Deterministic Sovereign RAG via Signed-Hash Projection

Formalization of FNV-1a 64-bit + L₂ spherical projection at D=128 for reproducible retrieval on sovereign cloud infrastructure.

Author	David Mireles
ORCID	0009-0008-4374-2254
Contact	[email protected]
Venue	arXiv (cs.IR + cs.DS)
Status	Pre-submission draft, June 23 2026
License	Apache-2.0
Canonical	https://github.com/LOUST-PRO/deterministic-sovereign-rag

What this paper formalizes

A four-formula pipeline for deterministic, sub-millisecond retrieval over sovereign cloud corpora:

1. FNV-1a 64-bit feature hashing into a fixed D=128 vector.
2. Sign-flipping cancellation to eliminate embedding drift across re-indexes (signed-hash projection).
3. L₂ spherical normalization to unit length on the 128-sphere.
4. Cosine similarity for top-k retrieval, with formal bounds from the Johnson-Lindenstrauss lemma and Chernoff concentration.

The architecture is reproducible end-to-end: the same corpus hashes to the same vectors on the same substrate, with bit-identical output. No external embedding model. No GPU dependency. No SaaS.

Repository layout

deterministic-sovereign-rag/
├── README.md              # this file
├── LICENSE                # Apache-2.0
├── CITATION.cff           # machine-readable citation
├── paper/
│   ├── deterministic-sovereign-rag.tex   # canonical LaTeX master
│   ├── f80-references.bib    # BibTeX
│   ├── ieee_arxiv.sty        # IEEEtran-style arXiv preamble
│   ├── sections/             # 8 modular subfiles (\input{})
│   └── figures/              # 10 TikZ source + baked PDFs
├── verification/
│   ├── f80-build.sh          # reproducible 4-pass compile
│   ├── arxiv-allowlist.txt   # 24 paths, binding
│   └── anonymization-lexicon.md   # public-safe IP-boundary summary
├── docs/                    # submission plan, related work, review log
├── companion/               # reproduction notes (pseudocode pointers)
└── .github/workflows/
    └── build-paper.yml      # CI: paper compiles + figures OK

How to build the paper

cd verification
bash f80-build.sh           # 8-stage validation + 4-pass compile

The build script enforces the IP boundary: any content matching the anonymization-lexicon.md patterns is rejected before compilation. The arXiv allowlist restricts the tarball to the 24 paths in verification/arxiv-allowlist.txt — nothing else ships.

How to cite

@misc{mireles2026deterministic,
  author       = {David Mireles},
  title        = {Deterministic Sovereign RAG via Signed-Hash Projection:
                  An Operator-Stack Formalization of FNV-1a 64-bit +
                  L\textsubscript{2} Spherical Projection at D=128 for
                  Reproducible Retrieval on Sovereign Cloud Infrastructure},
  year         = {2026},
  month        = jun,
  howpublished = {arXiv preprint cs.IR / cs.DS},
  note         = {ORCID: 0009-0008-4374-2254},
  url          = {https://github.com/LOUST-PRO/deterministic-sovereign-rag}
}

Companion artifacts

The paper's formalization is implemented in production at the companion repository LOUST-PRO/dsvh-verification-suite. That implementation provides the reference reproducer for the mathematical claims in §3 (preliminaries) and §8 (DET keep-alive).

The gap between the formalization (cosine) and the production runtime (Jaccard) is documented in §12 of the paper (Adversarial: When NOT to Use DET).

Submission status

• arXiv submission: pending. The submission gate requires the operator's approval (per the operator's local-first policy).
• Endorsement: required per arXiv policy for first-time submitters. Plan documented in docs/submission-plan.md (forthcoming).
• Zenodo DOI: a Concept DOI will be reserved after the first arXiv publication, and the manuscript will be updated to reference the canonical citation identifier.
• ORCID linkage: public record at https://orcid.org/0009-0008-4374-2254.

License

Apache-2.0. See LICENSE.