Citing this work¶

Every repository carries a CITATION.cff, so GitHub's "Cite this repository" button works and tools like Zenodo and cffconvert can read it directly.

One tool¶

Use the repository's own CITATION.cff. On GitHub: Cite this repository → BibTeX.

@software{harris_gridlock,
  author  = {Harris, Nick},
  title   = {gridlock: certify a wait-for relation cannot wedge},
  url     = {https://github.com/nickharris808/gridlock},
  license = {Apache-2.0},
  year    = {2026}
}

The portfolio¶

@software{harris_verification_portfolio_2026,
  author  = {Harris, Nick},
  title   = {A measurement you cannot check is a press release:
             twenty-five verification tools that refuse to guess},
  url     = {https://github.com/nickharris808},
  license = {Apache-2.0},
  year    = {2026}
}

The datasets¶

The Hugging Face datasets carry their own cards with licence and provenance:

abstain-corpus — 32 inputs a verifier must not pass
kv-tenant-isolation-bench — including four rows marked uninterpretable
kv-reuse-econ-traces
llm-precision-fingerprints

If you are citing a negative result¶

Several results here are honest negatives and are more useful cited as such:

a pre-registered null: default vLLM 0.24.0 was exactly batch-invariant under static co-batching on the configuration tested (dynamic batching untested);
a headline of ours that turned out to be an arithmetic identity rather than a measurement;
no speedup claim: a live comparison came out at ≈0.997×.

These are documented in the repositories that produced them, not summarised away.