Proof and documentation
Verifier: Open-Source ML-DSA-44 Certificate Verifier
What the open-source certificate verifier recomputes, the standardized rules it checks, the sample verdict and self-test behind it, and the honest limits stated plainly.
What it is
The Verifier is an open-source reference tool, written in plain Python over a reference ML-DSA-44 implementation, that validates a post-quantum certificate. It does more than call a library's verify() and trust the answer. It decodes the ML-DSA-44 signature per FIPS 204 and recomputes the public acceptance rules anyone can independently check: the signature encoding, the norm bound on z, and the hint density, along with identity-digest recomputation, issuer trust, validity window, revocation, and policy digest. It is downloadable so a reviewer can run it against their own certificates. It holds FAIL (contradictory evidence) separate from INCONCLUSIVE (missing evidence), which is the design choice that keeps a verifier from failing open.
What this proves to you (plain English)
You do not have to take our word that a certificate is valid. This tool recomputes the public math from the signature bytes and shows its work, so your own reviewer can confirm the result independently. Just as important, it is honest about uncertainty: when it is not given enough information to establish issuer trust, it says INCONCLUSIVE rather than quietly passing or failing. A verifier that recomputes the rules and refuses to fake a trust decision is one you can audit, which is the whole point of shipping it open source.
Standards and parameter set
- Signature scheme: ML-DSA-44 (CRYSTALS-Dilithium), NIST FIPS 204, security category 2. Public key 1,312 bytes, signature 2,420 bytes. FACT (FIPS 204).
- Signature structure decoded: c_tilde (32 bytes) concatenated with z and the hint h, totaling 2,420 bytes. FACT (FIPS 204 sigDecode).
- Public acceptance bounds recomputed: norm bound gamma1 - beta = 130,994 (gamma1 = 131,072, beta = 78) and hint weight at most omega = 80. FACT (FIPS 204 constants for ML-DSA-44).
- Context: NIST standardized ML-DSA in FIPS 204 (August 2024). Federal post-quantum migration is directed under EO 14412 and OMB M-26-15. FACT (public record).
Correctness evidence (concrete, verifiable)
Positive case (SAMPLE run on the bundled sample certificate):
- Parameter set matched: ML-DSA-44. FACT (demo output).
- Signature verification: ACCEPT, reason code none. ML-DSA-44 verifies over the recomputed identity digest. FACT (demo output).
- Signature encoding: PASS, 2,420 bytes, c_tilde | z | h decoded. FACT (demo output, matches FIPS 204 size).
- Norm bound: PASS, ||z||inf = 130,863, below the bound gamma1 - beta = 130,994. FACT (demo output; bound is a FIPS 204 constant). This figure is from one run on the sample certificate.
- Hint density: PASS, hint weight 60 of omega 80. FACT (demo output; omega is a FIPS 204 constant).
- Identity match: TRUE, the recomputed canonical identity digest equals the one bound in the certificate. FACT (demo output).
- Overall result on the sample: INCONCLUSIVE, because no issuer trust anchor was supplied. The signature is cryptographically valid, but the tool declines to assert a full trust decision it cannot support. FACT (demo output). This is the tool behaving correctly, holding FAIL apart from INCONCLUSIVE.
Negative case (the rejection battery, tamper leads to a correct non-accept):
- The tool ships a self-test that exercises each rejection reason and confirms each returns the right outcome: valid certificate passes; a tampered subject returns identity_mismatch (fail); an untrusted issuer key fails; a policy-digest mismatch fails; an expired or revoked certificate fails; a zeroed or truncated signature returns fail or malformed; a violated norm bound or hint density fails; and a missing trust anchor returns INCONCLUSIVE, not fail. Result: 20 of 20 vectors pass, with FAIL and INCONCLUSIVE held apart. FACT (demo self-test output).
Honesty caveats stated plainly:
- The verdict on the page is a real run of the open-source tool on an included sample certificate. It is a SAMPLE run, not a live in-browser verification. In-browser verification is a disabled affordance in this build and is flagged as such; the tool runs locally from the command line.
- This demo does NOT compare against the published NIST Known-Answer-Test (KAT) / ACVP vectors. The battery above is the tool's own 20-vector self-test, not a NIST KAT run. We do not claim KAT agreement here.
- The page makes no security claim beyond what the tool computes on the sample input. No certification, assurance level, or production-readiness is asserted.
Measured performance
Omitted. This demonstration does not report measured timing or throughput numbers, so none are claimed here.
Client FAQ
1. Why not just trust a library's verify()? A wrapper around verify() only tells you a library returned true. This tool recomputes the public acceptance rules (encoding, norm bound, hint density, identity digest) from the signature bytes, so your own reviewer can confirm the result independently.
2. What is the difference between FAIL and INCONCLUSIVE? FAIL means the evidence contradicts the certificate (for example a tampered subject). INCONCLUSIVE means evidence is missing (for example no issuer trust anchor was supplied). Collapsing the two is how a verifier silently passes something it should not. This tool keeps them separate.
3. Why did the sample come back INCONCLUSIVE if the signature is valid? Because no issuer trust anchor was provided. The signature itself verifies, but the tool will not assert a full trust decision without the anchor. Supply one and the result resolves.
4. Can we run it ourselves against our own certificates? Yes. It is open source, pure Python over a reference ML-DSA-44 library, and downloadable with a sample certificate. Run the self-test, then verify your own files locally.
5. Does it leak anything sensitive? No. It emits only public validation evidence. It does not emit seeds, raw entropy, rejected candidates, timing traces, or generator internals.
6. Is the in-browser verify button live? No. In-browser verification is intentionally disabled in this build and flagged. It requires a reviewed, constant-time WebAssembly build of ML-DSA-44, which is gated. Until then, run the downloadable tool locally.
See it run