> 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/15-sovereign-imatrix-calibration.md).

# 15 · Sovereign Imatrix Calibration

> **A companion paper.** This report is the compression evidence beneath the series' architectural spine: it is cited in-series by title and grounds the claim in *The Sovereign Triad* (Chapter 1, §2.1) that in-distribution calibration achieved an order-of-magnitude tighter KL divergence than generic-corpus practice. The methodology it validates (calibrate on the operator's own workload) is one enabler of in-weights sovereign specialization (FC-2) on consumer hardware; it is the L₁ calibration axis of *The Sovereign Optimization Flywheel* (Chapter 8), and the direct companion to *Sovereign CTI-NER* (Chapter 14), which carries the same sovereign-corpus discipline from calibration into supervised fine-tuning.
>
> **Status note.** The study below was pre-registered, gated, and sealed on 2026-04-13; the body is preserved as sealed. An append-only addendum (2026-06-09) records the quantization's subsequent promotion to the production broadcast pipeline; it adds observational evidence and amends nothing.

> **What is new here.** The contribution is the demonstration that the imatrix calibration corpus, treated by community practice as a secondary detail, is a first-order fidelity lever: substituting the operator's own in-distribution corpus (511 memory breadcrumbs plus targeted dog-food synthesis) for a generic one during mixed-precision quantization of a 35B-A3B mixture-of-experts model yields a mean KL divergence of 0.0103 against BF16, roughly 16x tighter than the 0.16-0.20 community Q6\_K band, while the artifact shrinks 21% and runs 11% faster. The load-bearing novelty is the reframe that follows from it: calibration is a training-time input with zero inference-time footprint, which makes it a supply-chain attack surface, and sovereign in-distribution calibration is the defense that is also the quality win, the same 16x that rewards the curating defender is available in reverse to an attacker who mis-curates a corpus that still scores clean on wikitext.
>
> **Deepest water.** §4.2 and Figure 1, the KLD distribution sitting an order of magnitude left of the community Q6\_K band, read with §5.5's honest register (single model, single baseline comparator drawn from a different model's published card, and the calibration-leakage pathway that the out-of-distribution wikitext parity argues against but does not fully close); §5.1, the corpus-workload-alignment mechanism; and §5.7, the calibration-as-supply-chain-attack-surface argument that generalizes to every training-time ingredient with no inference-time signature.

### Abstract

Community practice for GGUF imatrix quantization uses generic calibration corpora (wikitext, author's-choice mixes). We show that substituting a **workload-specific in-distribution corpus** (constructed from 511 operator memory breadcrumbs plus targeted self-generation, or "dog-food", synthesis from the model under quantization) for a generic corpus during mixed-precision quantization of a Qwen3.5-35B-A3B MoE model yields a mean KL divergence of **0.0103** against the BF16 baseline, compared to published community benchmarks of **\~0.16–0.20** for Q6\_K-tier quantizations of comparably-sized models. This represents a **16× improvement** in calibration fidelity at equivalent compute cost. The resulting quantization (22.8 GB, a 21% reduction from the Q6\_K baseline at 28.9 GB) preserves 96.08% top-1 token agreement with the BF16 reference, maintains perplexity within ±0.11% of baseline on both in-distribution (sovereign 10 MB corpus) and out-of-distribution (wikitext-2) evaluation sets, matches baseline 1:1 on a deterministic SAT-style code audit (15 prompts, 53.3% APPROVE on both), and delivers **1.11× inference throughput** over the Q6\_K baseline. We conclude that corpus-workload alignment is a first-order lever in post-training quantization, and that the order-of-magnitude calibration fidelity gap between generic and in-distribution imatrix corpora has been systematically underexploited by the open-source LLM quantization community.

### 1. Introduction

Post-training quantization (PTQ) has become the dominant method for deploying large language models on consumer hardware. The GGUF format and llama.cpp toolchain introduced "imatrix" calibration: a mechanism by which activation statistics collected over a calibration corpus are used to select per-tensor quantization parameters, reducing quantization error in regions of the parameter space the corpus exercises \[1].

