> 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/references-and-provenance.md).

# References & provenance

### References

\[1] llama.cpp. llama-perplexity tool and --kl-divergence-base flag. ggml-org/llama.cpp, examples/perplexity/perplexity.cpp. <https://github.com/ggerganov/llama.cpp> (retrieved 2026-04-13).

\[2] kalomaze (2023). "Perplexity / PPL, as a quantization loss benchmark, is inaccurate - KL divergence seem to be a better data point." ggml-org/llama.cpp Discussion #4110, opened 2023-11-17. <https://github.com/ggml-org/llama.cpp/discussions/4110> (verified 2026-04-14).

\[3] localbench. Gemma 4 31B quantization scan, per-quant KLD reference table. <https://github.com/christianazinn/localbench> (retrieved 2026-04-13; approximate values, exact pull pending).

\[4] HuggingFace Blog. KLD-guided quantization series on imatrix calibration trade-offs (specific post URLs to be pinned at submission).

\[5] bartowski. Published per-quant KLD tables in HuggingFace model cards; e.g. Llama-3.3-70B-Instruct quants reporting Q6\_K mean KLD \~0.18–0.22 against BF16 reference.

\[6] Sonatype. *State of the Software Supply Chain* (2024 edition). <https://www.sonatype.com/state-of-the-software-supply-chain>

\[7] NIST SP 800-204D (2024). *Strategies for the Integration of Software Supply Chain Security into DevSecOps CI/CD Pipelines.* <https://csrc.nist.gov/publications/detail/sp/800-204d/final>

### Data and Code Availability

The validation harness, quantization artifacts (22.8 GB GGUF), imatrix file (imatrix\_apex\_v6.dat), live tracking log with full provenance, final report, and raw data with parsed JSON are archived in the OSINTelligence repository and are available to qualified reviewers on request. The calibration corpus contains operator memory content and is not shareable; the construction methodology (§3.2) is fully specified for replication.

### Author Contributions

**Jamey Kistner** is the sole author: he conceived the sovereign-imatrix hypothesis, owns the OSINTelligence pipeline and Hindsight memory store that provided the Phase A calibration corpus, defined the four-gate acceptance criteria, operated the quantization run, verified every claim against the sealed artifacts, and provided domain direction throughout. Large language models were used as research tools in support of the work (validation-harness implementation, parsing of llama-perplexity output into the metric pipeline, methodology research that surfaced --kl-divergence-base as the correct full-vocab KLD tool, experiment tracking, and draft composition under operator review); their role is detailed in the AI-assistance disclosure below.

### Acknowledgments

Methodology guidance (specifically the existence of llama-perplexity --kl-divergence-base as the correct full-vocab KLD tool) came from llama.cpp Discussion #4110 and the localbench project's Gemma-4 calibration benchmarks. The author thanks the llama.cpp maintainers and the broader open-source quantization community.

**Evidence & seal.** The four-gate validation behind this chapter, reproduced verbatim in the record's evidence section with the V1–V4 all-green table and the GGUF verification provenance: [APEX Imatrix Calibration (P5)](/osintelligence/evidence-and-seals/apex-imatrix-calibration-p5.md). (Report-class: this run pre-dates the canonical-prefix sealing discipline, so it carries a validated report rather than a self-hash, stated as such on the page.)

### AI-assistance disclosure

Large language models were used as research tools in the preparation of this chapter: Claude Opus 4.6 (Anthropic). The model versions and roles named above keep the provenance of this chapter auditable. No AI system is listed as an author or credited as a contributor, in line with COPE and ICMJE guidance: an AI system cannot take responsibility for the work, cannot assert competing interests, and cannot enter a licence agreement. The author verified every claim in this chapter against the sealed artifacts and is solely accountable for it.

**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.

**License:** CC BY 4.0 (text). Code and data artifacts MIT per repository license.

**Corresponding author:** Jamey Kistner, <jamey.kistner@osintelligence.io>, OSINTelligence LLC (Columbus, OH).

***

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