| 2026-09-03T12:00:00Z |
missing-identity |
runtime distribution |
NInfer publishes source and a Dockerfile but no digest-pinned runtime image or packaged binary. |
Source-build exact revision inside the managed recipe using a digest-pinned CUDA base image and revision-specific build volume. |
Add a promotion-grade baked-image preparation surface with pinned package inputs before promotion. |
| 2026-09-03T12:01:00Z |
failure |
Windows/WSL GPU-lane cross-check |
The default WSL distribution could not start because its registered virtual disk was absent; Windows and Docker inventory remained available. |
Continue through the product's Docker Desktop path and retain the WSL discrepancy as an environment limitation. |
Repair or unregister the stale WSL distribution outside this model campaign. |
| 2026-09-03T12:04:28Z |
reproducibility |
managed source build |
The CUDA base image is digest-pinned and NInfer source revision is verified, but apt package resolution is floating. |
Treat the recipe as candidate intake only. |
Bake and publish a digest-pinned runtime with immutable package inputs before promotion. |
| 2026-09-03T12:10:00Z |
setup-cost |
first managed load |
The first load compiled NInfer from the exact source revision before serving. |
Retain the revision-specific build cache for the matched second arm and future local reruns. |
Measure and publish cold build/load time separately if operational readiness becomes a gate. |
| 2026-09-03T12:24:00Z |
admission |
20-way shared-prefix tools |
Both no-spec and MTP3 arms completed 17/20; three requests per arm received explicit HTTP 429 server_overloaded responses under C1. |
Preserve the failures and continue bounded C1 functional checks. |
Resolve or explicitly accept the C1 admission contract before promotion; do not report 20/20 tool concurrency. |
| 2026-09-03T12:29:00Z |
measurement |
nominal 244480 retrieval fixture |
The harness's nominal filler target produced 201746 API-reported prompt tokens. |
Report both values and use 201746 for all actual-context claims. |
Add tokenizer-aware target generation if an exact deeper prompt point is required. |
| 2026-09-03T12:34:00Z |
policy |
selected MTP3 memory state |
MTP3 left 2354 MiB free: above the campaign's explicit 1 GiB model-only floor but below the ordinary 3 GiB reserve. |
Keep the profile challenger-only and restore the incumbent after qualification. |
Adjudicate reserve policy or reduce memory demand before any promotion. |
| 2026-09-03T12:36:00Z |
coverage |
direct qualification scope |
No router, real client, multimodal, thinking-enabled, broad agentic/SWE, or sustained thermal gate was run. |
Bound the finding to direct text/tools performance. |
Run the missing gates only after the runtime, admission, and reserve blockers are resolved. |