Data and per-query results for the paper Signed but Steerable: Measuring and Closing the Text-Vector Integrity Gap in Offline Retrieval-Augmented Generation (Georgios Feretzakis, Alexandros Papaspyridis, securagen.ai P.C.).

Paper: arXiv:2611.06498 · Hugging Face paper page
Code: github.com/securagenai/signed-but-steerable
Archive: Zenodo release v1.0.3, 10.5281/zenodo.24031109. The paper cites v1.0.2, 10.5281/zenodo.23948839, whose data and results are identical.

Summary

Offline retrieval-augmented generation (RAG) on edge devices is updated through signed knowledge packs that carry text and precomputed embedding vectors. A signature proves who published a pack, not that its vectors encode its text. The paper measures this text-vector integrity gap (the margin law), shows that tolerance-based audits of compressed vector stores can leave queries exposed, and evaluates code-level verification that closes the gap.

This repository contains the data/ and results/ folders of release v1.0.3, together with LICENSE and DATA_LICENSES.md, with two exclusions: the answers generated by the language model (results/e2/gen_*.jsonl) and the outputs of the authors' earlier study (data/prior/). Both are available in the Zenodo record and on GitHub. The 223 data and licence files here are byte-identical to release v1.0.3 (SHA-256 in MANIFEST.json of the release).

data/: the SQuAD 2.0 development set (unmodified), the SQuAD-derived corpus, sentences, windows and questions, and precomputed BGE-M3 and all-MiniLM-L6-v2 vectors.
results/: per-experiment outputs (JSON, JSONL and NumPy files), including the signed test packs in results/e8/packs/. The README on GitHub maps each number in the paper to the file that holds it.

The files have different schemas per experiment, so the dataset viewer is disabled. Files can be loaded directly, for example:

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download("securagenai/signed-but-steerable", "results/e7/beir_margins.json", repo_type="dataset")
with open(path) as f:
    beir_margins = json.load(f)

The scripts that produced and analyse these files are in the GitHub repository.

Licences

Code and our results: MIT.
SQuAD 2.0-derived data, vectors and result files that contain SQuAD text: CC BY-SA 4.0 (SQuAD 2.0, Rajpurkar, Jia and Liang, 2018).
SciFact, NFCorpus, ArguAna and FiQA-2018 (obtained through BEIR) and MS MARCO: only derived numbers and identifiers are included, no text or vectors; each dataset keeps its own terms.

File-level details are in DATA_LICENSES.md, which also describes the files of the full release that are not included here. securagen.ai P.C. used all third-party datasets and models for non-commercial research.

Citation

@misc{feretzakis2026signed,
  title         = {Signed but Steerable: Measuring and Closing the Text-Vector Integrity Gap in Offline Retrieval-Augmented Generation},
  author        = {Feretzakis, Georgios and Papaspyridis, Alexandros},
  year          = {2026},
  eprint        = {2611.06498},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CR},
  doi           = {10.48550/arXiv.2611.06498}
}

Please also cite SQuAD 2.0 (Rajpurkar et al., 2018) and BEIR (Thakur et al., 2021) when you use the corresponding parts.