative control. If random expert removal also preserves TTS, our screen is not load-bearing for the verdict; if random ablation degrades TTS, our conjoint screen is empirically discriminative. This is the primary auditing question Phase-B was launched to answer (the answer, delivered in Appendix §10.4, is honest and surprising).

#### 6.4 HumanEval secondary regression as the marginal residual

The Δ pass\@1 on HumanEval (164 full) is −5.49 pp (ratio 0.9274). At N=164 the 95% Wilson CI on the per-arm rates is approximately ±7 pp, so the regression is within one CI-half-width of zero. Three response paths were pre-committed: (a) Phase-B Component A (bf16 cross-validation) audits whether the regression is a quantization artefact rather than a pruning artefact; (b) Phase-B Component B (random-expert ablation) audits whether the regression is screen-discriminative or an inevitable consequence of removing 197 experts at any selection; (c) future multi-seed runs would harden the N=164 statistical floor. The HumanEval regression is therefore flagged as a Phase-B audit primary test case rather than a Phase-2 verdict-relevant signal.

#### 6.5 Mixed-backend eval-bank under shared pre-registration

The eval-bank backend split is a pre-registered, single-pre-reg-document evaluation-engineering choice documented in the amendment deviation log, the machine-readable citation envelope, and the decision ledger. All three surfaces declare: hypotheses unchanged, thresholds unchanged, weights unchanged, aggregate scoring unchanged. The amendment is engine-only. The choice is empirically justified by the four-point substrate envelope in §5.5: at 35B-A3B-class on this hardware, HF-bf16 sampling at 0.35 tok/s would consume approximately 340 h alone for one arm of the verbatim Wang grid, infeasible at the operator's 240 h ceiling. Backend partitioning is the only path that fits, grounded in HELM + StratPPI + BenchBench precedent for pre-committed engine architecture as a peer-review-clean evaluation-engineering decision.

#### 6.6 Substrate-matching production-deployment posture

The Q6\_K configuration used for GSM8K + HumanEval + long-form TTS is byte-identical to the production architect preset in the live OS-INTelligence deployment. The eval-bank backend split is therefore **backend-faithful for the Q6\_K path**: what is being evaluated is exactly what is being deployed. The forward-pass HF-bf16 probes do *not* reflect production deployment; they are evaluation-only paths chosen for substrate feasibility. The Q6\_K↔bf16 paired-agreement audit is the closure protocol for the precision-gap question, deferred to Phase-B Component A. The substrate-matching recommendation for similar consumer-12 GB MoE deployments: Q6\_K with CPU-MoE offload for autoregressive generation; HF-bf16-in-child viable only for forward-pass workloads at ≈ 10 s/row latency.

#### 6.7 Three-axis substrate telemetry as the publishable envelope

Phase-2 emitted host (psutil timeline) ⊕ GPU (nvidia-smi dmon 1 Hz) ⊕ engine (llama.cpp Prometheus metrics + slots) telemetry on every arm, plus a one-shot observables snapshot per arm (router-entropy, router-L2-norm, expert-top-k, final-layer logits-top-k, NaN/Inf flags, attention-entropy per block). The discriminative power of three-axis telemetry over single-axis monitoring is the ability to distinguish **memory-bandwidth-bound** from **compute-bound** from **engine-queue-bound** from **host-disk-bound** regimes, each with distinct mitigation paths. Combined with the cumulative **\~219 h zero-event sustained AI workload** across three workload classes (SFT 50.47 h; Q6\_K-only MoE inference 117.55 h; mixed-backend concurrent residency 50.85 h, detail in Appendix §10.8), this is the publishable substrate envelope offered to the Blackwell-tuning-guide audience.

