> 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/20-mrope-serving-path-rca/upstream-attribution-references-and-provenance.md).

# Upstream attribution, references & provenance

### Upstream attribution

Every fix composed in this paper originates with a named member of the llama.cpp community, not with the present author. The author's contribution is the falsification isolation, the composition onto a pinned build, the attribution bundle, and the 2,004-row production validation, *not* the fixes themselves. Readers are encouraged to open the sources and confirm the credit directly:

* The heap over-read and its **two-part fix** (the ubatch\_reserve resize and the state\_read\_meta M-RoPE broadcast) were diagnosed and authored, in full and verbatim as applied here, by the reporter of [llama.cpp issue #20093](https://github.com/ggml-org/llama.cpp/issues/20093) (reproduced on an RTX 5090; the reporter's handle is shown at the link).
* The K-shift assertion was independently reported by **KernelFreeze** in [llama.cpp issue #19915](https://github.com/ggml-org/llama.cpp/issues/19915) (reproduced on an RTX 3080), and fixed by **Georgi Gerganov (@ggerganov)**, reviewed by **@ngxson**, in [PR #19928](https://github.com/ggml-org/llama.cpp/pull/19928) (merged as commit [99bd67c](https://github.com/ggml-org/llama.cpp/commit/99bd67c9b29851f758c7d22caa8fc57fd5af3e4f)), which this deployment cherry-picks verbatim.

The llama.cpp project is maintained by Georgi Gerganov and contributors under the MIT license.

### Author Contributions

**Jamey Kistner** is the sole author: operational diagnosis, CUDA toolchain management, rebuild pipeline, production validation gating, and editorial control. Large language models were used as research tools in support of the work (web-research synthesis, upstream-issue correlation, patch-bundle composition, specification authorship, and paper drafting under operator review); their role is detailed in the AI-assistance disclosure below.

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

The defect class this paper documents outlived its original binary. At the fleet-modernization arc (June 2026) the serving binary this paper's forensics were run on was retired for a mainline build carrying the model family's vision path natively, and the upstream M-RoPE heap-over-read this paper's patch closed was **verified still open upstream on 2026-06-25**, six-plus weeks after this paper's build pinned it. The rebuild itself was governed in this paper's own register: operator-approved under the infrastructure rule, prior binary preserved as instant rollback, and the sealed deploy record states the new production binary runs **zero local serving patches** (the mainline vision merger having absorbed the affected path; the patch-disposition note and the deploy record are both carried in the working ledger for the operator's confirmation rather than silently reconciled).

The paper's deeper thesis, that RoPE-class serving defects are *silent* and only coherence-checking catches them, became standing fleet doctrine at the same arc: the new-generation cutover smoke was expanded from load-checking to **output-coherence checking**, precisely because a wrong-RoPE-parameter serve of the new model family corrupts silently. Provenance: the June 2026 fleet-modernization roadmap seal footers and the 2026-06-25/26 deploy and cutover handoff records. Append-only; the sealed forensic record above is unmodified.

### References

Alibaba Cloud / Qwen team (2025). Qwen structured-output documentation. <https://qwen.readthedocs.io/>

JJJYmmm and contributors (2026). "Revisiting Multimodal Positional Encoding in Vision-Language Models." *ICLR 2026.* [arXiv:2510.23095](https://arxiv.org/abs/2510.23095). MRoPE-Interleave design-space survey.

Kistner, J. (2026). "P7.5 Negative Result: Falsification of the Blackwell-MMQ Hypothesis for a Sovereign llama.cpp Serving Wedge." OSINTelligence LLC. The negative-result companion bundle to this paper.

