> 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/19-watcher-kl-drift-floor/references-and-provenance.md). # References & provenance ### References \[1] Xiong, Y., & Xie, X. (2026). "OPLoRA: Orthogonal Projection LoRA Prevents Catastrophic Forgetting during Parameter-Efficient Fine-Tuning." *AAAI 2026*; [arXiv:2510.13003](https://arxiv.org/abs/2510.13003). Body-verified. \[2] Hoy, W., & Celik, N. (2025). "STABLE: Gated Continual Learning for Large Language Models" (a gated continual self-editing framework). [arXiv:2510.16089](https://arxiv.org/abs/2510.16089). Body-verified. \[3] Biderman, D., et al. (2024). "LoRA Learns Less and Forgets Less." *TMLR*; [arXiv:2405.09673](https://arxiv.org/abs/2405.09673). Abstract-verified; body-fetch gate noted before any module/rank-specific citation. \[4] Wortsman, M., et al. (2022). "Model Soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time." *ICML 2022*; [arXiv:2203.05482](https://arxiv.org/abs/2203.05482). \[5] llama.cpp (2024–2026). ggml-org/llama.cpp, llama-server --lora adapter overlay (all six Gate-D evaluations). \[6] Hugging Face PEFT (2024–2026). LoRA developer guide (all training-side adapter construction). \[7] kalomaze (2023). [llama.cpp Discussion #4110](https://github.com/ggml-org/llama.cpp/discussions/4110), KLD as quantization metric; community substrate for the H4 protocol. \[8] Cochrane Collaboration / AllTrials. Pre-registration discipline precedent, adapted to engineering arcs. \[9] Kerr, N. L. (1998). "HARKing: Hypothesizing After the Results are Known." *Personality and Social Psychology Review* 2(3):196–217. The concept-anchor for the corrigenda discipline. **In-series companions:** [*The Drift Taxonomy*](/osintelligence/part-iii-the-evidence-what-broke/9-the-drift-taxonomy.md) (the cascade catalogue) · [*The Guard Changes at 23:26Z*](/osintelligence/part-iii-the-evidence-what-broke/10-the-guard-changes-at-23-26z.md) §10.17 (the watcher-firmware thesis) · [*Sixteen Practices*](/osintelligence/part-ii-the-discipline/5-sixteen-practices.md) §5.37 (the methodology-tier landing) · [*The Sovereign Triad*](/osintelligence/part-i-the-architecture/1-the-sovereign-triad.md) (the Governor role this daemon instantiates at within-session granularity) · [*Sovereign Imatrix Calibration*](/osintelligence/part-iv-the-evidence-what-worked/15-sovereign-imatrix-calibration.md) (the imatrix technique behind checkpoint 5) · [*Sovereign IOC Classifier*](/osintelligence/part-iv-the-evidence-what-worked/13-sovereign-ioc-classifier.md) (the trainer-template provenance: the classifier trained here descends from the first sovereign micro-agent's recipe, selected as the canonical template by operator decision on 2026-05-22, carrying the same rank-16, alpha-32, seven-projection LoRA geometry and the same causal-LM-plus-grammar output discipline onto a 2B base). **Evidence & seal.** The pre-registration behind this chapter, the frozen H1-H6 hypothesis battery reproduced verbatim with the canonical self-hash quoted from its sidecar (invariant across two ratified §99 corrigenda), is in the record's evidence section: [Watcher L4 Cascade Classifier (cycle22)](/osintelligence/evidence-and-seals/watcher-l4-cascade-classifier-cycle22.md). ### AI-assistance disclosure Large language models were used as research tools in the preparation of this chapter: Claude Opus 4.7 (Code; scaffold and §1 through §4.1); Claude Opus 4.7 (web; relayed peer review: 3 corrections + 4 reframes); Claude Opus 4.8 (completion pass: §3.2, §5, §6). 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). *An Architectural-Inherent KL-Drift Floor in Narrow-Classification Fine-Tuning of a 2B Mamba2-HYBRID Model: Six-Checkpoint Targeted-Attack Characterization*, version 1.0.0. OSINTelligence LLC. **License:** CC BY 4.0 (text). Code and data artifacts MIT per repository license. **Corresponding author:** Jamey Kistner, , OSINTelligence LLC (Columbus, OH). *** *The Sovereign Stack · Watcher KL-Drift Floor · Chapter 19 · Part IV · v1.0.0 · License CC BY 4.0 · © Jamey Kistner, OSINTelligence LLC*