The boundary worth studying is the one the digest crosses. The CLI hashes the source and the captured execution record with SHA-256, passes the digest to the attestation service, and the service signs that digest with RSA-PSS-SHA256, so the experiment itself never leaves the researcher's machine [8]. Datasets and model weights stay outside even the open disclosure mode, which at most carries a source bundle [10]. Selective-disclosure proofs push on the obvious hole: filenames and contents can stay redacted while their hashes are committed, which stops a file being removed after the run without the record showing it [10]. The delivery vehicle is a signed PDF, which is the one artifact a program committee is guaranteed to be able to open.
Keita's own research is speech recognition and machine translation for low-resource African languages, including SALAN, a speech dataset covering eight languages spoken in Niger, and he works in Rochester Institute of Technology's computing Ph.D. group under Christopher Homan [3]. That is a setting where "just publish the corpus" is not always available, which is visible in the design. On the project homepage he puts the goal plainly: "What is written in the paper should be what was truly gotten from the experiments" [9].
The narrowness is the point. Artifact evaluation shows that submitted code runs, trackers such as Weights & Biases, MLflow and Neptune preserve the metrics and configuration the user chose to log, and preregistration files a plan; none of them establishes that the number printed in the paper came out of a particular execution [7].
Where it stops is documented, and the documentation is honest about it. A valid report establishes the integrity of the captured record, and nothing more: not whether the design was sound, whether the code had bugs, whether the statistics were appropriate, or whether the author sealed the most favorable run out of a larger set [14]. Against a wrapper being handed a fabricated record, the repository leans on heuristics rather than proofs. It captures CPU, memory, storage and, where available, GPU telemetry, ships an execution-coherence check called HMCA, and adds compute-cost checks that compare declared work against physical limits derived from elapsed time, energy and memory [11]. Those raise the price of manufacturing a plausible record; they do not turn a signature into a statement about the science.
On maturity, the arithmetic is small. The visible release history opens with an initial CLI commit on March 23 and stood at 74 commits in the supplied August capture, with a cleanup on August 28 [12]. From that first commit to the August 31 introduction is 161 days, which works out to about 0.46 commits a day [15]. That is one person's side project pace, and the load-bearing parts for a reviewer are the protocol and the verifier rather than the feature list.
My read, in my context: the wrapper is cheap enough to try on runs I already instrument, because the marginal cost is a session init and a command prefix [5]. The claim a seal supports is small and checkable, which is why it is useful. This number came out of this execution, and the record has not moved since.