### 7. Limitations and Phase-B Follow-Ups

* **HumanEval marginal regression** at N=164; multi-seed audit path queued.
* **MMLU at the 1,995-row cap** (70%-power @ 3-pp); a full 14,042-row run would elevate to 80%-power @ 2-pp Δ.
* **Random-197-expert ablation baseline:** CLOSED at Phase-B (Appendix §10.4): the discriminative-screen hypothesis mechanically FAILED (random arm worse on only 1/4 probes vs the ≥ 2/4 criterion), with a peer-review-grade asymmetric finding: the screen wins on GSM8K math-reasoning by +4.09 pp, but HumanEval pass-count is **byte-equal (115/164) across two distinct 197-expert removal sets** at 25.4% overlap, selection-invariant at this prune fraction. Implication: the Phase-2 GREEN verdict is *robust to selection method*; pruning at this depth works regardless of which 197 experts are chosen, with GSM8K as the one material capability where the conjoint screen demonstrably outperforms random. The single-seed caveat bounds (does not rescue) the falsification.
* **Super-expert activation-magnitude proxy** used router-gate max as proxy for Su's full magnitude signature; the full proxy is approximate.
* **bf16 stratified cross-validation:** CLOSED at Phase-B (Appendix §10.3): the precision-gap hypothesis mechanically FAILED on its conjoint criterion (Spearman ρ 0.204 vs ≥ 0.95), while the Bland-Altman bf16-NLL limits-of-agreement axis **passed with margin** under the Kurtic protocol (LoA width 0.030 nats vs ≤ 0.10 threshold). The Spearman failure is attributed to axis-incomparability at cell grains far below the \~250-example reliability floor (Yauney 2025). No rescue by redefinition; the bf16-NLL fidelity result is reported as a separate positive finding.
* **GPQA recovery:** CLOSED at Phase-B (Appendix §10.5): with authentication fixed forward, the full 45-cell Wang grid ratio is **0.9866**, inside the pre-registered \[0.9326, 1.0326] PASS interval, so the AMENDMENT\_4 reduced-scope verdict was robust to scope-completion.

### 8. Reproducibility Envelope

Hardware and software per §3.1. Eval-bank sources are hash-pinned in the source manifest (MMLU, GSM8K, HumanEval, AIME-2024, GPQA (gated; token required for the Phase-B recovery), MATH-500, and the sovereign held-out slice). The grader is a local Qwen3.5-2B judge (sovereign local-first). Pinned artifact hashes: pruned Q6\_K GGUF 744f22ab…2133 (28.5 GB); baseline Q6\_K GGUF 57a2c173…61bd; stratified-subset definition 8d68b4e1…6aab. A reference rebuild seal exists for the full Phase-2 cascade; the complete artifact tree, decision ledger (D-P8-001 → D-P8-027), and per-file SHA-256 sidecars are archived in the sovereign repository's results directory.

### 9. Conclusions

The Architect 35B-A3B MoE → CTI-specialised model pruning pipeline validates at the pre-registered Gate-B with peer-review-clean integrity discipline. Pre-registration body and amendments are cryptographically frozen at γ.1 (three-file co-seal) and γ.2 (four-file re-seal); the H5 sovereign-perplexity primary passes at 2.4× margin under threshold (pruned arm slightly better than baseline); the H6 mandatory long-form-reasoning pass-rate ratio lands at 0.9826, decisive GREEN at the ≥ 0.90 floor and outside the CI-straddle band even at the AMENDMENT\_4 reduced scope; and the weighted Δ accuracy is −1.20 pp net (modest), with the HumanEval secondary regression flagged as marginal at one CI-half-width and routed to Phase-B as the primary audit test case.

The pruned model is deployable as the Architect at production Q6\_K with sovereign CTI specialisation. The Phase-B companion bundle (Appendix §10) closed the residual audit surface in three independent components, and its verdict pattern (two ancillary-axis falsifications reported without rescue, one scope-axis pass with margin, and the Phase-2 GREEN primary unfalsified across all three adversarial lenses) is itself the methodological payload: companion-validation as inferential strategy, executed under falsification-bar preservation contracts that forbid rescue by redefinition.

**Forward integration.** The VRAM and capability-preservation headroom established by this result is amplified, not consumed, by the other optimization layers of the Sovereign Optimization Flywheel: KV-cache compression is multiplicatively compatible with the pruned model; working-context reduction shrinks the byte budget that compression then compresses; and the in-weights specialization cadence reduces the in-context governance overhead the pruned model would otherwise spend on doctrine compliance. Further pre-registered prune-fraction increases (\~2.5%, \~5%) can be admitted under the same H6 ≥ 0.90 long-form floor at a less-binding VRAM constraint than this paper's design budget allowed.

