Status: closing record, written 2026-08-09. The track ran from the first retained entry through to a catalog that holds five AI-platform shapes, each pinned by a digest and guarded by lanes that fail when a claim stops being true. This says what exists, what it proves, what it refuses to claim, and what is deliberately left undone.
The reader-facing entrance is the AICR catalog overview. This page is for whoever picks the work up next.
What exists
Five entries, each with a digest-bound index compiled from committed bytes.
| Entry | Provenance | Ladder rungs with receipts |
|---|---|---|
eks-h100-training-kubeflow | AICR v0.14.0, retained | 2 |
eks-h100-training-kubeflow-v0-18-0 | AICR v0.18.0, retained | 0 |
eks-h100-inference-nim | AICR v0.14.0 with platform: nim, retained | 1 |
kserve-nim-inference | nim-deploy KServe subtree at commit 3ef33472 | 2 |
cpu-starter | derived from the v0.14.0 training entry by recorded rules | 3 |
Behind them: eleven committed receipts, nineteen generated data surfaces, eight hand-authored registers, eight demo pages, three reference documents, and forty-one steps in the verify chain. Every one of those steps runs in CI now, which was not true when the track started.
What it proves
A platform shape can be pinned. One digest covers the upstream source, the recipe criteria, and every rendered Application. Change a byte anywhere and the digest moves.
Provenance can run from an upstream signer to bytes on disk. The v0.18.0 release signature verifies offline against a committed sigstore trust root, its attested subject is recomputed from the two files the catalog retains, and the binary that generated the entry reports the same build commit the signature names. The trust root itself is reviewed against the one sigstore publishes, and it cannot change without that review.
Governed change works on platform configuration. Entries import into ConfigHub, carry reviewed overrides, promote through environments, and deliver to a cluster at the config plane. The blast-radius checker holds each declared control point to the documents it actually reaches, in both directions.
Refusals are evidence. Nine candidate changes run through the shipped commands in a throwaway worktree, and the verdicts are published. Two of them must be accepted, because lanes that refused everything would look identical to lanes that refused the right things.
Counted claims fail when they stop being true. Every number on an AICR page is bound to a quantity computed from committed bytes, and a counted claim nobody declared fails the lane by name.
What it refuses to claim
No GPU workload ran, anywhere, for any receipt in this track. Every proof is config-plane: import, render, digest pinning, variant, promotion, delivery wiring, acceptance by an API server. The boundary is stated in each receipt rather than left to a reader to infer.
No gated artifact was mirrored. Ten gated references are enumerated, one has its terms read and recorded, and the other nine say plainly that their catalog pages sit behind bot detection and were not read. The gap is published rather than filled with a guess.
The v0.18.0 entry has climbed no rung beyond retention, and the evidence record lists it with an empty ladder for exactly that reason.
What the track taught, beyond AICR
Four lessons outlived the entries that produced them, and all four are doctrine or shared machinery now.
A parity check should verify the property, not the shape it took in the version you wrote it against. Upstream moved to parallel deployment, and two of our checks refused an entry for doing what upstream now intends. Both compared against a total order that had never been the claim. Ordering is now checked against the dependency edges the recipe carries.
A rule discovered in one place usually belongs everywhere. The credential guard the inference compiler carried since the NGC license read now applies to every producer in the repository. The gated-artifact rules moved from a planning note into doctrine.
A cadence keyed to a date goes stale quietly; key it to the thing that changes. Gated references are keyed to their exact tag, so a version bump produces a reference nobody enumerated and the lane refuses it. Nothing fails when a date passes.
An unanswered question is not a negative answer. Provenance that could not be asked about is recorded as unknown, never as unsigned. Receipts that carry no date are counted, not assumed fresh.
What is deliberately left
The remaining backlog is real work rather than tidying, and it splits three ways.
Surfacing. Theme 6 is untouched: the Platforms section on the site, the buyer-facing page, an AICR route on the Examples page, and the CPU starter as a runnable first exercise. The evidence exists and is published in site/catalog.json under platformEvidence; what is missing is the reading experience on top of it.
Deeper parity and generation. Field-level parity against platform shapes and the first Pilot generation gated by it are the largest remaining pieces. The refusal corpus and the blast-radius locators are the groundwork; neither finishes the job.
The AICR CLI comparisons. aicr trust, aicr mirror and aicr skill each are now compared against v0.20.0 in the maintained reference. The comparison distinguishes TUF acquisition from reviewed root retention, recipe image discovery from retained chart locks, and CLI skills from chart playbook selection. Full mirror discovery and authenticated trust-root acquisition remain separate work.
Two smaller items stay open with reasons rather than dates. The per-artifact terms for nine gated references need a person to read nine catalog pages. NIM_TELEMETRY_MODE needs a structural editing path that ConfigHub's single-token substitution does not offer today, which is a product gap rather than an entry gap.
One thing worth fixing early
The credential-boundary lane hashes every document in its scope, so any concurrent merge that adds one makes it stale. It tripped three times in a single afternoon for that reason alone. Scoping it to a manifest would keep the guarantee and drop the fragility, and whoever touches it next should do that rather than rebase around it.
Where the record lives
- The catalog overview, for a reader.
- The task backlog, with the progress log of what each increment found.
- The platform evidence record, which names every rung climbed and every rung not.
- The refusal corpus and the claim-integrity lane, which are how the pages stay honest as they age.