From e865ffdf259e7cc2230fafd063f980dc504e4037 Mon Sep 17 00:00:00 2001 From: Timothy Wayne Gregg Date: Tue, 29 Sep 2026 22:00:05 -0400 Subject: [PATCH] Record verified canonical FP32 and graph repair delivery --- docs/RELEASE_READINESS.md | 20 ++++++++++++++++++- docs/acceptance-status.json | 12 +++++++++++ .../short-fp32-canonical-v10-20260929.md | 1 + 3 files changed, 32 insertions(+), 1 deletion(-) diff --git a/docs/RELEASE_READINESS.md b/docs/RELEASE_READINESS.md index 8257e01..d3a8a07 100644 --- a/docs/RELEASE_READINESS.md +++ b/docs/RELEASE_READINESS.md @@ -1,5 +1,23 @@ # Release readiness — intentionally incomplete +## Verified numerical/source delivery (September 29, 2026) + +[PR #56](https://github.com/Jev-Engineering/layev/pull/56) merged at +`4f31cc76e38ef9848950e1667875571f8e190903`, from verified SSH-signed source +`4ffd9f1f1263d225e4f130f174cf6070c635d235`. All five required +[PR checks](https://github.com/Jev-Engineering/layev/actions/runs/36656915788) +and all five [post-merge checks](https://github.com/Jev-Engineering/layev/actions/runs/36657405767) +passed. The primary checkout synchronized cleanly with main. The CUDA graph +ownership regression passed again on the merged checkout (one pass, 80 deselected). + +This completes delivery of the canonical FP32 numerical and BF16 graph lifetime +repairs, public reproducers, and scoped receipts. Historical numerical and rank-8 +receipts retain their precommit source identities; the graph repair records its +later training source separately. Complete rank-0 resume and exact-limit FP32 +training still fail on the measured configuration. Issues #3/#4 remain open; +representative quality, Jev benefit and remote publication remain unverified. +Overwatch remains archived by user instruction. + ## Full-weight mechanics and graph lifetime supplement (September 29, 2026) The [graph lifetime repair and measured rank-0 limits](evidence/native/graph-lifetime-fullweight-limits-20260929.md) @@ -14,7 +32,7 @@ The [canonical FP32 repair](evidence/native/short-fp32-canonical-v10-20260929.md passes the short independent HF adapter oracle, four generated cached/batched/full cases, 8191/8192 state-prefix boundaries, and a 10520-token state numerical gate. This is candidate source evidence; the exact-limit FP32 training resource probe exhausted the fixed allocator cap -during forward, leaving backward unverified. Delivery readback remains pending. Issues #3/#4 remain open; representative quality and Jev benefit +during forward, leaving backward unverified. Delivery is recorded above; these receipts retain their original source identities. Issues #3/#4 remain open; representative quality and Jev benefit are unknown. Overwatch remains archived by user instruction. Older readbacks below are historical and do not override these scoped receipts. diff --git a/docs/acceptance-status.json b/docs/acceptance-status.json index 2848fcc..6d96205 100644 --- a/docs/acceptance-status.json +++ b/docs/acceptance-status.json @@ -1,4 +1,16 @@ { + "canonical_fp32_graph_delivery_2026_09_29": { + "pull_request": "https://github.com/Jev-Engineering/layev/pull/56", + "source_commit": "4ffd9f1f1263d225e4f130f174cf6070c635d235", + "merge_commit": "4f31cc76e38ef9848950e1667875571f8e190903", + "pr_ci_run": 36656915788, + "postmerge_ci_run": 36657405767, + "required_jobs_passed_each_run": 5, + "merged_cuda_graph_regression_passed": 1, + "scope": "Verified source delivery; historical numerical receipts retain their original source hashes; native resource and quality acceptance remain incomplete", + "issue_3": "open", + "issue_4": "open" + }, "fullweight_graph_lifetime_supplement_2026_09_29": { "receipt": "docs/evidence/native/graph-lifetime-fullweight-limits-20260929.json", "graph_lifetime": "CUDA positive passes; proper old-source negative fails; 14 BF16 tests pass", diff --git a/docs/evidence/native/short-fp32-canonical-v10-20260929.md b/docs/evidence/native/short-fp32-canonical-v10-20260929.md index 3a4e407..0be9aac 100644 --- a/docs/evidence/native/short-fp32-canonical-v10-20260929.md +++ b/docs/evidence/native/short-fp32-canonical-v10-20260929.md @@ -7,6 +7,7 @@ records checkpoint, source-file, protocol-script and private-result hashes. requests, fixed precision settings and numerical tolerances. The source parent is recorded with a dirty candidate tree; delivery commit and post-merge verification must be recorded separately by the delivery owner. +The [verified delivery readback](../../RELEASE_READINESS.md) records those revisions and checks. CUDA FP32 projections use fixed 64-row calls. Attention dispatch follows absolute query bands: positions 0–255 use fixed 256-row eager arithmetic, positions