orchestrate制御

初期live evidence (2026-09-10): Ryzen NPU FLM=gpt-oss:20b, status HTTP 200。Ryzen Vulkan=Ornith-1.5-35B-A3B Q4_K_M, profile=fast, healthy。B550=priority queue active、GPU0 Qwen3VL-8B-Instruct-Q4_K_M常駐、GPU1 text lease OFF、pending/running 0。B550 GPU status観測ではGPU0 busy 0%、VRAM約9.18GB/17.16GB、power約5W。よって現在の役割仮説は NPU=高頻度監督/分類、Ryzen Vulkan=意味レビュー/再計画、B550=bulk/visual、ChatGPT=Teacher/高影響判断。#11242でfixture benchmarkして確定する。旧llm.chatのdefault canaryはServer disconnectedが出たため、Routerはcanonical adapter/profile経路を使用し旧直結を依存にしない。

2026-09-11 live benchmark plan: compare the same bounded structured-decision fixtures across (A) B550 q8 via b550-qwen3-8b, (B) Ryzen Ornith via ryzen-api-semantic after #11244 reasoning=none fix, (C) Ryzen NPU ryzen-npu-supervisor. Score schema-valid rate, deterministic decision consistency, latency, confidence calibration, and provider/runtime failures. NPU currently health=online but structured completion unstable/timeout; do not promote until repeated canary passes. B550 one structured canary already passed schema. Ornith direct HTTP with reasoning=none returned valid JSON.

2026-09-11 live benchmark summary / Routing Matrix v1 evidence:

1. deterministic progress fixtures (4 fixtures x2): Ryzen Ornith ryzen-api-semantic 8/8 success, schema 100%, decision 100%, median 9.80s, max 11.48s. B550 b550-qwen3-8b 8/8 success, schema 100%, decision 100%, median 3.21s, max 12.74s. NPU ryzen-npu-supervisor health=online but 2 short structured calls did not complete within 120s combined; earlier raw tests also showed empty final content / reasoning-content domination. Do not promote NPU to production primary.
2. semantic root-cause fixtures (3 x2): Ornith 6/6 correct/schema-valid, median 11.32s; B550 q8 6/6 correct/schema-valid, median 9.51s.
3. Japanese editorial decision fixtures (2 x2): Ornith 4/4 correct/schema-valid, median 11.95s; B550 q8 4/4 correct/schema-valid, median 12.17s.
4. confidence behavior: B550 tends to emit 0.95-1.0 across deterministic/semantic fixtures; Ornith varies 0.55-1.0, giving more useful uncertainty signal.
   Routing v1: obvious progress/resource/state rules = deterministic code (no LLM); semantic_review/replan/ambiguous evidence = Ornith primary, B550 q8 fallback; bulk extraction/classification/article preprocessing = B550 primary; visual QA = B550 Qwen3VL lane; NPU = experimental/non-critical only until repeated schema+SLA threshold passes; high/critical impact, confidence<0.65, disagreement, teacher/final strategy = ChatGPT normal Chat. ChatGPT Work forbidden.