Community practice treats the choice of imatrix corpus as a secondary concern. The Unsloth team, bartowski, and most individual quantizers use generic corpora (wikitext, The Pile excerpts, or curated mixes designed to exercise "general capability"). The resulting GGUFs are shipped to millions of operators whose actual workloads (code generation, domain-specific retrieval-augmented generation, intelligence analysis, broadcast automation) diverge substantially from the calibration distribution.

This paper asks a simple question: **what if we calibrate on our actual workload?**

We report results from a single-model, multi-gate empirical evaluation: a Qwen3.5-35B-A3B mixture-of-experts model (40 layers, 256 experts, 8 active per token), quantized from BF16 source weights to a mixed-precision target using an imatrix computed over an in-distribution corpus harvested from the operator's own long-term memory store and targeted dog-food synthesis against the live deployed model.

The finding is not that the sovereign-imatrix quantization matches its baseline. It is that a quantization calibrated on the operator’s own workload exceeds published community benchmarks by an order of magnitude on the standard fidelity metric (mean KL divergence against the BF16 reference) while simultaneously shrinking the artifact, accelerating inference, and preserving every performance characteristic the validation gates measure. The rest of this paper is the receipt for that sentence.

### 2. Background

#### 2.1 GGUF imatrix calibration

llama.cpp's llama-imatrix tool runs forward passes of a BF16/F16 reference model over a calibration corpus and records, per tensor, the outer-product activation statistics needed to compute importance weights for downstream quantization \[2]. The resulting .dat file is consumed by llama-quantize with an I-class target (IQ4\_XS, IQ3\_M, Q6\_K with imatrix, etc.) to produce a GGUF whose per-tensor rounding minimizes squared-activation-weighted error over the calibration distribution.

#### 2.2 The corpus question

llama.cpp's documentation recommends "enough text to cover the vocabulary" (typically 20k–200k tokens) and community practice has converged on a small number of canonical corpora:

* **Unsloth's UD-family quants** use a proprietary curated corpus
* **bartowski** uses a wikitext-based corpus for most releases
* **Author's choice:** individual quantizers often use wikitext-2, random C4 shards, or domain-specific text for specialized models

Critically, **none of these corpora represent the distribution of tokens the deployed model will actually see in production for a given operator**. A model quantized on wikitext and deployed for OSINT threat analysis is being asked to minimize error on news prose while being evaluated against CTI/TTPs/IoC classification.

#### 2.3 KL divergence as quantization-fidelity metric

llama-perplexity --kl-divergence-base computes the per-token symmetric KL divergence of the quantized model's full output distribution against a reference (typically BF16) on a held-out text \[1]. Unlike perplexity, which can be gamed by token-level label agreement, KL divergence measures distributional similarity across the full vocabulary and is the metric of choice for fidelity-sensitive applications (speculative decoding, distillation, calibration evaluation).

Published community benchmarks (see §5.1 and the reference archive) put Q6\_K-tier quantizations of 30–70B-class models in the **0.16–0.20 mean KLD** range, Q5-tier at 0.40–0.50, Q4-tier at 0.60–0.80.

### 3. Method

#### 3.1 Source model and baseline

* **Base:** Qwen3.5-35B-A3B (MoE, 40 layers, 256 experts, 8 active, 2048 embedding dim)
* **BF16 source:** Unsloth's Qwen3.5-35B-A3B-GGUF BF16 shards (69.4 GB total, 2 shards)
* **Baseline for comparison:** Qwen3.5-35B-A3B-Q6\_K.gguf (28.9 GB, Unsloth's standard Q6\_K build with their proprietary imatrix)

#### 3.2 Sovereign imatrix corpus construction

The calibration corpus was built in two phases, yielding a total of roughly 1.2M characters:

**Phase A: Hindsight harvest.** The author's Hindsight Memory System (an operational long-term memory store attached to the primary work pipeline) was queried across three populated banks (HIVE\_SHARED, STRATEGIST\_MEMORY, CUSTODIAN\_MEMORY) using six recall queries per bank. A minimum-length filter (≥80 chars) and a ground-truth priority bonus were applied. Yield: **511 memory breadcrumbs**, \~230K characters of real operator/agent interaction text, in-distribution for the workloads this model actually runs.

