Skip to content

Latest commit

 

History

320 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

LayerCake

Release-tag status: SEALED. Current post-release HEAD: UNSEALED.

The eight-phase LayerCake campaign is locally sealed at the annotated Git tag layercake-moonshot-final (0537cbb9e93cd7ebd4ba01c0bf641414ecebb1c3). The tag is the release boundary. The canonical campaign state, raw evidence, and machine validation are under moonshot/ and results/moonshot/.

The immutable tag was locally revalidated again on 2026-08-16: 602/602 tests passed, verify-all reported every phase valid, and the final verifier returned PROVEN. See moonshot/postrelease_sealed_tag_local_revalidation_result_v99.json.

Current HEAD is a separate post-release host lineage. Commits 8f49fe1 and 662c5a9 changed layercake/cake/**, which the frozen invalidation matrix classifies as the precision_contract from Phase 2 onward. The current tree therefore correctly fails the old sealed verifier and has not earned a successor Phase 2-8 seal. See moonshot/postrelease_current_head_seal_invalidation_v96.json. Run sealed verification from a detached layercake-moonshot-final checkout, not from post-release HEAD.

All phase lifecycle entries are sealed. Phase 3 has the explicit disposition RETIRED_BY_GOVERNANCE: it preserved the bounded training controls and retired the original faster-training requirement with zero headline scientific claims. It was not skipped, but it also must not be described as a training-efficiency or host-certification proof.

Full-core training speed is a separate, currently open gate. This release does not claim faster full-core training or training dominance.

LayerCake is a capability host: an English core executes selected, signed, non-executable capability packages over a byte-facing interface. It is designed for fast CPU and GPU inference, immutable packages, persistent incremental state, selected-only execution, lossless package transfer, and safe dynamic routing. ABI extraction is a separate project and is not bundled into this repository or release.

Foreign-teacher capability acquisition is maintained separately in the ABI repository. External ABI research does not change this sealed release; see the artifact handoff boundary.

Post-release host work has construct-certified a Unicode-safe external English- core interface, lc-direct-neural-core/2. It is not part of the sealed Phase 0–8 product and has not accepted a real external English artifact. V1 remains historical after a local UTF-8 conformance failure. See the v2 construct report.

The latest post-release execution boundary is the construct-only lc-direct-neural-core/25 signed route-isolated host. It preserves V24's allocation-bounded activation and original four residual routes while adding exactly one physically isolated clarification route. Clarification alone maps to route 4; all other capability mappings remain unchanged, and no token can activate more than one residual route. The construct passes CPU/CUDA fixture identity, strict signed-package and storage-adoption checks, 689 tests, and the unchanged sealed verifier. No real external V25 artifact has yet earned English quality, information-minimum, speed, memory, or TTFT certification. Earlier post-release hosts and their negative evidence remain preserved.

The separately versioned authoritative-destination extension also passes its bounded mechanical construct: an immutable outer destination selects only its registered package, while missing, unknown, and quarantined destinations fail closed without falling through to the English core. Prompt text cannot override that outer label. This additive extension leaves the sealed routing component byte-identical, passes 9 focused and all 693 repository tests, and does not inherit any external acquisition, quality, semantic-purity, or performance claim. See postrelease_authoritative_destination_control_decision_v95.json.

Certified release scope

The final independent clean-room verification reran the sealed Phase 7 product from a detached checkout and passed every Phase 8 gate.

Area Certified result
Campaign lifecycle Phases 0 through 8 are sealed and verify as one ordered campaign.
Clean-room regression 597 tests passed in the detached Phase 7 checkout.
Functional quality suite LayerCake passed 100/100 CPU and 100/100 GPU prompts; the locked Qwen comparator passed 23/100 and 24/100.
Quality uncertainty Paired bootstrap 95% lower bounds for the LayerCake-minus-Qwen success delta were +0.68 on CPU and +0.67 on GPU.
CPU comparison 9.91x LayerCake/Qwen output-byte throughput; LayerCake median latency ratio 0.0179x.
GPU comparison 8.25x LayerCake/Qwen output-byte throughput.
CPU versus transformer GPU 7.67x output-byte throughput; LayerCake median latency ratio 0.0222x.
Domain retention 384/384 held-out cases passed on CPU, 384/384 on GPU, with 384/384 identical outputs.
Core safety 100/100 core-only abstentions; no receiver training or calibration.
Package portability Three fresh hosts installed, verified, removed, and reinstalled identical package bytes.
Routing and catalog 1,980/1,980 routing rows passed; the management catalog exercised 500 entries.
Adversarial verification 32 hostile checks across 24 categories reached their intended boundaries and were resolved.

These are deliberately bounded claims for the exact sealed hardware, runtime, packages, checkpoints, data, and Qwen deployment digest. They do not establish physical mobile performance, calibrated energy dominance, GPU-training dominance, latent neural fusion, external-laboratory independence, or faster from-scratch English acquisition.

Phase map

Phase Objective Release status
0 Permanent governance and campaign state Sealed
1 Benchmark truth Sealed
2 Matched-quality CPU core performance Sealed
3 Governed retirement of the original training-efficiency objective and preservation of its controls Sealed as RETIRED_BY_GOVERNANCE; no training-efficiency or host-certification claim
4 One useful lossless portable Python capability Sealed
5 Generic multi-domain extensibility Sealed
6 Safe routing and catalog scalability Sealed
7 Integrated CPU, GPU, and CPU-versus-GPU evidence Sealed
8 Independent hostile verification and release Sealed at layercake-moonshot-final

See the phase-status document for evidence anchors and the governing tags.

Verify the release

Use the declared Python 3.10 environment from a detached checkout of layercake-moonshot-final. These commands are read-only except for normal temporary test caches:

C:\Python310\python.exe -m layercake.moonshot_campaign verify-sealed 8
C:\Python310\python.exe -m layercake.moonshot_campaign verify-all
C:\Python310\python.exe -m layercake.moonshot_final verify

verify-all must report completed_phases_valid: true. V99 records the latest such exact-tag run. Do not expect these sealed commands to pass from current post-release HEAD until a successor recertification closes V96. The final tag is local in this checkout; remote publication is intentionally not claimed by the verifier.

Post-release external English cores may target the construct-certified lc-direct-neural-core/4 interface when their declarative tokenizer requires normalization and sequence decoding. V4 preserves the UTF-8 byte boundary, shifts external token IDs above LayerCake's four reserved actions, prohibits pointer actions, and performs no receiver learning. This is a hosting interface only; an external artifact must independently prove English quality and performance before LayerCake can inherit either claim.

The construct-certified lc-direct-neural-core/5 interface is available for portable token-plan cores that need raw UTF-8 BPE compression plus stable source-pointer boundaries around digit-or-underscore identifiers. It likewise inherits no quality or speed claim until the same external artifact is tested.

Architecture at a glance

UTF-8 request
  -> sealed English host with persistent incremental state
  -> authenticated package registry and archive-bound routing profile
  -> core-only abstention, one selected cake, top-k, or structured multidomain plan
  -> selected neural capability execution only
  -> UTF-8 response and auditable execution trace

The release uses a frozen direct neural-decoder attachment path for the promoted packages. That is a byte-facing capability-host contract, not a claim of latent semantic fusion. See the architecture guide and the cake registry specification.

Documentation

Older North Star, tokenizer, byte-patch, and V2 development documents remain in the repository as historical research. They are not the release source of truth unless a current phase certificate imports and recomputes their raw evidence.

About

Modular language models via a fixed-dimension ABI bottleneck — domain modules are bit-exactly portable across model sizes

Topics

Resources

Contributing

Security policy

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages