This document is the human-readable companion to moonshot/campaign.yaml.
The campaign file, phase certificates, raw evidence, and Git tags are
authoritative; prose never promotes a claim by itself.
| Field | Value |
|---|---|
| Campaign | layercake-moonshot |
| Final status | SEALED |
| Final tag | layercake-moonshot-final |
| Final tag commit | 0537cbb9e93cd7ebd4ba01c0bf641414ecebb1c3 |
| Final release-evidence commit | ebf594073c7364df9249e75ee9833861a7804992 |
| Phase 8 framework commit | 22501a3a4ef16c3cbb390aef565ec502391d8768 |
| Governing Phase 8 contract | moonshot/phase8_independent_verification_preregistration.json |
Run C:\Python310\python.exe -m layercake.moonshot_campaign verify-all to
recompute the campaign-level lifecycle state. A valid release reports
completed_phases_valid: true.
| Phase | Sealed result or disposition | Primary evidence | Seal tag |
|---|---|---|---|
| 0 | Permanent governance, contracts, invalidation rules, and campaign state | results/moonshot/phase0/ |
layercake-moonshot-phase0 |
| 1 | Auditable benchmark truth with matched quality/speed accounting | results/moonshot/phase1/ |
layercake-moonshot-phase1 |
| 2 | One integrated English-host lineage cleared the locked matched-quality CPU gate | results/moonshot/phase2_recertification/ |
layercake-moonshot-phase2-r3 |
| 3 | Original training-efficiency objective retired by governance; bounded controls and negative evidence preserved | results/moonshot/phase3/ |
layercake-moonshot-phase3 |
| 4 | A useful signed Python capability package transferred losslessly without receiver learning | results/moonshot/phase4/ |
layercake-moonshot-phase4 |
| 5 | Three real capability packages established generic multi-domain extension | results/moonshot/phase5/ |
layercake-moonshot-phase5 |
| 6 | Archive-bound routing, selected-only execution, multi-mode orchestration, and catalog scaling | results/moonshot/phase6/ |
layercake-moonshot-phase6 |
| 7 | Locked CPU, GPU, and CPU-versus-GPU performance for the exact integrated product | results/moonshot/phase7/ |
layercake-moonshot-phase7 |
| 8 | Detached clean-room reproduction, hostile falsification, and final release integrity | results/moonshot/phase8/ |
layercake-moonshot-final |
Phase 3 ran bounded training-efficiency experiments and preserved their positive
and negative evidence. Its final certificate has scope
governance_retirement_no_training_efficiency_claim, disposition
RETIRED_BY_GOVERNANCE, zero headline claims, and
scientific_training_efficiency_passed: false. The five promoted records prove
only that the retirement was authorized, historical evidence was preserved, no
scientific training pass was claimed, the Phase 2 parent was sealed, and a future
ABI artifact would require recertification.
Therefore Phase 3 is a valid sealed campaign transition, not a skipped lifecycle step and not a scientific pass of faster training, English acquisition, ABI bootstrapping, or host certification. Ordinary LayerCake training remains a research and fallback capability.
The final certificate derives three required metrics from fresh raw evidence:
| Required gate | Result |
|---|---|
| Prior required-gate retention | 1.0 |
| Clean-room reproduction | 1.0 |
| Adversarial findings resolved | 1.0 |
The detached verifier reran a 597-test clean checkout, all Phase 2-7 typed gates, a 100-distinct-prompt plus 20-repeat four-system performance matrix, 384 held-out domain cases per device, 100 core abstentions, 1,980 routing rows, a 500-entry catalog, and a three-host package lifecycle. It also executed 32 hostile checks spanning 24 categories.
The release certificate is
results/moonshot/phase8/independent_verification_certificate.json; the
human-readable report is results/moonshot/phase8/release_report.md.
The sealed phase tags are immutable historical boundaries. Documentation,
tooling, runtime, ABI, package, checkpoint, data, evaluation, or precision
changes made after a seal do not inherit the corresponding scientific claim.
Use moonshot/invalidation_matrix.yaml and open a governed recertification
before modifying a component that affects a sealed result.
Failed and invalidated Phase 8 attempts are deliberately preserved under
results/moonshot/phase8/history/. They are negative systems evidence, not
release evidence.