K-Veritas gives machine-learning papers cryptographic receipts for their numbers

Why it matters

Keita is targeting a specific weakness in computational research: reviewers can inspect code and logs without knowing whether those artifacts produced the published result. K-Veritas makes that claim tamper-evident while keeping raw training data local, though conference adoption and independent attestation infrastructure remain essential.

Mamadou K. Keita publicly introduced K-Veritas on August 31, offering researchers a way to attach a cryptographically signed execution record to the numerical claims in a paper. The reviewer can inspect that record without rerunning the experiment or receiving the researcher's training data.

Keita is listed in Rochester Institute of Technology's computing Ph.D. research group, where computer science professor Christopher Homan serves as his advisor. His other research includes speech recognition and machine translation for low-resource African languages. Keita, Homan and collaborators also published SALAN, a speech dataset covering eight languages spoken in Niger.

That work gives Keita a direct view of the trust problem K-Veritas targets. Modern machine-learning papers can depend on long training runs, private datasets, specialized hardware and code that changes between experiments. Reviewers generally receive a table of results and whatever artifacts an author chooses to share. They rarely have the time, compute budget or data access required to reconstruct the run.

"What is written in the paper should be what was truly gotten from the experiments," Keita says on the K-Veritas homepage.

K-Veritas turns that principle into an execution wrapper. A researcher initializes a session, prefixes an existing command with kveritas run, and seals the completed session into a signed PDF. The command-line tool captures source hashes, standard output and error, metrics, environment details, timestamps and hardware activity. The researcher can submit the resulting report beside a paper, while a reviewer checks its signature and recorded claims through a local or web verifier.

A receipt for the run

Keita and Homan formalized the idea in a position paper submitted on May 9, calling the missing property "experiment nonrepudiation." Their definition requires a record that connects a paper's reported number to a specific computation and prevents the author from quietly changing that record later.

The distinction matters because common reproducibility measures answer related questions. Artifact evaluation can establish that submitted code runs. Experiment trackers such as Weights & Biases, MLflow and Neptune preserve metrics and configuration selected by the user. Preregistration records an experimental plan. None inherently establishes that a particular number printed in a paper came from a particular execution.

K-Veritas attempts to create that link by hashing the code and captured execution record with SHA-256. The attestation service receives a digest rather than the raw experiment data, then signs the digest using RSA-PSS-SHA256. The generated PDF carries the verification material, allowing another party to detect changes to the committed record.

The current repository documentation extends the original paper's design with source snapshots, optional checkout bundles and selective-disclosure proofs. Researchers can keep filenames and contents redacted while committing hashes that prevent files from being removed invisibly after the run. An open disclosure mode can include a source bundle, while datasets and model weights remain outside it.

K-Veritas also records CPU, memory, storage and, when available, GPU telemetry. The repository includes HMCA, an execution-coherence heuristic, along with compute-cost checks that compare declared work against physical limits derived from elapsed time, energy and memory.

Keita began the repository's visible release history with an initial CLI commit on March 23, five months before the August 31 public introduction. The repository had reached 74 commits in the supplied August capture, with a recent cleanup committed on August 28. Keita added an Apache-2.0 license for the CLI and protocol and an AGPL-3.0 license for the attestation server on August 27.

The seal has limits

A valid K-Veritas report establishes the integrity of the captured record. It cannot determine whether the experiment was well designed, the code contained bugs, the statistics were appropriate or the author selected a favorable run from a larger set of undisclosed attempts.

The observer also runs in user space. A compromised operating system, malicious firmware or falsified hardware counters could feed it fabricated telemetry.

The authors describe K-Veritas as a testbed. That positioning matters. A verification standard controlled by one research lab, university or vendor would give its operator unusual influence over which experimental records receive a trusted seal. K-Veritas says its intended governance model excludes investment-based control and calls for an independent, community-maintained standard.

Adoption is the harder problem. The paper proposes starting with voluntary badges, moving toward attestations that conferences expect, and eventually requiring them for empirical submissions. Keita and Homan estimate that progression could take three to eight years after initial adoption. Conference integrations, persistent storage, audits, key management and support across research frameworks still have to follow the command-line release.

K-Veritas therefore arrives as working code attached to an institutional argument. Keita wants experimental claims to carry evidence generated while the computation happens, instead of relying entirely on artifacts assembled after the result is known. The tool will earn its place only if researchers can use it without exposing sensitive work and conferences decide that signed execution receipts belong in peer review.