**Seven peer-review-grade methodology contributions** are extracted from this sealed corpus and companion-published in the series' methodology reference: hash-first-before-edit + four-file re-seal discipline; document-as-found symmetric fail-soft scope-reduction; multi-version aggregator audit trail; mixed-backend eval-bank under shared pre-registration; three-axis substrate telemetry; pre-registration internal-contradiction resolution via interpretation-clarification amendment; and the tooling-orchestration-verification-gap class (Appendix §10.7). The contributions are not retrospective rationalisation: they were exercised simultaneously in this running pre-registered experiment, and the full decisional provenance trail (including the spurious-RED v1 aggregator output and the degenerate v1 merge output, both preserved as forensic anchors) is available for audit at the sovereign git forge.

### Appendix §10 · Phase-B Companion Validation (full γ-seal closure 2026-04-30)

> Phase-B audits three residual surfaces *around* the Phase-2 primary claim without modifying it. All three closed: two honest FAILs with substantive characterization, one PASS with margin. **Phase-2 D-P8-009 GREEN unchanged throughout**: the sealed-corpus contract held at every recompute (7/7 cryptographic anchors unchanged at every operational moment). Operator-ruled: this appendix extends the Phase-2 paper rather than standing as a separate publication.

#### §10.1 Posture and scope

Phase-B is structurally distinct from Phase-1 + Phase-2 along three peer-review-relevant axes: (a) **inferential purpose**: Phase-1 + Phase-2 establish a primary capability-preservation claim under destructive expert pruning; Phase-B audits residual surfaces around that claim without modifying it; (b) **statistical instrument**: Phase-B introduces a conjoint-test architecture (Bland-Altman 95% limits-of-agreement + Spearman ρ simultaneously; random-arm-worse-on-≥2/4-probes; full-grid ratio within a pre-specified PASS interval) where each component pre-registers *two* PASS conditions whose conjunction is the verdict; (c) **inferential posture**: Phase-B verdicts *bound* the Phase-2 primary claim by probing distinct ancillary assumptions whose mechanical FAIL produces substantive characterization rather than rescue. Phase-B findings update the Limitations and this appendix only; **Phase-2 sealed bytes never change as a function of Phase-B outcome.** The Phase-B pre-registration (canonical SHA 33a20d02…, sealed 2026-04-25, before any component executed) and AMENDMENT\_5 (SHA c766def8…, an interpretation-clarification sealed before its component's launch) carry falsification-bar preservation contracts: no rescue by redefinition, no threshold relaxation post-data, no hypothesis migration, no post-hoc weight change, no asymmetric arm treatment.

#### §10.2 The three components at a glance

| Component                                                                     | Hypothesis                                                                    | Mechanical verdict                                      | Substantive finding                                                                                                                                                                                                                                                                                                 |
| ----------------------------------------------------------------------------- | ----------------------------------------------------------------------------- | ------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| A · bf16 cross-validation 50.85 h mixed-backend, 1,200 paired cells, 0 errors | H\_pb1 PRECISION-GAP: is Q6\_K a faithful surrogate for bf16?                 | **FAIL (DIVERGENT)**: Spearman ρ 0.204 vs ≥ 0.95        | Bland-Altman bf16-NLL LoA **PASSES with margin** (width 0.030 nats vs ≤ 0.10 Kurtic envelope; bias 0.016; bootstrap CI95 \[0.014, 0.019]). Spearman fail attributed to axis-incomparability at cell grains far below the Yauney \~250-example floor. Reported as falsification + separate positive fidelity result. |
| B · random-197 ablation 32.43 h, seed 20260425, jaccard 25.4% vs screened set | H\_pb2 DISCRIMINATIVE-SCREEN: does the conjoint screen beat random selection? | **FAIL (1/4 worse)** vs ≥ 2/4 criterion                 | Asymmetric: screen wins GSM8K by **+4.09 pp**; sovereign PPL and MMLU random-comparable-or-better; **HumanEval byte-equal 115/164 across both distinct 197-expert sets**, selection-invariant. Implication: Phase-2 GREEN is robust to selection method at this prune fraction.                                     |
| C · GPQA recovery 16.78 h, auth fixed forward, 900 generations                | H\_pb3 GPQA-RECOVERY: did the AMENDMENT\_4 scope cut mask a scope effect?     | **PASS**: v4 full-grid ratio 0.9866 ∈ \[0.9326, 1.0326] | Scope-completion-effect-immateriality: full 45-cell Wang grid confirms the reduced-scope verdict (+0.004 delta). GPQA-diamond near-flat (0.5867 → 0.5844); AIME24 the only non-flat task (−0.0244); MATH-500 byte-equal.                                                                                            |