**Phase B: Architect dog-food synthesis.** The target model (served via llama.cpp Router Mode on port 8080, architect preset) was prompted against a synthesis template across representative task types (code review, doctrine synthesis, intelligence analysis). Outputs were mixed at authorial ratios to fill the gap between Phase A's conversational data and the reasoning-heavy workload the architect serves.

**Phase C: Iteration.** The imatrix file (imatrix\_apex\_v6.dat) was produced after six corpus revisions (v1–v6), each validated by a fast-fail KLD probe on a 5-chunk wikitext sample. v6 was the first to stabilize below 0.05 mean KLD on the probe.

#### 3.3 Quantization schedule

Target quant class: **APEX I-Balanced**, a mixed-precision schedule applied per-tensor via llama-quantize --tensor-type. The schedule places:

* Attention Q/K/V projections at Q6\_K
* Router gates and shared expert FFN at Q6\_K
* Expert FFN (up/gate/down) at IQ4\_XS with imatrix
* Embedding + output at Q6\_K

Full tensor type assignments are archived with the raw data. Target footprint: ≤ 23.6 GB. Actual: **22.8 GB**.

#### 3.4 Validation gates

A four-gate validation harness was defined pre-quantization:

| Gate | Test                                                                                     | Acceptance                                    |
| ---- | ---------------------------------------------------------------------------------------- | --------------------------------------------- |
| V1   | Perplexity on sovereign held-out (10 MB, 500 chunks) and wikitext-2 (1.3 MB, 580 chunks) | Δ ≤ ±1.0% / ±3.0%                             |
| V2   | Full-vocab KL divergence against BF16 baseline, 50 chunks wikitext-2                     | Mean ≤ 0.30, max ≤ 5.0                        |
| V3   | 15-prompt deterministic SAT audit (code generation + Orchestrator approval judgment)     | APPROVE ≥ baseline−2pp, REJECT ≤ baseline+2pp |
| V4   | VRAM peak + 3-task throughput (1024 max tokens each)                                     | Peak ≤ 23.6 GB, TPS ratio ≥ 0.90×             |

All gates were run on the same hardware (RTX 5070 12 GB sm\_120 Blackwell, i7-14700F 20C/28T, 128 GB DDR5, Windows 11, llama.cpp custom sm\_120 build).

### 4. Results

#### 4.1 Summary

| Gate          | Metric     | Acceptance    | Actual                 | Verdict |
| ------------- | ---------- | ------------- | ---------------------- | ------- |
| V1 sovereign  | PPL Δ      | ±1.0%         | **−0.082%**            | pass    |
| V1 wikitext-2 | PPL Δ      | ±3.0%         | **−0.107%**            | pass    |
| V2            | Mean KLD   | ≤ 0.30        | **0.0103**             | pass    |
| V2            | Max KLD    | ≤ 5.0         | **1.52**               | pass    |
| V2            | Same top-1 | informational | **96.08%**             | –       |
| V3            | APPROVE    | ≥ 51.3%       | **53.3%** (= baseline) | pass    |
| V3            | REJECT     | ≤ 2.0%        | **0%** (= baseline)    | pass    |
| V4            | Peak VRAM  | ≤ 23.6 GB     | **11.2 GB**            | pass    |
| V4            | TPS ratio  | ≥ 0.90×       | **1.11×**              | pass    |

