Skip to content

Commit 04cf292

Browse files
author
Sam Yoder
committed
docs: finalize sealed phase accounting
1 parent f574ccb commit 04cf292

20 files changed

Lines changed: 98 additions & 12 deletions

‎AGENTS.md‎

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -343,6 +343,10 @@ reaches the intended boundary. No unresolved finding of any severity may be
343343
waived to complete the campaign.
344344

345345
Phase 3's training-efficiency objective remains retired by governance.
346+
Its sealed certificate has zero headline claims and must not be described as
347+
training-efficiency, ABI-bootstrapping, capability-acquisition, or host-
348+
certification proof. It is a valid campaign-lifecycle retirement, not a skipped
349+
phase and not a scientific pass of the retired objective.
346350
Phase 8 must preserve its historical controls and must not claim faster
347351
foundation or domain training. Physical mobile, calibrated-energy dominance,
348352
latent fusion, and third-party laboratory independence also remain outside

‎BLOCKERS.md‎

Lines changed: 5 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1,5 +1,10 @@
11
# Experimental Blockers
22

3+
> Historical research backlog. These items do not reopen or redefine the sealed
4+
> Moonshot. Current certified limits are in
5+
> `docs/VERIFICATION_AND_LIMITS.md`; post-release work is governed by
6+
> `ROADMAP.md`.
7+
38
The software acceptance path and selected small-scale research gates pass. The following
49
broader scientific claims remain blocked by resources or evidence:
510

‎BREAKTHROUGH_SWEEP.md‎

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1,5 +1,9 @@
11
# LayerCake Breakthrough Sweep
22

3+
> Historical predecessor protocol. It is retained for reproducibility and does
4+
> not represent an open Moonshot gate or the final release verifier. Use
5+
> `BENCHMARKS.md` and `docs/PHASE_STATUS.md` for the sealed campaign.
6+
37
This is the locked engineer-facing gate for promoting the LayerCake breakthrough claim.
48

59
The sweep is intentionally strict: the top-level certificate passes only when every

‎BYTE_NEXTGEN_LADDER_100M_225M_500M.md‎

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1,5 +1,9 @@
11
# Byte NextGen Ladder: 100M -> 225M -> 500M
22

3+
> Historical byte-model scaling plan. It is not an active Moonshot phase,
4+
> production commitment, or certified release path. Its artifacts remain
5+
> research controls.
6+
37
This plan locks in fast iteration first, then scale only when quality gates require it.
48

59
## Core principle

‎BYTE_PATCH_LAYERCAKE.md‎

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1,5 +1,9 @@
11
# Byte-Patch LayerCake
22

3+
> Historical architecture-selection record. The sealed public interface remains
4+
> byte-facing, but the release does not claim tokenizer-free superiority. See
5+
> `docs/MOONSHOT_ARCHITECTURE.md` and `docs/VERIFICATION_AND_LIMITS.md`.
6+
37
The v2 path is:
48

59
`UTF-8 bytes -> byte embeddings -> causal patches -> patch core -> canonical ABI -> sparse brick -> byte logits`

‎CLAIMS.md‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -6,6 +6,7 @@ below is bounded by the exact release lineage and raw evidence paths.
66
| Claim | Evidence anchor | Certified result |
77
| --- | --- | --- |
88
| Campaign completion | `moonshot/campaign.yaml`, Phase 8 seal | All eight phases sealed; `verify-all` validates the completed campaign. |
9+
| 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. |
910
| Independent reproduction | `results/moonshot/phase8/raw_runs/cleanroom_environment.json` | Detached Phase 7 checkout, 597 passing tests, all prior typed gates recomputed. |
1011
| CPU performance | `phase8/raw_runs/reproduction_performance.json` | 9.91x LayerCake/Qwen output-byte throughput; 0.0179x median latency ratio. |
1112
| GPU performance | same | 8.25x LayerCake/Qwen output-byte throughput. |

‎DOMAIN_PASTE.md‎

Lines changed: 5 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1,5 +1,10 @@
11
# Domain Paste Guide
22

3+
> Historical direct-paste workflow. The certified Phase 4-6 package lifecycle,
4+
> compatibility, identity, routing, and portability contracts are documented in
5+
> `docs/CAKE_AUTHORING.md`, `docs/CAKE_REGISTRY_SPEC.md`, and
6+
> `docs/PHASE_STATUS.md`.
7+
38
## V2 canonical transfer note
49

510
This document primarily describes the original tokenized direct-paste workflow. Exact

‎DOMINANCE_GATES.md‎

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1,5 +1,9 @@
11
# Dominance gates
22

3+
> Historical experimental framework. Its smoke and tier gates are not the
4+
> sealed Moonshot acceptance criteria. The final promoted evidence is governed
5+
> by `BENCHMARKS.md`, `CLAIMS.md`, and the phase certificates.
6+
37
Dominance gates are a framework for serious experiments, not a smoke-CI claim.
48

59
Tracked dimensions:

‎EMBER_ROAD_BYTE_DOMAIN_WORKFLOW.md‎

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1,5 +1,9 @@
11
# Ember Road Byte-Domain Workflow
22

3+
> Historical project-specific training workflow. It is not part of the sealed
4+
> Moonshot product lineage and does not alter the current capability-package
5+
> claims.
6+
37
This workflow uses the tokenizer-free byte-patch LayerCake path only.
48

59
## 1) Train the 500M byte core

‎FAST_225M_CURRICULUM_PLAN.md‎

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1,5 +1,9 @@
11
# Fast 225M Curriculum Plan (Days-Scale)
22

3+
> Historical training proposal. It is not an active Moonshot gate or a
4+
> certified production recipe; Phase 3's original training-efficiency objective
5+
> was retired by governance.
6+
37
This plan is designed for fast iteration and companion quality in days, not weeks.
48

59
## Why this pivot

0 commit comments

Comments
 (0)