***Table 10.** D-P8-024 OVERALL closed 2026-04-30: 3/3 audit surfaces resolved.*

#### §10.3 Component A detail: precision gap

AMENDMENT\_5 R1 operationalises bf16 inference as forward-pass-on-Q6\_K-generated-sequence per the Kurtic protocol: Q6\_K-side autoregressive sampling via llama-server; bf16-side single-shot forward-pass via HF transformers with full CPU residency. Concurrent envelope: \~72 GB CPU (HF bf16) + \~12 GB GPU (llama.cpp Q6\_K) = 84 GB on 128 GB DDR5, \~44 GB headroom, three-axis telemetry active throughout, 0 errors across 183,073 s. The unit re-interpretation (pp → nats, per Kurtic Table 4's quantization-fidelity envelope) was sealed *before* the component launched; the Spearman threshold is rank-invariant under monotone transformation, so the re-interpretation does not weaken the rank-correlation criterion. Verdict at §10.2 Table 10: the conjoint fails on the Spearman axis; the load-bearing Kurtic axis passes with margin. Binary task-success and continuous per-token log-likelihood are not commensurable axes at n\_problems\_per\_cell ∈ {15, 25}: the pass-rate axis is degenerate at small N while the log-likelihood axis carries fine gradation. Higher-n replication above the Yauney floor is queued (§10.9).

#### §10.4 Component B detail: three-way scorecard

| Probe             | Baseline         | BI-screened      | Random-ablation                 | Random worse?                       |
| ----------------- | ---------------- | ---------------- | ------------------------------- | ----------------------------------- |
| Sovereign PPL     | 12.6418          | 12.5899          | **12.5626** (lower)             | no                                  |
| MMLU              | 0.8521           | 0.8396           | 0.8416 (within noise)           | no                                  |
| GSM8K strict      | 0.8840           | **0.9014**       | 0.8605                          | **YES (+4.09 pp screen advantage)** |
| HumanEval pass\@1 | 0.7561 (124/164) | 0.7012 (115/164) | **0.7012 (115/164, identical)** | tied, selection-invariant           |

***Table 11.** H\_pb2: 1 of 4 worse (< 2/4 criterion) → FAIL as pre-registered. Long-form (informational): screened 0.6898 vs random 0.6922.*

The interpretation is consistent with the super-expert literature: residual expert redundancy dominates capacity-loss at moderate prune fractions: at 197 removed, the residual expert pool recovers most capability-relevant paths regardless of selection methodology, with GSM8K's multi-step arithmetic dependency the exception where conjoint-screened experts are demonstrably load-bearing. The HumanEval byte-equal invariance across two sets at 25.4% overlap is striking enough to warrant cross-benchmark replication (MultiPL-E, BigCodeBench) before generalizing. Single-seed caveat pre-declared: it bounds (does not rescue) the falsification.

#### §10.5 Component C detail: the completed Wang grid

With authentication fixed forward (two distinct root causes: invalid token + un-accepted gated terms), GPQA-diamond ran 15 cells × 30 generations per arm (\~16 h 47 m, 0 errors, 1.43M tokens). Merged v4 full-grid long-form scorecard: AIME24 0.4467 → 0.4222 (the only non-flat task); MATH-500 byte-equal at 0.9573; GPQA-diamond 0.5867 → 0.5844 (near-flat); **45-cell aggregate ratio 0.9866**, inside the pre-registered PASS interval with ≈ 0.05 margin, +0.004 from the reduced-scope ratio. **Forensic chain disclosed honestly:** the first v4 merge orchestration produced silently degenerate output (a source-data-structure mismatch reading an empty field and aggregating on a non-existent one); the defect was discovered by post-hoc inspection of cascade-frozen artifacts *before* any verdict-claim authorship, fixed at a named commit, and the degenerate v1 cascade is preserved as the forensic discovery anchor rather than deleted.