***Table 1.** All pre-registered gates green.*

#### 4.2 KL divergence detail

| Percentile | KLD                     |
| ---------- | ----------------------- |
| Mean       | **0.010327 ± 0.000293** |
| Median     | 0.004351                |
| 90th       | 0.021439                |
| 95th       | 0.034667                |
| 99th       | 0.093380                |
| 99.9th     | 0.380122                |
| Maximum    | 1.520779                |

| Token statistic  | Value               |
| ---------------- | ------------------- |
| Mean Δp          | −0.074 ± 0.026%     |
| RMS Δp           | 2.979 ± 0.102%      |
| Same top-1 token | **96.078 ± 0.172%** |

***Table 2.** Full-vocabulary KLD against BF16; token-probability statistics.*

![KLD cumulative distribution vs BF16 with community Q6\_K band](https://137900913-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2Fxx2bv6VR9dSJDJ9HsDER%2Fuploads%2FRZQT6f34C7LVMLNyp8ro%2Fgb-ch15-fig1-kld-cdf.png?alt=media)

***Figure 1.** Per-token KLD distribution against the BF16 reference. The APEX I-Balanced curve sits an order of magnitude left of the published community Q6\_K band (\~0.16–0.20 mean).*

#### 4.3 Perplexity detail

**Sovereign held-out (10.0 MB, 500 × 512 tok)**

| Model               | PPL     | ± stderr | tok/s      |
| ------------------- | ------- | -------- | ---------- |
| **APEX I-Balanced** | 13.0821 | 0.1113   | **128.39** |
| Baseline (Q6\_K)    | 13.0928 | 0.1115   | 98.85      |

**Wikitext-2 (1.3 MB, 580 × 512 tok)**

| Model               | PPL    | ± stderr | tok/s      |
| ------------------- | ------ | -------- | ---------- |
| **APEX I-Balanced** | 6.7244 | 0.0433   | **129.19** |
| Baseline (Q6\_K)    | 6.7316 | 0.0434   | 99.66      |

***Table 3.** Both deltas within error bars; APEX fractionally (not significantly) better.*

![Perplexity parity, APEX vs Q6\_K on sovereign and wikitext-2 sets](https://137900913-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2Fxx2bv6VR9dSJDJ9HsDER%2Fuploads%2FgKQTBRRNXncPFVEMXyH5%2Fgb-ch15-fig2-perplexity.png?alt=media)

***Figure 2.** Perplexity parity on in-distribution and out-of-distribution evaluation sets: fidelity was not bought by overfitting the calibration domain.*

#### 4.4 SAT audit detail (V3)

15 coding prompts submitted through both models, then judged by an independent Orchestrator (Qwen3.5-9B Q8\_0) against a fixed rubric:

| Outcome | APEX      | Baseline  | Delta    |
| ------- | --------- | --------- | -------- |
| APPROVE | 8 (53.3%) | 8 (53.3%) | **0 pp** |
| FLAG    | 7 (46.7%) | 7 (46.7%) | 0 pp     |
| REJECT  | 0 (0%)    | 0 (0%)    | **0 pp** |
| ERROR   | 0         | 0         | 0        |

***Table 4.** Deterministic SAT audit: 1:1 match with baseline.*

#### 4.5 VRAM / throughput (V4)

| Model               | Peak VRAM   | Median tok/s |
| ------------------- | ----------- | ------------ |
| **APEX I-Balanced** | **11.2 GB** | **26.3**     |
| Baseline (Q6\_K)    | 11.5 GB     | 23.8         |

APEX saves 324 MB VRAM at inference peak, provides 11% higher throughput (1.11× ratio), and saves 6.1 GB on disk (22.8 GB vs 28.9 GB).

