Research: reproduce and validate single-token model fingerprinting

[toby-bridges]

Context

Bruckner proposes verifying LLM endpoint identity from distributions of short, one-word responses. This may complement the project's existing model-substitution signals, but it should remain a research candidate rather than provider-level proof.

Validation scope

• Reproduce the published JSD pipeline from the pinned Zenodo dataset and software.
• Check normalization behavior, query budget, cross-provider variance, temporal drift, and hidden system-prompt effects.
• Report match / anomaly / inconclusive; keep riskMatrixImpact=none until independently validated.

Deliverable

A reproducible note with pinned artifact versions, checksums, commands, results, and limitations. Any code or documentation integration should follow in a separate atomic PR after evidence review.

Citations

• Bruckner, T. (2026). One Token Is Enough: Fingerprinting and Verifying Large Language Models from Single-Token Output Distributions. arXiv:2607.10252v1. https://arxiv.org/abs/2607.10252
• Bruckner, T. (2026). Single-token output distributions as behavioral fingerprints of large language models [Data set]. Zenodo. https://doi.org/10.5281/zenodo.21278557
• Bruckner, T. (2026). Single-token output distributions as behavioral fingerprints of large language models — software [Computer software]. Zenodo. https://doi.org/10.5281/zenodo.21278793

[github-actions]

Claude encountered an error after 2s —— View job

---

I'll analyze this and get back to you.