Reproducing the record.
The record is published so it can be attacked. These answers cover where the paper and the data live, how to check nothing was edited, how to re-judge the transcripts with your own panel, and what result would falsify the claim.
The paper
Where can I find LLM agent safety testing and auditing?
Gubernaut publishes its own safety and auditing artifacts rather than describing them. The engineering receipts folder in the public repository holds the raw run artifacts for the fail-safe battery, the concurrency isolation run, the edge soak and the ablation, published unedited apart from stripped local paths. The fail-safe battery is published together with the round that failed it: the pre-hardening run recorded 4 fail-open leaks in Gubernaut's own code and the hardened run records 0. Four reproduction levels let you re-run the record yourself, and the first three need no API key and cost nothing.
Source: HARDTEST v1.0 release gates, 2026-07-24.
Where can I read the paper?
The paper is on arXiv as 2607.24339, cross-listed cs.AI, cs.CL, and archived on Zenodo under DOI 10.5281/zenodo.21303518. Both carry the same text, and the Zenodo record also holds the complete validation release.
Which DOI should I cite?
Cite 10.5281/zenodo.21303518, the concept DOI. A concept DOI resolves to the newest version of a Zenodo record, so a citation written today still lands on the current release. Version-specific DOIs exist inside the archive and are not used in published copy.
Is the paper peer reviewed?
The paper is a preprint on arXiv and has not completed journal peer review. The evaluation was pre-registered before data collection, and the transcripts, judge panels and analysis scripts are public, so the result can be checked directly rather than taken on the strength of a review.
How do I cite Gubernaut?
Cite the paper by its concept DOI, 10.5281/zenodo.21303518, or by its arXiv identifier 2607.24339. A ready BibTeX entry sits on the research page of this site, and the repository carries a CITATION.cff file that reference managers read directly.
What licence covers the paper and data?
The paper and the validation release are CC-BY-4.0, so reuse and redistribution are permitted with attribution. Software carries Apache-2.0 instead. Figures on this site are CC-BY-4.0 and ship machine-readable licence metadata alongside them.
Who wrote the paper?
Dushyant Sharma at Gubernaut Research, ORCID 0009-0007-6534-0347. Gubernaut Lab is an independent lab, and the same author wrote the controller, ran the evaluation and published the record.
The data
What is in the data release?
The release holds every endurance and evaluation transcript for four frontier families in both arms verbatim, all judge panels with SHA-256 provenance, the sealed matrices, tidy per-turn telemetry tables, and the extraction, statistics and figure scripts that regenerate every published number.
Where is the data hosted?
The sealed release is archived on Zenodo under DOI 10.5281/zenodo.21303518 and mirrored in the public GitHub repository thegubernaut/Gubernaut_Validation. Zenodo is the citable archive of record; GitHub is the working copy that carries the scripts.
How do I know the transcripts were not edited?
Each transcript and judge panel carries a SHA-256 hash recorded in the sealed release, so any later edit changes the hash and becomes visible. Verifying is a local hash of the file against the manifest, which requires no trust in this website.
Can I re-judge the results myself?
Re-judging is the intended path. All transcripts and four-family judge panels ship with SHA-256 provenance, and the extraction scripts that turn raw panels into every published number are in the public repository. A different panel or a different rubric can be run against the same replies.
Can I reproduce the spend numbers?
The spend figures are outputs of a scoring script rather than hand-tallied estimates, and the receipts harness is part of the published record. Running the same battery against your own keys produces your own prices, and the governed-to-ungoverned ratio is the portable part of the result.
Source: Pre-registered receipts benchmark, 2026-07-19. Scored output of harness/report.py.
Method and correction
Is the evaluation pre-registered?
The evaluation was pre-registered: the design, the scoring protocol and the analysis were fixed and written down before data collection began, and the pre-registration documents ship inside the public release. Every fix to a documented failure mode was likewise declared before it was written.
What result would falsify the claim?
A fresh panel scoring the published transcripts and finding the regulated arm no calmer would falsify the central claim, as would a governed run that fails to hard-stop the loop battery, or a replication in which recovery does not return to baseline. The materials to attempt each of those are published.
What happens if I find an error in the record?
Report it as a public issue on the validation repository, or write to the lab directly. Corrections are made in the open with a dated note, and the published numbers on this site trace to a single register file, so one correction propagates to every page at once.
The other sets
These four sets divide the same subject. The short form of all of them is on the questions hub.
- Agent guardrailsWhat the governor does when an agent starts looping, what it leaves alone, and where its resistance stops. Every answer states its scope, because a guardrail claimed wider than it was measured is worse than no guardrail.18 questions
- Proxy and installGubernaut ships as five packages and adopts in one configuration line. These answers cover the install itself, what it is compatible with, and what it costs you in latency and operational surface.14 questions
- The measurementsEach figure Gubernaut publishes, with the qualifier it is only true under. Where a measurement came back null or inconclusive, the answer says so in the same paragraph as the headline.14 questions