Qwen3.8 27B RTX 5090 quant bakeoff evidence

This directory contains the sanitized evidence bundle for the 2026-09-03 single-GPU RTX 5090 Qwen3.8 27B quantization and speculative-decoding bakeoff. Native benchmark artifacts are authoritative; the files here provide the campaign-level index.

Campaign boundary

• Campaign ID: 2026-09-03-qwen38-27b-rtx5090-quant-bakeoff
• Capability: text LLM serving
• Repository revision: aed0214e03d2fe0e8220e3249a12f90994fd89b5 at campaign start; clean isolated worktree
• Evidence labels: functional, capacity, matched-performance, bounded-quality, compatibility-failure, exact-restoration
• Decision labels: challenger, no-promotion
• Promotion boundary: benchmark evidence does not authorize a router or live-alias change

Common campaign artifacts

• artifact-manifest.json - role ledger and explicit evidence gaps
• source-registry.json - dated source provenance
• summary.json - bounded machine-readable decision
• friction-log.md - failures and operational friction
• restoration.json - starting state and final restoration proof
• configuration.json - sanitized starting identity and policy envelope

Workload and plan

• run-plan.json declares the candidate order, paired no-speculation/speculation controls, repetitions, context points, concurrency checks, gates, and ranking order.
• workload-manifest.json defines the deterministic smoke, timing, durable-context, and concurrency workloads.

Raw run evidence

• Every loaded arm has retained *-preflight.json, warm/cold *-timing-4k-*.json, and *-capacity-*.json artifacts.
• Gittensor and CometKim have repeated *-quality.json artifacts.
• The advertised Gittensor DSpark startup failure is retained in gittensor-dspark-load-failure.json.
• vLLM speculative acceptance counters are retained in the four *-runtime-metrics.json files.
• The invalid CometKim per-request thinking-control attempt remains labeled *invalid-thinking-control.json; it is not included in performance claims.

Decision and publication

The Gittensor target-only SGLang arm won the primary TTFT metric at 50.9 ms warm median while passing repeated bounded checks and a 244,002-token actual- prompt request. CometKim MTP3 won decode at 228.0 tok/s but failed strict tools 0/3. The exact Unsloth incumbent was restored. See the decision summary, publication summary, and dated finding.