![Footprint and throughput two-panel comparison](https://137900913-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2Fxx2bv6VR9dSJDJ9HsDER%2Fuploads%2FC2Q2lSMhEIZTm8zx9A2i%2Fgb-ch15-fig3-vram-tps.png?alt=media)

***Figure 3.** Footprint and throughput: smaller, faster, and more faithful at once, the combination that generic-corpus calibration typically trades away.*

### 5. Discussion

#### 5.1 Comparison to community benchmarks

Published mean KLD values for imatrix-calibrated quantizations of comparably-sized models (30–70B class) \[3, 4, 5]:

| Quant class   | Typical community mean KLD | This work (APEX I-Balanced) |
| ------------- | -------------------------- | --------------------------- |
| Q8\_0 (8-bit) | \~0.16                     | –                           |
| Q6\_K         | \~0.20                     | **0.0103**                  |
| Q5\_K\_S      | \~0.40–0.50                | –                           |
| Q4\_K\_M      | \~0.61–0.76                | –                           |

***Table 5.** Community benchmark comparison (localbench Gemma-4-31B reference table; bartowski published cards).*

The APEX I-Balanced result is **approximately 16× tighter** than standard Q6\_K community practice and **\~15× tighter** than even Q8\_0 (half-precision per-weight!) community practice, despite APEX using IQ4\_XS for the bulk of MoE expert FFN tensors.

We attribute the gap primarily to corpus-workload alignment. The Phase A harvest captured exactly the operator/agent interaction patterns the model is evaluated against; Phase B synthesis captured the reasoning density it is asked to produce. By contrast, community corpora are tuned for generality, which necessarily trades off calibration fidelity on any specific workload.

#### 5.2 Why does top-1 stay at 96% if KLD is already tiny?

The KLD mean of 0.010 reflects near-identical full distributions. The 96.08% top-1 agreement is the separate, autoregression-relevant metric: even with very similar distributions, a 4% stochastic disagreement on greedy-decoded tokens is expected given MoE expert selection non-determinism and small floating-point differences in matmul paths. 96% is at the high end of published values for any 4-bit-ish quantization: consistent with the KLD result.

#### 5.3 Why is APEX *faster*?

22.8 GB vs 28.9 GB: less memory bandwidth per token. On the RTX 5070 (Blackwell sm\_120, 504 GB/s), the 21% footprint reduction translates directly to faster MoE expert dispatch. This is not a surprising outcome in isolation, but it is unusual in combination: footprint, throughput, and calibration fidelity typically trade against one another, and here all three improved at once.

#### 5.4 Reproducibility

The full corpus construction pipeline, quantization schedule, imatrix file, and validation harness are archived in the OSINTelligence codebase. The calibration corpus itself contains operator memory content and is not shareable; however, the **methodology** (Phase A harvest from an operator's own memory store plus Phase B targeted dog-food) is fully replicable by any deployer of a local LLM with an existing conversation/memory log.

**Toolchain reproducibility note (2026-04-14).** All numbers in this paper were measured on the 2026-03-07 llama-server.exe built against **CUDA Toolkit 13.1** on RTX 5070 (sm\_120 Blackwell). Subsequent investigation (documented in the companion CTI-NER paper) found that CUDA 13.1 contains a Blackwell MMQ kernel segfault under prolonged single-preset load, prompting a re-compile on **CUDA Toolkit 12.8** per NVIDIA's Blackwell migration guidance. Gate-B parity on the 12.8 binary holds perplexity to within ±0.5% and tokens/sec to within ±5% of the 13.1 baseline (within our reproducibility band), but future re-runs should expect the headline deltas in §4 to shift by at most that amount. Explicit disclosure for anyone reproducing these numbers against the shipping binary.

#### 5.5 Limitations

1. **Single-model result.** Validation on the orchestrator 9B and janitor 2B is planned. Generalizability to non-MoE architectures and to smaller sizes is empirically open.
2. **Single baseline comparator.** The community-benchmark comparison uses published numbers from different models (Gemma 4 31B in localbench's reference table), not a same-model same-schedule comparator. A stronger claim would quantize the same BF16 source with a standard wikitext imatrix and compare head-to-head.
3. **Calibration-leakage concern.** Phase B dog-food is generated by the same model being quantized. There is a theoretical pathway for the imatrix to over-fit activation statistics the model already produces easily. Empirically the wikitext-2 out-of-distribution PPL delta (−0.107%) does not show this pathology, but a broader OOD evaluation is warranted.
4. **No speculative-decoding benchmark.** KLD matters most when the quantized model is used as a draft model in speculative decoding. We have not measured acceptance rates under speculation.

