{
  "observedAt": "2026-10-19T03:26:49.220197+00:00",
  "workflowVersion": "1.5.6",
  "machine": {
    "model": "Mac16,10",
    "chip": "Apple M4",
    "physicalCpuCores": "10",
    "logicalCpuCores": "10",
    "memoryBytes": "17179869184"
  },
  "macOSVersion": "27.0.1",
  "macOSBuild": "26A434",
  "architecture": "arm64",
  "reportedVirtualMachineFlag": 0,
  "dedicatedRentedPhysicalMac": true,
  "differentHostVerifiedByLocalHashComparison": true,
  "operator": "Same Kingy operator; no independent external lab claim",
  "pythonVersion": "3.12.14",
  "pypdfVersion": "6.10.0",
  "model": {
    "basename": "Qwen3-4B-Q4_K_M.gguf",
    "bytes": 2497280256,
    "sha256": "7485fe6f11af29433bc51cab58009521f205840f5b4ae3a32fa7f92e8534fdf5"
  },
  "runtime": {
    "version": "version: 0.6.0-dev (build 11429, commit d81235049)\nbuilt with AppleClang 21.0.0.21000101 for Darwin arm64",
    "binarySha256": "1bf5602a2be50097c44282a6ecc63234d94ee8f96d39fd45fed6eff4e74b6b41"
  },
  "runtimeSource": "https://github.com/ggml-org/llama.cpp/releases/download/b11429/llama-b11429-bin-macos-arm64.tar.gz",
  "runtimeArchiveSha256": "740288ec6887be94280a5dfa25b5e23a78285cab104519e6c7e218904ee82459",
  "requestedAndObservedBackendRecorded": true,
  "gpuLayersRequested": 99,
  "observedBackend": "Metal computation sampled during original session 3",
  "observedOffloadedLayerCount": null,
  "allThreeSessionLogsReportActualLayerPlacement": false,
  "backendObservation": {
    "duringSession": 3,
    "observedAt": "2026-10-19T03:20:21.983048+00:00",
    "sampleDurationSeconds": 1,
    "computeSymbols": [
      "ggml_metal_synchronize"
    ],
    "metalLibraryLoaded": true,
    "modelGraphToMetalCommandBufferChain": [
      "llama_context::graph_compute",
      "ggml_backend_sched_graph_compute_async",
      "ggml_metal_synchronize",
      "_MTLCommandBuffer waitUntilCompleted"
    ],
    "originalRecipeSettingsAndClocksUnchanged": true,
    "timingIncludesReadonlyObserverOverhead": true,
    "noDocumentQuestionRepeatedForObservation": true
  },
  "originalSessionRuntimeHashes": [
    "1bf5602a2be50097c44282a6ecc63234d94ee8f96d39fd45fed6eff4e74b6b41"
  ],
  "settings": {
    "contextTokens": 8192,
    "parallel": 1,
    "threads": 10,
    "gpuLayersRequested": 99,
    "flashAttention": "on",
    "kvK": "f16",
    "kvV": "f16",
    "reasoning": "off",
    "temperature": 0,
    "maxOutputTokens": 256,
    "cachePromptRequested": false,
    "retrievalPages": 8,
    "passageCandidatesPerField": 12,
    "answerMode": "extractive-source-ID-selection",
    "explicitAbsenceGuard": false,
    "semanticSufficiencyCheck": true,
    "sufficiencyMaxOutputTokens": 192,
    "modelPhasesWallDeadlineSeconds": 90,
    "generationSeed": 271828,
    "scoring": "kingy-common-document-grade/1.5.0; conservative unresolved bounds",
    "maximumRequestedFacts": 2,
    "repetitions": 3,
    "questionsPerRepetition": 40,
    "serverRestartBetweenRepetitions": true
  },
  "limitations": [
    "Offloaded layer count was not exposed in default original runtime logs.",
    "A one-second read-only process sample in session 3 identified active Metal compute; its overhead was not subtracted from clocks.",
    "OS/runtime binary/background load differ from the original M4 Pro, so no hardware-only speed claim follows."
  ]
}