Status date: 2026-09-15
LayerCake has a sealed scientific snapshot and a newer host-development branch. They must not be presented as one artifact lineage.
| Surface | Identity | Status | What may be said |
|---|---|---|---|
| Sealed research release | layercake-moonshot-final at 0537cbb9e93cd7ebd4ba01c0bf641414ecebb1c3 |
Locally revalidated; immutable release boundary | Only the bounded claims in CLAIMS.md and docs/VERIFICATION_AND_LIMITS.md |
| Default branch | Post-release HEAD | Unsealed development lineage | Mechanical construct results backed by their individual decision records |
| Latest route-isolated host construct | lc-direct-neural-core/25 |
Construct-certified | Five physically isolated rank-16 routes, selected-route execution, signed-package lifecycle, and fixture CPU/CUDA identity |
| Authoritative-destination extension | Post-release additive extension | Construct-certified | Immutable outer destination selects only its registered package and fails closed otherwise |
| External English artifact | None accepted into a successor release | Open | No external artifact quality, speed, memory, TTFT, or acquisition claim |
| Foreign-teacher acquisition | Separate ABI repository | Outside LayerCake | No ABI research result is inherited here |
Post-release changes touched registered package/runtime components. The frozen
invalidation matrix therefore rejects inheritance of the old
precision_contract and dependent Phase 2-8 seals. This is expected behavior,
not evidence that the immutable tag regressed.
Use the exact tag for release verification. Use current HEAD for development
and its narrowly scoped construct tests. A future successor release must run
the governed recertification required by moonshot/invalidation_matrix.yaml.
Exact-lineage evidence tests skip their positive replay on an incompatible
development checkout and instead require the stale evidence to fail closed.
Phases 0, 1, 2, 4, 5, 6, 7, and 8 have sealed scientific results for the exact
release lineage. Phase 3 is sealed as RETIRED_BY_GOVERNANCE; its original
faster-training objective was retired with zero headline training-efficiency
claims. It was not skipped, and it must not be described as proof of faster
training or external capability acquisition.
See Phase status for the evidence map.
The evidence certifies the exact local tag and retained assets described by the campaign. It does not itself assert that every tag, checkpoint, or release asset is published remotely. Users should distinguish source availability, artifact availability, local reproducibility, and independent reproduction.
The post-release host constructs need one real, independently validated, immutable external English artifact. The same final payload must then pass output conformance, English quality, collapse, CPU/GPU, throughput, TTFT, memory, package-identity, and hostile-verification gates before any successor release claim changes.