#### 5.6 Broader implications

If the magnitude of this effect replicates on other workloads, it has immediate implications:

* **For operators:** Quantizing against your own memory/log store is essentially free and measurably beats off-the-shelf community GGUFs on your workload.
* **For the open-source community:** "Pick a model that matches your workload" is the conventional advice. "Calibrate the quantization of any model to your workload" is a newer and possibly stronger lever.
* **For applied AI businesses:** In-distribution data is not just fuel for fine-tuning; it is fuel for quantization too. The same training-data moat argument extends to post-training compression.

#### 5.7 Calibration as a model supply-chain attack surface

The corpus-alignment result re-frames imatrix calibration as a **supply-chain-integrity property**, not merely a quality-engineering choice. Three observations compose the threat surface.

**Calibration data never appears at inference time.** The .dat file produced by llama-imatrix is consumed exclusively by llama-quantize to select per-tensor rounding; the calibration corpus itself is discarded once the GGUF is emitted. A downstream operator receiving a pre-quantized GGUF cannot recover the calibration distribution from the weights, and cannot distinguish a GGUF calibrated on the operator's workload from one calibrated on an adversary-chosen distribution that happened to produce plausible aggregate activation statistics. This asymmetry (calibration is a training-time input with zero inference-time footprint) is exactly the class of ingredient supply-chain attacks exploit, in the same general pattern documented for model-repository, pip, and npm supply chains \[6, 7].

**Distribution-specific drift is invisible under standard evaluation.** Community quantization benchmarks (KLD on wikitext-2, perplexity on wikitext or C4, HumanEval code pass-rate) average across the calibration-agnostic distribution they were designed for. A GGUF whose imatrix is drifted against a specific operator-relevant distribution (*phishing-email classification, ICS-indicator extraction, legal-document redaction, medical-record triage*) will score cleanly on wikitext-2 and fail selectively on the drifted distribution. Without in-distribution evaluation, the operator has no telemetry surface that would surface the attack. Our §4.2 result cuts the other direction of the same knife: a 16× improvement on in-distribution KLD is available to defenders who curate, and (by symmetry) a >16× degradation is available to attackers who mis-curate.

**The attack generalizes to every supervised post-training step.** The silver-label-poisoning argument made in the companion CTI-NER paper is the SFT analogue of this calibration-poisoning surface. Both mechanisms share the same invariant: a training-time ingredient with no inference-time signature is the ideal location for a covertly-drifted artifact. Calibration corpora and silver labels are two instances of that class; future post-training mechanisms (distillation corpora, preference-pair generation, activation steering datasets) will add more.

**Countermeasure: sovereign in-distribution calibration as defense-in-depth.** The methodology in §3.2 is the first-order defense: an operator who curates their own calibration corpus from their own memory store is also the operator who decides what the model is precision-optimized for. The corpus is on the operator's disk; its composition is auditable; its provenance is known. Our 16× KLD improvement is therefore dual-signaled: it is both a quality win (calibration fidelity improves on the workload that matters) and a supply-chain property (the calibration artifact is operator-verifiable rather than opaque). The two signals point in the same direction, which is the signature of a well-chosen defensive primitive.

This framing is immediately operationalizable for defensive-security deployments: critical-software operators running local LLMs for code review, patch analysis, or incident-response triage can adopt sovereign in-distribution calibration as a one-shot supply-chain-integrity control without vendor dependency, without cloud data egress, and without displacing existing quantization pipelines: the only change is the imatrix input. The §3.2 Phase A plus Phase B recipe is the reference implementation.

