{
  "schema": "anvil-serving.benchmark-decision-summary/v1",
  "campaign_id": "2026-09-21-mimo-v26-qualification",
  "complete": true,
  "identity": {
    "candidate_repository": "XiaomiMiMo/MiMo-V2.6-Flash-RL",
    "candidate_revision": "3b38d063180c3e4aed9691fdc735f3d10b266ee4",
    "runtime": "SGLang v0.5.20",
    "runtime_revision": "94602c9c2b7cbdb8efd5c52802dac6a1c180089e",
    "hardware": "2x RTX PRO 6000 Blackwell Max-Q",
    "baseline": "GLM-5.3-Flash EXL3 r10 APC, TP2/C4"
  },
  "workload": {
    "baseline_capacity": {
      "requests": 100,
      "concurrency": 4,
      "nominal_context": 32768,
      "response_words": 16,
      "max_tokens": 2048,
      "prompt_cache_mode": "unique within each cell; first eight final requests may reuse the scout cache",
      "request_canaries": true,
      "strict_output": true
    },
    "candidate_correctness": {
      "same_greedy_preflight_first_check": "failed",
      "official_sampling_probe": {
        "temperature": 1.0,
        "top_p": 0.95,
        "requests": 1,
        "visible_budget": 1024,
        "reasoning_headroom": 8192
      },
      "partial_batch": "interrupted during JSON after first coding smoke failed"
    }
  },
  "measurement": {
    "path": "online direct",
    "warm_cold_state": "Baseline already running; candidate cold load followed by engine warmup and correctness probes",
    "sample_count": 100,
    "statistics": [
      "Baseline request percentiles use nearest rank. No performance-eligible MiMo population."
    ]
  },
  "gates": [
    {
      "gate": "baseline capacity C4",
      "status": "passed",
      "evidence": "glm-capacity-c4-16word-final.json"
    },
    {
      "gate": "Candidate native-scale cold load",
      "status": "passed",
      "evidence": "mimo-v3-startup.log"
    },
    {
      "gate": "candidate correctness",
      "status": "failed",
      "evidence": "mimo-v3-sampling-smoke.json"
    },
    {
      "gate": "Candidate capacity, vision, long context and repository coding",
      "status": "not-run",
      "reason": "correctness gate failed"
    },
    {
      "gate": "GLM final restoration",
      "status": "passed",
      "evidence": "restoration.json"
    }
  ],
  "headline_results": [
    {
      "name": "GLM controlled capacity",
      "eligible": 100,
      "requests": 100,
      "evidence": "glm-capacity-c4-16word-final.json"
    },
    {
      "name": "MiMo official-sampling diagnostic",
      "pass_count": 0,
      "requests": 1,
      "visible_characters": 0,
      "finish_reason": "length",
      "completion_tokens": 9216,
      "latency_ms": 74834.97632504441,
      "performance_eligible": false
    }
  ],
  "failures": [
    "V1 rejected an ambiguous legacy CUDA-graph flag.",
    "V2 hit CUDA OOM before weight loading; source review identified avoidable expert-scale widening.",
    "V3 loaded with compact scales but failed correctness at temperature 0 and official temperature 1/top_p 0.95. Exact corruption cause is unknown.",
    "GLM SWE officially resolved 0/1. The bug-specific test passed; two HTTP 502 failures and a legacy pytest API error confound model attribution."
  ],
  "limitations": [
    "This runtime/recipe failure does not rank upstream model quality.",
    "Configured context was 327,680 tokens; V3 allocated 91,342 aggregate full-attention KV tokens.",
    "MiMo C4, vision, long-context and SWE tests were not run after correctness failed.",
    "Failed responses and engine token counters do not support a speed claim.",
    "GLM p99 at N=100 is one order statistic; cache warmth of the first eight requests is unverified.",
    "The greedy preflight batch was interrupted after its first failed check; no final native JSON exists for that partial batch.",
    "SWE was one smoke task. Its official unresolved result was retained without regrading."
  ],
  "decision": {
    "evidence_labels": [
      "compatibility-only",
      "failed-correctness",
      "measured-baseline"
    ],
    "decision_labels": [
      "retain-baseline",
      "no-promotion"
    ],
    "promotion_authorized": false,
    "next_gate": "Establish a correct SM120/TP2 runtime or checkpoint path, then pass preflight before capacity, vision and long-context qualification."
  },
  "evidence": [
    "glm-capacity-c4-16word-final.json",
    "mimo-v3-startup.log",
    "mimo-v3-preflight.log",
    "mimo-v3-sampling-smoke.json",
    "glm-swe-native.json",
    "glm-swe-official-report.json",
    "glm-swe-test-output.txt",
    "restoration.json"
  ]
}