Qwen3.8 Huihui NInfer qualification evidence — 2026-09-19

The exact no-spec8K/C1 profile is rejected for promotion after repeated strict tool failures. The campaign reached a hard-gate decision and restored media; full performance, long-context, MTP and endurance qualification did not run.

• Dated finding, decision, coverage, publication.
• Identity and run binding, workloads, run plan, feasibility input and result.
• Source registry, research, producer manifest.
• Initial text recipe, vision reload recipe, owning log excerpts.
• Friction, independent intake review, gate review, restoration.
• Repository verification, infrastructure disposition.
• Campaign ledger, bounded assignments, hash manifest.

Native evidence

Artifact	What it proves
Initial smoke	Two HTTP400 failures from unsupported request field.
Server-default smoke	Smoke and JSON2/2.
Protocol preflight	Tools, streaming tools, continuation, Responses,4096-token needle5/5.
Quality off	Diff0/3, triage3/3, string ZIP tools0/3.
Quality requested low reasoning	Diff3/3, triage3/3, string ZIP tools0/3.
Initial images	Configuration disabled vision;0/12.
Vision reload	Image/OCR/chart/UI/spatial/multi-image12/12.
Incumbent preflight	Smoke and JSON2/2.
Incumbent quality	Diff, triage and string ZIP tools3/3 each.

All probes use direct loopback on the measured Secondary Node RTX5090. Native latency fields are incidental mixed warm/cold samples, not controlled performance evidence. Launcher1.0.0 comes from repository commit 2b99e3353eb2e5a2813195af85050b1da0754332 plus this campaign's recipe/docs additions. Model grading uses deterministic assertions, not self-evaluation.

Absolute operator corpus paths were redacted in native image JSON; public text line endings were normalized to LF. Private originals remain. Runtime binaries/linked dependencies are not baked into an immutable candidate image; no deployment recreation claim is made.