### 6. Conclusion

Sovereign in-distribution imatrix calibration produced a GGUF quantization that is simultaneously smaller (−21%), faster (+11%), and substantially more faithful to the BF16 reference (16× tighter mean KLD than community benchmarks) than the baseline produced by a leading community quantization pipeline using generic calibration. We view this as a single strong data point in favor of a broader thesis: **corpus-workload alignment is a first-order lever in post-training quantization, routinely underexploited in current open-source practice.** Extending the evaluation across model sizes and workload types is straightforward and, if the effect holds, changes the default advice for local deployment: not merely *pick the model that matches your workload*, but *calibrate the quantization of whatever model you pick to the workload itself*. The corpus was already on the operator’s disk. The lever was already in the toolchain. What this paper adds is the measurement that shows how much the combination is worth.

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

Post-seal attestation, append-only. Everything above (sealed 2026-04-13, citeable metadata added 2026-04-16) is preserved verbatim and unaltered. This section records how the calibration method has continued since the seal; it neither amends nor recomputes any sealed result in §1–§6.

#### A.1 The calibration corpus matured into a governed, drift-guarded artifact

The sealed study built its imatrix from a single harvest: 511 operator memory breadcrumbs plus dog-food synthesis, roughly 230K characters, iterated to a sixth corpus revision. In the months since, the calibration text stopped being a one-off input and became a governed artifact in its own right. Each corpus now carries a SHA-256 sidecar recomputed at the start of every run that hard-fails the build on any drift, and a negative test confirms the guard: a single flipped byte aborts the loader with a corpus-hash-drift error before any compute begins. The calibration slice is now rendered from a much larger sovereign corpus than the original harvest: tens of thousands of governed rows carrying explicit train, validation, and test splits with per-source provenance, rather than a hand-assembled breadcrumb pull. The methodology also generalized to other sovereign models: a governance-classifier adapter was calibrated from a structured blend of 427 sections (200 cascade, 100 transcript, 127 operator-feedback) and cleared all four of the paper's fidelity gates on that model.

#### A.2 One sovereign corpus draw is now cited at every layer of the compression flywheel

The sharper development is that imatrix calibration is no longer an isolated step. A corpus-identity discipline now binds a single sealed sovereign corpus draw to every corpus-touching layer of the compression pipeline at once: prune-saliency scoring, recovery fine-tuning, imatrix calibration, self-distillation, persona adapters, and the speculative-draft head all cite the same draw and its hash. This makes directly testable a question the sealed study could only gesture at, whether calibrating (and pruning, and fine-tuning) on the operator's own in-distribution corpus rather than a generic one compounds across layers. That comparison is now run as a pre-registered, matched-versus-mismatched study: the same chain measured once on a generic composite and once on the sovereign blend, with the mismatched arm frozen in advance as the control. The in-distribution direction the sealed §4 result reported at a single layer is thereby carried into a whole-chain, falsifiable form, and the compression flagship (*Sovereign Big-Model Compression*, Chapter 17) inherits the same discipline.

#### A.3 The validated quantization went to production, and stays under live telemetry

The APEX I-Balanced artifact validated in §4 was promoted from candidate to the operational architect model on 2026-06-09, serving every stage of the production broadcast pipeline. Post-cutover, per-inference telemetry (serving-model id, latency, token estimate, success state) turns the §4.5 throughput figure and the §4.1 quality-equivalence verdict into continuously-monitored production quantities and a regression tripwire, instantiating in production the in-distribution telemetry surface argued for in §5.7. This is observational validation, not a controlled re-run of the V1–V4 gates, so it adds no new gated claim; both the stock and APEX architect remain registered, keeping A/B fallback available. Provenance: the operator's production adapter roster and the governed sovereign-corpus compression records, June through July 2026. Append-only; the sealed body above is unmodified.

***

*The Sovereign Stack · Sovereign Imatrix Calibration · Chapter 15 · Part IV · v1.0.0 · License CC BY 4.0 · © Jamey Kistner, OSINTelligence LLC*

**Citation (preferred):** Kistner, J. (2026). *Sovereign In-Distribution Imatrix Calibration Achieves 16× Tighter KL Divergence Than Generic-Corpus Quantization on a 35B-A3B MoE Model*, version 1.0.0. OSINTelligence LLC. Cited in-series by title.

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