The eight-phase Moonshot campaign is complete and sealed. There is no automatic
"Phase 9" and no permission to alter the certified product under the release
claim. Work after layercake-moonshot-final belongs to one of the following
explicit tracks.
- Re-run
verify-allon clean supported environments. - Publish the annotated tag only through an explicit release decision; remote publication is not currently certified.
- Maintain the package signing, registry, and vulnerability-response process.
- Record environment-specific verification results without claiming that a new host inherits the declared performance ratios.
- Author new signed, non-executable packages using the canonical cake ABI.
- Evaluate each package for package identity, semantic retention, core immutability, routing eligibility, and selected-only execution.
- Treat Python, SQL, and regex as sealed examples—not as a claim that any new domain is automatically useful or portable.
ABI extraction owns source-model capability extraction, minimization, imported information accounting, and independent artifact validation. ABI code and evidence do not belong in this LayerCake repository. A future ABI artifact may enter only as an independently validated, signed, content-addressed package.
The separate ABI repository owns all
foreign-teacher acquisition research. LayerCake remains in release stewardship
unless an independently validated external artifact enters the governed handoff
defined in docs/ABI_HANDOFF_STATUS.md.
Any change to the host, ABI, package format, package payload, router, runtime
numerics, checkpoint, data, evaluation suite, benchmark protocol, precision, or
hardware claim invokes moonshot/invalidation_matrix.yaml. The change must:
- create a versioned governance amendment;
- preserve earlier evidence as historical;
- declare affected phase gates before experiments;
- rerun every required dependent phase; and
- receive a new annotated seal before a public claim changes.
The current evidence-supported candidate for such a future track is a signed,
byte-facing, self-causal direct neural core artifact interface. The post-release
V2 audit found that the existing direct interface is capability-cake-only and
the semantic-residual family is not autonomously qualified for an external
English core. See
docs/POSTRELEASE_EXTERNAL_CORE_HOST_INTERFACE_AUDIT.md. No implementation or
claim is authorized without a new preregistration.
The current construct implementation is lc-direct-neural-core/25, isolated
under layercake_extensions/. It adds the fifth clarification-only residual
route required by the current external handoff while retaining V24's
allocation-bounded execution. The next governed recertification input is one
real, independently validated, immutable five-route English artifact. That
same payload must earn quality, routing, collapse, CPU/GPU, memory, TTFT, and
package-identity evidence; the host construct supplies none of that credit.
Older byte, token, hybrid, training-efficiency, and architecture experiments can continue as research only. They must not consume the sealed release identity or be described as an unfinished Moonshot phase. Faster from-scratch English training remains a control, not a LayerCake product gate.