#### §10.6 Multi-lens robustness: the substantive finding

**Phase-2 D-P8-009 GREEN is unfalsified by Phase-B.** Each audit surface probed a distinct assumption: precision-axis fidelity (H\_pb1: the Kurtic-protocol NLL axis holds; the conjoint fails at sub-floor aggregation grain), selection-method discriminativeness (H\_pb2: capability-specific, not general), and scope completion (H\_pb3: immaterial). The two FAILs concerned *ancillary validation surfaces*, not the primary capability-preservation claim, which was authored over capability-aggregate weighted accuracy + long-form ratio, not over precision-fidelity or screen-discriminativeness. The falsification-bar contracts forbid rescue by redefinition; Phase-B reports the asymmetric FAIL+FAIL+PASS pattern as the report-of-record. The pattern (two ancillary falsifications that bound the claim without rescuing it, plus one scope-axis pass that confirms the reduced-scope verdict at full scope) is the peer-review-grade methodology contribution: **companion-validation as inferential strategy.** Sealed-corpus contract: all 7 cryptographic anchors (four Phase-2 files + manifest + Phase-B pre-registration + AMENDMENT\_5) re-verified MATCH at every one of the four integrity sweeps.

#### §10.7 The tooling-orchestration-verification-gap class

A secondary methodology contribution: six silent-mismatch defects (design assumptions about artifact shape diverging from the actual artifact, producing output *internally consistent at type-check level but semantically empty or degenerate*) were discovered during Phase-B execution by post-hoc inspection of cascade-frozen artifacts, not by authoring-side guards. Four were corrected at audit moment (named fix commits: missing engine-metrics flag; a manifest generator whose no-write flag silently wrote; a stale-boilerplate report emitter; the §10.5 merge defect); two are queued (an argparse mode-mislabeling worked around by sequential arm dispatch; a stdout-encoding divergence that is non-blocking). Authoring-side guards would have caught five of the six; the sixth requires property-based testing because the empty-but-typed read was nominally well-formed. The class anchors a planned orchestration-verification harness: where the falsification-bar contracts forbid redefining *what* is measured, this contract enforces that *the instrument actually measures what the design intends*.

#### §10.8 Substrate envelope: \~219 h zero-event on consumer hardware

| Workload class                                                           | Wall-clock | Errors / OOM / thermal / NaN-Inf |
| ------------------------------------------------------------------------ | ---------- | -------------------------------- |
| (a) SFT 9B QLoRA (SSD-pilot arms)                                        | 50.47 h    | 0 / 0 / 0 / 0                    |
| (b) Q6\_K-only 35B MoE eval-bank inference (Phase-2 + β.3 + Component C) | 117.55 h   | 0 / 0 / 0 / 0                    |
| (c) Mixed-backend concurrent residency (Component A)                     | 50.85 h    | 0 / 0 / 0 / 0                    |
| Cumulative                                                               | \~219 h    | 0 / 0 / 0 / 0                    |

***Table 12.** Three workload classes, three-axis telemetry (host ⊕ GPU ⊕ engine) on every run: the publishable Blackwell substrate envelope.*

#### §10.9 Five pre-registered follow-up research questions

1. **RQ-Φ-1:** Conjoint-test sensitivity to aggregation grain. Replicate Component A at ≥ 250 problems per cell (the Yauney floor) over the full Wang grid: if Spearman recovers at higher n, the FAIL was grain-bound; if not, the conjoint architecture needs reformulation.
2. **RQ-Φ-2:** Random-ablation seed variance. Ten-seed bootstrap at the same prune fraction: does the GSM8K-only asymmetry hold across ≥ 8/10 seeds, and the HumanEval byte-equal pattern across ≥ 7/10?
3. **RQ-Φ-3:** Selection-invariance cross-benchmark replication. MultiPL-E + BigCodeBench at the same prune fraction: HumanEval-specific artifact, or a general property of code-completion under deep MoE pruning (a stronger finding deserving its own paper)?
4. **RQ-Φ-4:** AIME24 per-budget localization. Where in the {512…8192} budget space does the mild degradation live, one tier, distributed, or inverted at short budgets?
5. **RQ-Φ-5:** Orchestration-verification-harness coverage. Replay the six §10.7 defects against a property-tested harness: ≥ 5/6 caught means the compass artifact is operational.

