Collect exact GLM-5.3 runtime evidence

[FujitsuPolycom]

Outcome

This evidence-only collection preserves three independent GLM-5.3 artifact records without changing runtime code, profiles, configuration, commands, pins, cache identity, or stored formats. Qualification never transfers between the recorded images.

Exact recurrent publication and shared exact prefix

Status: qualified only for the bounded cases recorded for local image sha256:ed60be066d6d9eadea267bc4597a0687869f3ddb95a3e5c6f86649893a838eb8.

• pull request Add exact GLM-5.3 adaptive-MTP and DFlash7 SparkCache runtimes #146 evidence base: 74a6a472e7f11a56029d52b879c9b3665aa60e01, tree 5d8a87fb545a0cd6316d054dd9a2c35b186603e1
• SparkRing runtime-contract commit: d93cb3d98305041081cf572521602625185112ae
• SparkCache commit: 65b6642df1afc64366430d3aef9aca01f5c5e1c3
• image construction receipt SHA-256: 2c4a02efe91df5de21c5e3c92f65710b7d41680f25c22baed43ca96c1e5a51d3
• semantic canary, clean-restart 8K restore, 128K restore, appended 128K tail publication, verified 256K restore, and one C16 distinct-tail shared exact prefix cohort completed
• the exact receipt existed before all observed replacement containers were created

Restore timings are research-only. DFlash response quality, public OCI publication, and continuous availability of the preceding service are unsupported by this record.

The receipt schema is sparkring-glm53-pr146-live-validation/v2. Its evidence normalization classifies the C16 result as shared exact prefix reuse for block_pages_v1. Prompt hashes, HTTP outcomes, request IDs, per-rank timing values, and byte counts are unchanged. The artifact does not exercise per_token_rows different-root descriptor coalescing; that behavior is implemented with GPU-free coverage only and has no live qualification here.

DFlash7 image 9faa… with SparkCache 5d571018

Status: qualified only for the cases recorded for image sha256:9faa36a9f37aee16d97ab9214ef3153b4d200121126e6b2dee5ebb63109fea18.

• SparkRing revision label: e2d92fdc7d0306d664d6fd9f296dc2adcaf0fe05
• SparkCache commit: 5d571018de5b63a9a90e5c11e6d6e86bbff4a957
• startup, four-rank health, raw semantic smoke, flat restore, reconstructed page-delta correctness, persistent 128K-class restore, and one retained-prefix wave each at C2, C8, and C16 are preserved
• page-delta timing is research-only and DFlash response quality is unsupported

This receipt explicitly does not qualify image sha256:eef863d8bc578815a80b0e2d9f0d745102b6363415225101fd92171a2e5a55cb or image sha256:ed60be066d6d9eadea267bc4597a0687869f3ddb95a3e5c6f86649893a838eb8. Pull request #138's quickstart changes are intentionally excluded from this collection.

Adaptive-MTP and DFlash7 decode diagnostic

Status: research-only.

The diagnostic preserves bounded server-log observations for adaptive-MTP image sha256:d209cf986c1f14f320e53d8b425c5e3a255eef9320f25b632af50f5b5c977314 and the separately identified DFlash7 service. It retains exact runtime identities and UTC intervals. Prompt, generated output, request shape, and client timing receipts are absent, and the compositions differ in several components. The record therefore qualifies no runtime, image, benchmark, quality result, or SparkCache behavior. No SparkCache event appeared in the measured decode intervals.

Artifact separation

The ed60..., 9faa..., and eef... images have independent source and evidence boundaries. C2/C8/C16 observations from the PR25 9faa... image do not qualify the eef... PR30 artifact or the ed60... recurrent-publication artifact. The diagnostic comparison is not qualification evidence for any of them.

No file from pull request #138's quickstart delta is included. No runtime, profile, resolver, builder, or configuration file from pull request #132 or #138 is included.

Historical review note

Pull requests #132 and #138 retain their review history. They may be marked superseded only after reviewers accept this consolidated evidence collection. This change does not close, delete, or merge either pull request.

Validation

• focused evidence and quickstart coverage: 28 passed
• exact runtime and profile coverage: 47 passed
• complete repository: 2008 passed, 9 skipped
• Ruff E/F/W: passed
• repository-relative Markdown links: 258 checked
• release-safety scan: 0 findings
• JSON parsing and git diff --check: passed

[FujitsuPolycom]

This was generated by AI during triage.

The public GLM-5.3 R8 image and its main-branch validation record now provide the supported runtime evidence. This draft records earlier source-overlay artifacts and remains useful as research history, but it is not a candidate for main.

Closing the pull request to remove it from the active review queue. Branch codex/pr146-live-validation is retained.