Qwen3.8 Flash Next vision-promotion evidence¶
This directory is the sanitized evidence bundle for the dated finding. All artifacts bind the exact RadixArk revision, SGLang runtime, dual-PRO TP=2 topology, and c1 route contract. No artifact contains a credential, private address, GPU UUID, personal path, media bytes, or data URL.
Compact outputs¶
summary.json- machine-readable decision summarythroughput-summary.csv- compact six-size tablepublication-summary.md- platform variants, alt text, and claim ledger
Context and throughput¶
capacity-4096-c1.jsoncapacity-32768-c1.jsoncapacity-65536-c1.jsoncapacity-131072-c1.jsoncapacity-196608-c1.jsoncapacity-253952-c1.json
Each anvil-serving.benchmark/v1 artifact retains request-level prompt and
output tokens, TTFT, effective prefill, decode, inter-token, E2E, measurement
definitions, completion state, and exact identity.
Multimodal corpus¶
direct-multimodal.json- direct 30/30isolated-routed-multimodal.json- 27/30isolated-routed-multimodal-r2.json- 30/30 repeatlive-routed-multimodal.json- 29/30live-routed-multimodal-r2.json- 28/30 repeat
The artifacts retain every strict assertion and model output. The five routed misses across four runs are visible rather than cherry-picked; every miss was a correct observation that omitted one literal rubric word.
Router and regression gates¶
isolated-router-edges.json- 8/8live-router-edges.json- 8/8live-primary-regression-preflight.jsonlive-vision-ocr-preflight.jsonmini-client-vision-acceptance.json
The edge artifacts deliberately retain admitted/overflow/malformed status, SSE termination, and grounded tool-call results without retaining the media payloads.