#### §10.10 Phase-B-specific limitations

Single seed across all three components (bounds, does not rescue, every verdict); Component A cell grain below the Yauney reliability floor (the proximate cause of the Spearman FAIL); the bf16 subset at 600 cells per arm vs 1,650 full (chosen for the \~33 h budget; subset-vs-full reliability expected to preserve ≥ 95% ordinal agreement per BenchBench); GPQA recovery required operator token-action (a reproducibility burden the Phase-2 paper does not carry); the HumanEval selection-invariance and AIME24 degradation both pending their queued follow-ups; and multi-lens robustness is an inferential claim that *bounds* the Phase-2 verdict against three adversarial lenses; it does not extend Phase-2's pre-registered scope to claims that were not sealed at γ.2.

#### §10.11–§10.12 Evidence chain

The full Phase-B evidence chain (pre-registration, AMENDMENT\_5, tier-zero verdict SITREPs (collaborator-authored, operator-ratified at each close-out), gate verdicts, post-run integrity report, per-component cascade manifests with self-SHAs, the preserved forensic v1 cascades, per-probe arm artifact hashes, the decision-ledger closures (D-P8-018/019/024/025/026/027), the recipe version bump, rebuild seals, and the two Forgejo seal tags) is archived in the sovereign repository's results tree under the pruning project's phase-B companion directory, with every artifact carrying a SHA-256 sidecar. The 13 Phase-B-specific literature anchors are merged into this paper's References.

### System Update: July 2026 (appended; the sealed body above is unmodified)

The pruning program has advanced two model scales since this paper sealed, under the same pre-registration discipline, sealed across two arcs in July 2026. **At 122B:** a REAP-style one-shot expert prune of Qwen3.5-122B-A10B at 25 % (256 → 192 experts across all 48 layers; 23.3-hour observer pass on the operator's workstation, zero thermal events) produced a 186.2 GB checkpoint whose GGUF + matched-imatrix chain (built on a Q8\_0 host, a measured 29× wall-clock reduction over BF16-direct calibration at 2h04m versus 118h, with expert coverage 9,215/9,216) served at **16.65 tok/s versus the unpruned model's 15.60 at two quantization tiers better fidelity**. A pre-registered, sealed quality gate then returned **verdict PASS**, with pooled-math +2.5 pp (67.5 vs 65.0, n = 80), GPQA-198 +1.01 pp, and a MATH-500 tie, and the pruned model was **promoted to the production research-brain preset**, with before/after receipts taken on the deployed serving surface. Vision was subsequently re-enabled by extracting the multimodal projector from the base weights: the pruned text trunk needed zero rework, because vision enters at the embedding interface expert pruning never touches.

**At 35B:** two cut geometries (25 %/31 % → 26.61B/24.59B effective) extended this paper's destructive-only method with a prune-then-recover pattern of 200-step recovery fine-tunes (val CE 1.4681 → 1.2100 and 1.5113 → 1.2430) whose training receipts double as the on-box-trainability proof: **a 24.6B model trained at \~10.3 GiB resident on the 12 GB card**. The honest negatives are carried at full fidelity per the series' discipline: the fully-resident serving hypothesis FAILED its bar at these cut depths (IQ4\_XS actuals 12.68/13.69 GiB spilled past dedicated VRAM; the WDDM "dedicated-VRAM-lies" spill signature is itself a sealed finding), and a matched-corpus re-calibration experiment concluded at composite FAIL with an operator-ruled reframe: the role corpus had been mis-layered into the base-model slot, and the pre-registered forgetting guard caught it.

Provenance: the July 2026 pruning-arc roadmap seal footers (with commits and brain anchors), the sealed 122B quality-gate results (pre-registration through final seal and post-run integrity), and the 35B prune-drafter receipts (amendments A1 through A8). This update is append-only; §1 through §10 above remain the sealed record.

***

*The Sovereign Stack · Sovereign Domain Pruning · Chapter 16 · Part IV · v1.0.0 · License CC BY 4.0 · © Jamey Kistner, OSINTelligence LLC*

**Citation (preferred):** Kistner, J. (2026). *Sovereign Domain Pruning: Destructive Expert-Pruning of a 35B-A3B MoE Against a Sovereign CTI Corpus*, version 1.0.0. OSINTelligence LLC research whitepaper. Cited in-series by title.

*The reference list and provenance follow as a sub-page of this chapter.*