Qwen3.8-27B NInfer NVFP4 RTX 5090 evidence

This directory contains the sanitized evidence bundle for the 2026-09-03-qwen38-ninfer-nvfp4-rtx5090 campaign. Native benchmark artifacts remain authoritative; the common files provide the campaign index.

Campaign boundary

• Campaign ID: 2026-09-03-qwen38-ninfer-nvfp4-rtx5090
• Capability: text generation and tool use
• Repository revision: 3fe82086592830ac1707ceaa8184428d2622ac7e, clean isolated codex/ worktree at campaign start
• Evidence labels: local-functional, local-capacity, local-performance, local-quality, exact-restoration
• Decision labels: verified, challenger, no-promotion
• Promotion boundary: this isolated benchmark did not authorize or perform a route, serve-role, or client-catalog change

Outcome

The NInfer NVFP4 MTP3 arm is the preferred measured direct text/tools performance challenger on this RTX 5090. At 4K/C1 it produced 0.430-second median TTFT, 165.9 tok/s median decode, and 0.720-second median E2E versus 0.421 seconds, 75.3 tok/s, and 1.085 seconds for the exact no-speculation control. A 201746-token API-reported prompt passed with an 8192-token completion cap.

This is not promotion evidence. The selected arm left 2354 MiB free, below the ordinary 3 GiB reserve, and each 20-way shared-prefix tool burst completed 17/20 because three requests received explicit C1 overload admissions. The exact incumbent was restored and passed a fresh smoke check.

Common campaign artifacts

• artifact-manifest.json - completed role ledger
• source-registry.json - dated source provenance
• workload-manifest.json - deterministic workload contract
• run-plan.json - preregistered execution order and gates
• configuration-start.json and configuration-end.json - sanitized before/after identity
• summary.json - bounded machine-readable decision
• friction-log.md - failures and operational friction
• restoration.json - verified incumbent restoration
• publication-summary.md - derivative public claims

Raw run evidence

• Matched capacity: capacity-nospec-4k-c1.json, capacity-mtp3-4k-c1.json, capacity-nospec-4k-c1-warm.json, and capacity-mtp3-4k-c1-warm.json
• No-spec functional/context: preflight-nospec-smoke.json, preflight-nospec-json.json, preflight-nospec-tools-c1.json, preflight-nospec-tools.json, and preflight-nospec-needle-244k.json
• MTP3 functional/context: preflight-mtp3-smoke.json, preflight-mtp3-json.json, preflight-mtp3-tools-c1.json, preflight-mtp3-tools.json, preflight-mtp3-needle-244k.json, and preflight-mtp3-needle-244k-output-reserve-8192.json
• Repeated bounded quality: quality-nospec.json and quality-mtp3.json
• Restoration check: restoration-smoke.json

Decision and publication

• Dated finding
• Feasibility record
• Managed recipe: configs/qwen38-27b-ninfer-nvfp4-rtx5090-252k-recipes.toml

The intake source-builds the exact NInfer revision on a digest-pinned CUDA base, but Ubuntu package resolution is not immutable. That limitation, the C1 admission result, the narrower memory reserve, and the unrun broader gates must be resolved before a separate human promotion decision.