Skip to content

Latest commit

 

History

History
62 lines (50 loc) · 3.16 KB

VERIFICATION_AND_LIMITS.md

File metadata and controls

62 lines (50 loc) · 3.16 KB

Verified Claims and Explicit Limits

What the sealed release verifies

The final claim is scoped to the exact LayerCake product lineage, Qwen 2.5 0.5B comparator deployment, data, checkpoints, packages, runtime settings, and laptop hardware named in the Phase 8 raw evidence.

  • The full eight-phase campaign is sealed and independently reproducible from its raw evidence.
  • LayerCake passed the locked 100-prompt functional suite on CPU and GPU and retained every promoted held-out domain result across CPU and GPU.
  • LayerCake's output-byte throughput exceeded the locked optimized Qwen comparator by 9.91x on CPU and 8.25x on GPU. LayerCake CPU also exceeded the comparator's GPU throughput by 7.67x in the declared matrix.
  • The final product installed, verified, removed, and reinstalled identical, signed, non-executable Python, SQL, and regular-expression capability archives on three fresh hosts without receiver training or calibration.
  • Routing was archive-bound and fail-closed: a missing, stale, unauthenticated, mismatched, or permission-ineligible package cannot enter automatic execution.
  • Only selected promoted capabilities execute. Core-only operation abstains safely when the request requires an unavailable capability.

What the sealed release does not claim

The following remain intentionally outside the certificate:

  • physical Android, iOS, ARM, NPU, battery, thermal, or mobile performance;
  • calibrated energy or carbon dominance;
  • GPU-training dominance;
  • faster from-scratch English acquisition or a general 5x training claim;
  • latent neural fusion of multiple direct-decoder cakes;
  • universal language-model superiority beyond the frozen functional suites;
  • arbitrary third-party host compatibility;
  • external human, laboratory, or remote-repository independence;
  • publication of the final Git tag to a remote.

The Phase 3 training-efficiency controls are preserved as historical science but are not product gates. Its sealed retirement certificate has zero headline claims and does not certify training efficiency, ABI bootstrapping, external artifact acquisition, or LayerCake host certification. ABI extraction, minimization, and source-model knowledge transfer are separate work and are not implemented, imported, or certified by this LayerCake release.

Correct interpretation of portability

Package transfer is mathematical at the archive and tensor boundary: the same authenticated bytes are installed on compatible hosts. Semantic retention is measured separately using frozen functional success sets. Neither property authorizes a claim that arbitrary neural networks share representations or that any package can run on any host.

Correct interpretation of speed

The primary cross-model rate is UTF-8 output bytes per complete request wall second. Transformer tokens are recorded through the deployment runtime but are not equated with LayerCake actions. Quality, speed, memory, and latency use the same accepted product lineage; values from an older checkpoint or another model cannot be mixed into a release claim.

For command-level reproduction and all raw values, use the certificates and raw files under results/moonshot/phase8/, not this summary.