Kistner, J. (2026). [*Sovereign In-Distribution Cyber Threat Intelligence Named Entity Recognition*](/osintelligence/part-iv-the-evidence-what-worked/14-sovereign-cti-ner.md) (Chapter 14). OSINTelligence LLC. The CTI-NER companion; supplies the 1,002-row A/B harness and the independent §4.4.1 observation of the content-class failures.

llama.cpp upstream (2026). [Issue #20093](https://github.com/ggml-org/llama.cpp/issues/20093), M-RoPE heap over-read in ubatch\_reserve, with a reporter-authored two-part fix (ubatch\_reserve resize + state\_read\_meta broadcast) applied verbatim; [Issue #19915](https://github.com/ggml-org/llama.cpp/issues/19915) (reporter: KernelFreeze), seq\_add() assertion on Qwen 3.5, fixed by Georgi Gerganov (@ggerganov), reviewed by @ngxson, via [PR #19928](https://github.com/ggml-org/llama.cpp/pull/19928) / commit [99bd67c9b](https://github.com/ggml-org/llama.cpp/commit/99bd67c9b29851f758c7d22caa8fc57fd5af3e4f) (cherry-picked); Issue #20345, grammar-constrained decoding and token-budget truncation; release tag b7992 (commit 612db6188); tools/server README (Router Mode flags, accessed 2026-04-14). Project by Georgi Gerganov and contributors, MIT license. <https://github.com/ggml-org/llama.cpp>

LM Studio (2025). Issue #1773, JSON response truncation under constrained decoding. <https://github.com/lmstudio-ai/lmstudio-bug-tracker/issues/1773>

Microsoft Corporation. Windows Error Reporting: Application Error (Event ID 1000). <https://learn.microsoft.com/en-us/windows/win32/wer/> · Miller, M., et al. (2008). *Debugging Heap Corruption with Application Verifier and Pageheap.* Microsoft. Canonical latent-heap-corruption diagnostics.

NVIDIA Corporation (2026). *CUDA Toolkit Blackwell Migration Guide.* Recommends CUDA 12.8 as near-term stable for sm\_120; cited at time-of-hypothesis, retained as valid for its narrower claim.

Ollama project (2025). Issue #14570, Qwen 3 structured-output truncation. <https://github.com/ollama/ollama/issues/14570>

Su, J., Lu, Y., Pan, S., Wen, B., & Liu, Y. (2023). "RoFormer: Enhanced Transformer with Rotary Position Embedding." *Neurocomputing.* Preprint [arXiv:2104.09864](https://arxiv.org/abs/2104.09864) (2021).

Wang, P., et al. (2024). "Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution." [arXiv:2409.12191](https://arxiv.org/abs/2409.12191). Introduces three-dimensional M-RoPE.

Willison, S. (2024). "Grammar-constrained JSON: LLMs and structured output." [simonwillison.net/2024/Aug/23/structured-generation-in-llms/](https://simonwillison.net/2024/Aug/23/structured-generation-in-llms/)

Liu, M., Zhong, S., Bi, W., Zhang, Y., Chen, Zhiyang, Chen, Zhenpeng, Liu, X., & Ma, Y. (2025). "A First Look at Bugs in LLM Inference Engines." *ACM Transactions on Software Engineering and Methodology.* [DOI 10.1145/3788873](https://doi.org/10.1145/3788873); preprint [arXiv:2506.09713](https://arxiv.org/abs/2506.09713). 929 bugs, 5 engines, 28 root causes, 65% Crash prevalence.

**Evidence & seal.** The forensic evidence bundle behind this chapter, the P7.5 cross-toolchain falsification (byte-identical crash signature) and the P7.6 root-cause + patch, with the upstream issues/PR/commit and the zero-crash-in-2,004-rows receipt, is in the record's evidence section: [mRoPE Serving-Path RCA (P7.5 / P7.6)](/osintelligence/evidence-and-seals/mrope-serving-path-rca-p7.5-p7.6.md). (An Admiralty-A1 evidence bundle, not a pre-registered experiment; it carries no canonical seal, as stated 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). *Cross-Toolchain Falsification and Surgical Patch-Pick Resolution of a Production M-RoPE Heap Over-Read in llama.cpp: A Sovereign-Deployment RCA Case Study*, version 1.0.0. OSINTelligence LLC research whitepaper. 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 · mRoPE Serving-Path RCA · Chapter 20 · Part IV · v1.0.0 · License CC BY 4.0 · © Jamey Kistner, OSINTelligence LLC*