Skip to content

CLAIMS.md

Latest commit

 

History

History
47 lines (38 loc) · 3.15 KB

CLAIMS.md

File metadata and controls

47 lines (38 loc) · 3.15 KB
Raw

LayerCake Claim and Evidence Map

The sealed moonshot release is defined by layercake-moonshot-final. Every row below is bounded by the exact release lineage and raw evidence paths.

Claim Evidence anchor Certified result
Campaign completion moonshot/campaign.yaml, Phase 8 seal All eight phases sealed; verify-all validates the completed campaign.
Phase 3 disposition results/moonshot/phase3/release_certificate.json RETIRED_BY_GOVERNANCE; zero headline claims and no training-efficiency, acquisition, ABI, or host-certification proof.
Faster full-core training TRAINING_NORTHSTAR.md OPEN; no training-dominance claim.
Independent reproduction results/moonshot/phase8/raw_runs/cleanroom_environment.json Detached Phase 7 checkout, 597 passing tests, all prior typed gates recomputed.
CPU performance phase8/raw_runs/reproduction_performance.json 9.91x LayerCake/Qwen output-byte throughput; 0.0179x median latency ratio.
GPU performance same 8.25x LayerCake/Qwen output-byte throughput.
CPU versus transformer GPU same 7.67x output-byte throughput; 0.0222x median latency ratio.
Functional quality same LayerCake 100/100 on CPU and GPU; paired bootstrap lower bounds +0.68 and +0.67.
Domain retention phase8/raw_runs/domain_retention.json 384/384 CPU, 384/384 GPU, and 384/384 identical outputs.
Portability phase8/raw_runs/lifecycle_portability.json Exact archives survive three-host install/verify/remove/reinstall with zero receiver learning.
Routing and catalog phase8/raw_runs/routing_catalog.json 1,980/1,980 correct rows and a 500-entry management catalog.
Attack resistance phase8/raw_runs/adversarial_falsification.json 32 attacks across 24 categories reached and resolved their intended boundaries.

Claims deliberately excluded

This release does not claim physical mobile performance, energy dominance, GPU-training dominance, general faster foundation training, latent cake fusion, universal model superiority, arbitrary host compatibility, or remote tag publication. See docs/VERIFICATION_AND_LIMITS.md.

All earlier claim tables are historical research context. They do not modify this evidence map or survive changes that trigger the invalidation matrix.

External ABI research is not an inherited claim

Capability acquisition from foreign teachers is researched in the separate ABI repository. No external ABI result changes or inherits this release's claims. See docs/ABI_HANDOFF_STATUS.md for the repository and recertification boundary.

Post-release host-interface audit

The read-only audit in docs/POSTRELEASE_EXTERNAL_CORE_HOST_INTERFACE_AUDIT.md establishes only that the sealed product originally had no canonical signed direct neural English- core artifact role. The isolated lc-direct-neural-core/1 extension now passes its mechanical host construct on CPU and CUDA. It does not invalidate the release or certify English, external acquisition, speed, memory, TTFT, domain compatibility, or ABI superiority. See docs/POSTRELEASE_DIRECT_NEURAL_CORE_HOST.md.