Skip to content

contract(trace-moe-gpu-sub-stages-v1): v1.5.0 → v1.6.0 — FALSIFY-MOE-SUB-004 DISCHARGED#1531

Merged
noahgift merged 1 commit into
mainfrom
contract/moe-sub-v1.6.0-sub-004-discharged
May 6, 2026
Merged

contract(trace-moe-gpu-sub-stages-v1): v1.5.0 → v1.6.0 — FALSIFY-MOE-SUB-004 DISCHARGED#1531
noahgift merged 1 commit into
mainfrom
contract/moe-sub-v1.6.0-sub-004-discharged

Conversation

@noahgift

@noahgift noahgift commented May 6, 2026

Copy link
Copy Markdown
Contributor

Summary

Status promotion amendment: FALSIFY-MOE-SUB-004 PROPOSED → DISCHARGED.

Rule

The M-GPU-MOE-1.4 fix PR title/body MUST mention one of:
{moe_router, moe_expert_gate, moe_expert_up, moe_expert_swigl, moe_expert_out, moe_ffn_out}.

Discharge evidence

aprender PR #1529 squash 89cb26af7 (M85) title:

fix(M-GPU-MOE-1.4 step c): qtype-aware dispatch in expert_swiglu_cuda — closes L6 moe_ffn_out NaN

→ explicitly cites moe_ffn_out (one of the 6 enumerated stages). The PR body further cites "moe_ffn_out at layer 6" multiple times in the Five-Whys analysis and per-layer bisection result table.

All four FALSIFY-MOE-SUB-* tests now DISCHARGED

Falsifier Discharged At
SUB-001 (parse) v1.4.0 (M82)
SUB-002 (byte-identity / heavy harness) v1.5.0 (M83)
SUB-003 (bisection-pinpoints-stage) v1.5.0 (M83)
SUB-004 (fix-PR-cites-stage) v1.6.0 (this PR)

M-MOE-SUB-4 stays PENDING

The contract's M-MOE-SUB-4 (per-expert sub-stages) was always optional — "only needed if MoeRouter+MoeFfnOut bisection is insufficient precision". M85's fix landed without it; M-MOE-SUB-3's precision was sufficient. M-MOE-SUB-4 remains a future enhancement if M-GPU-MOE-3 cosine-refinement work needs to bisect the ~7-8 cos<0.99 layers (L7, L9, L12, L20, L23, L29, L46) at per-expert granularity.

Test plan

🤖 Generated with Claude Code

…SUB-004 DISCHARGED

Status promotion amendment: FALSIFY-MOE-SUB-004 PROPOSED → DISCHARGED.

Rule: "The M-GPU-MOE-1.4 fix PR title/body MUST mention one of:
{moe_router, moe_expert_gate, moe_expert_up, moe_expert_swigl,
 moe_expert_out, moe_ffn_out}."

Discharge evidence: aprender PR #1529 squash 89cb26a (M85) title
is "fix(M-GPU-MOE-1.4 step c): qtype-aware dispatch in
expert_swiglu_cuda — closes L6 moe_ffn_out NaN" — explicitly cites
`moe_ffn_out` (one of the 6 enumerated stages) by name. The PR body
further cites "moe_ffn_out at layer 6" multiple times in the
Five-Whys analysis and per-layer bisection result table.

All four FALSIFY-MOE-SUB-* tests now DISCHARGED:
- SUB-001 (parse): DISCHARGED at v1.4.0 (M82)
- SUB-002 (byte-identity / heavy harness): DISCHARGED at v1.5.0 (M83)
- SUB-003 (bisection-pinpoints-stage): DISCHARGED at v1.5.0 (M83)
- SUB-004 (fix-PR-cites-stage): DISCHARGED at v1.6.0 (this amendment)

M-MOE-SUB-4 (per-expert sub-stages) stays PENDING — was optional;
M-MOE-SUB-3's MoeRouter+MoeFfnOut precision was sufficient for
M85's fix.

YAML-only — production hot paths byte-unchanged.

`pv validate` 0/0.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
@noahgift noahgift enabled auto-merge (squash) May 6, 2026 09:52
@noahgift noahgift merged commit 0bfcbc1 into main May 6, 2026
11 checks passed
@noahgift noahgift deleted the contract/moe-sub-v1.6.0-sub-004-discharged branch May 6, 2026 10:14
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant