Verify
Check our work.
You should not have to trust us.
Most supply-chain assurance asks you to believe a vendor questionnaire. We publish the auditor, publish what it found, and hand you the command that regenerates the result on your own machine. No account, no upload, and no network call, because the tool contains no network code at all. That is the property it exists to demonstrate, and you can check it by running the tool on itself.
This is deliberately not a demonstration you have to book.
git clone https://github.com/EntroLabs/EntroVouch && cd EntroVouchApache-2.0. Pure Python standard library: no dependencies, and no network code of its own. Python 3.11 or later, and nothing to install.
python examples/regenerate.pyRegenerates every report in examples/reports from the fixture tree beside it: the no-egress audit, its SARIF and in-toto forms, the cryptographic bill of materials, its CycloneDX 1.6 form, the key-provenance report (which has no findings digest: re-run it and read its findings), and the SBOM with its CycloneDX form.
compare one value: findings_digestIt is bit-identical across runs, processes and machines. If yours matches ours, the tool on your machine produced exactly what we published. The expected values are printed in the repository, beside the reports themselves.
What we publish that vendors usually do not
The repository carries a conformance file stating, for every standard and format we name, what backs the name: a test against that standard's own schema or vectors, or a plain statement that conformance is not claimed. Reports are signed by one signer, a hash-based Lamport-Merkle construction over SHA3-256. It claims conformance to no published standard, so there is no official vector set for it to fail. Separately, the repository ships an RFC 8554 implementation that verifies LMS and HSS signatures and makes LMS signatures; it is checked against the RFC's own test vectors and does not sign reports. It also verifies the RFC 9858 parameter sets and gives the published answer on NIST's ACVP LMS vectors, and the repository keeps an issuance log in C2SP tlog-tiles form that a buyer can check offline.
The repository also publishes how often each tool is wrong. The tools were run on public repositories drawn from search results and pinned to a commit. For every distinct claim they made, at least one site was read at its source lines, and the false ones are counted: none of 885 findings for the egress audit, none of 249 for the cryptographic inventory, 15 of 51 for key provenance. Findings that could be read either way are counted separately; counting them as wrong lowers key-provenance precision from 70.6% to 25.5% (and egress from 100% to 99.66%, inventory from 100% to 97.19%).
The content hash will NOT match, and that is correct.
It covers the issuance timestamp, so two honest runs produce two different values. A reader who compares it, sees a mismatch and concludes the report was forged has followed the most obvious instinct and reached the wrong answer. All three digests are printed on the report itself, each labelled with whether it reproduces.
Static analysis is detection-grade, not proof.
It audits the source you point it at. It cannot see egress introduced by a dependency it was not pointed at, by dynamic loading, or by a compiled extension. Detection is list-based: a third-party network library that is not on the list produces no finding, and the report names it as unlisted instead. Every report says this in its own body rather than leaving you to find it.
A signature proves origin and integrity, never correctness.
With a pinned public root it proves a report came from us and was not altered. It says nothing about whether the verdict is right. Omitting the root returns UNVERIFIED rather than ATTESTED, because a well-formed signature is not evidence of who wrote it.
Some findings in our own repository are deliberate.
The example fixture contains a real network import so the demo has something to find, and the test programs and the test workflow reach for the network on purpose. The shipped package, entrovouch/, audits clean for egress; the cryptography and key reports list the search strings the package itself holds. We did not teach the auditor to skip those folders, because a scanner with a built-in blind spot is what this tool exists to argue against.
The publisher's public root
Pin this value to check who issued a signed report. A customer's engagement letter carries the copy to pin; this page and the repository's README are convenience copies, and anyone who can change either can change the value.
eb309717634b3a9bd952aac903538bd9b3350d1b006f5176d1c87763ba87a96bThe repository holds one signed report, the same audit of the example tree. This prints ATTESTED; without the root it prints UNVERIFIED, and after any edit, TAMPERED.
python -m entrovouch.verify examples/signed/egress.json --root eb309717634b3a9bd952aac903538bd9b3350d1b006f5176d1c87763ba87a96b