The decision
The catalog verifies upstream AICR signatures with cosign 2.6.1, run from its published container image, against a sigstore trust root committed in this repository, with --use-signed-timestamps, and with the container's network disabled. It does not use the vendor's own verification command, and it does not install cosign on any machine.
Why each part of that
Cosign 2.6.1 specifically. Older cosign cannot read these bundles at all. Version 2.4.1 refuses the release's transparency-log entry with nil value in transaction log entry, with and without --insecure-ignore-tlog, because it expects a log entry shape this release does not use. Version 2.5.3 reads the entry and then fails leaf-certificate verification against its default trust root. Version 2.6.1 verifies the same bundle without any insecure flag. The version is therefore load-bearing and is pinned in the receipt.
Signed timestamps. AICR's release signature is Rekor v2 era material: its log entry is a hashedrekord version 0.0.2 carrying an inclusion proof, and it has no integrated time. Cosign needs a trusted time to check the signing certificate was valid when it signed, and without one it stops with threshold not met for verified log entry integrated timestamps. The bundle carries an RFC 3161 timestamp from sigstore's timestamp authority, and --use-signed-timestamps tells cosign to use it. That flag is what makes the verification work, and it is not a weakening.
A committed trust root and no network. Passing --trusted-root with a committed copy of sigstore's trust root removes the last network dependency, so verification runs with --network none. This matters twice over: it makes the check reproducible years from now, and it matches the air-gapped story upstream added in the same release. The trust root is the one input that ages, so refreshing it is a deliberate, reviewable change rather than a silent fetch.
A container rather than an install. The verifier is a pinned image reference, so the exact verifier is recorded in the receipt and nobody has to install anything to reproduce the result.
Not the vendor command. The aicr binary can verify its own artifacts, and using it would mean running a freshly downloaded vendor binary to check that same vendor's signature. The circularity is the problem, not the vendor. An independent verifier separates release verification from the vendor binary being evaluated, and stays available for other projects. This does not make our trust-root acquisition stronger: the review lane compares a raw HTTPS download and does not authenticate TUF metadata. AICR itself supports TUF-authenticated updates; see the command comparison.
What the lane proves
npm run aicr-signature:run performs the verification and writes a receipt. It proves three things:
- The signature over the release's recipe-catalog attestation verifies offline against pinned trust material.
- The signer is the tagged release workflow in NVIDIA/aicr, identified through GitHub Actions OIDC, and the statement carries a SLSA provenance predicate naming the repository, organization, runner, and build commit.
- The check can fail. The same command is run again with a different signer identity and must be refused, and the receipt records that refusal.
npm run aicr-signature:verify re-checks the committed receipt against the committed bundle without cosign, a container, or a network, so the ordinary verify chain stays free of this toolchain.
The subject is bound to bytes we hold, since 2026-08-08
This section previously said the opposite, and the change is worth stating plainly rather than quietly editing.
The signature covers a subject named recipe-catalog. Upstream computes that subject in pkg/recipe/catalog/digest.go as a length-prefixed SHA-256 over two files, the component registry and the validator catalog:
sha256( u64be(len(registry)) || registry || u64be(len(catalog)) || catalog )The length prefixes make the encoding injective, so two different splits of the same bytes cannot collide. Both files are now retained beside the bundle, and the lane recomputes that digest over the retained bytes on every run. It reproduces the attested subject exactly, and the lane refuses if it ever stops doing so.
That upgrades the receipt from a statement about a signer to a statement about an artifact this repository holds. An auditor can repeat the computation with nothing but the retained files and a hash function.
What it still does not prove
The two bound files are the whole of what the signature covers. The rest of the recipe tree, including the overlays and mixins that decide what a platform contains, is outside the attested subject, so this receipt says nothing about them. That is upstream's scope decision rather than ours, and it is worth knowing before treating the signature as covering the recipes themselves.
The receipt also concerns the upstream v0.18.0 release. The catalog still retains v0.14.0, which shipped no signature at all, so no retained entry carries a signature claim today.
The trust root is the one input that ages, and it is now reviewed
Passing a committed trust root is what removes the last network dependency, and it is also what makes that file the part of this lane that goes out of date. Sigstore publishes updates to its trust root, and our copy is a snapshot of one moment.
The refresh trigger is drift rather than expiry. Every active entry in the committed root carries a start date and no end date. Only two entries have ended, both years ago, and they are kept so signatures made while they were valid still verify. There is no deadline to count down to, so a lane that counted one down would be inventing it.
npm run aicr-trust-root:verifynpm run aicr-trust-root:run compares the committed trust root against the one sigstore publishes and records the result either way. At the review on 2026-08-08 the two were byte-identical, so the committed copy is not a local variant of anything. The verify lane is offline and refuses when the committed trust root has changed since the review that recorded it, which makes replacing trust material a reviewable step rather than a diff someone might scroll past.
It deliberately does not check the trust root against the network. That would put a network dependency back into the ordinary verify chain, which is the thing the committed copy exists to remove. The review record lists what the trust root contains and what it retains for older signatures.
What this unblocks
The refresh brief can proceed with signature verification as the new rung a second retained entry would carry, and the cost of that rung is now known rather than estimated. The binding step it was waiting on is done, so a v0.18.0 retained entry would inherit a provenance chain that already runs from the signer through the attested subject to bytes on disk.