{
  "state": "Complete",
  "baseModel": "gemma3_1b_it_q4_k_m",
  "scope": "Paired full Service generation after identical activation, including context process, planning, integrity, queue/admission, Metal inference and streamed output. Headless UI; no HTTP transport.",
  "buildProfile": "Rust development profile; Engine and sha2 opt-level=2 overrides; identical binary/build flags in both conditions",
  "datasetSha256": "618354fcaade2b7d2e494b43e66caf05c8d4e25b926a4bcbedaf49a1918544da",
  "protocolSha256": "ae280b13a202f89915e1eb2df821481cfae4f79bdc43f37177098c003b2f913f",
  "environmentEvidence": "experiment-environment.json",
  "environmentEvidenceSha256": "2215ac0723237890fcca0491c6a91d4fca604b34e573b7516aa1a8a8879790d1",
  "measurementsSha256": "3cb01ea37a42b1f59ee335bc573f6dde8550e24906c06185d8917ac6e44709ec",
  "analysisSha256": "d19dcef24b99bcccae1b2bcd44493032b36983830e863e88d01ea9e039cb6cca",
  "generations": 2040,
  "queryPairs": 1020,
  "repetitions": 3,
  "latency": {
    "baselineP95Ms": 4983.28801935,
    "candidateP95Ms": 2669.394663949996,
    "reduction": 0.4643306480410536,
    "reduction95CI": [
      0.32358956085708507,
      0.6251476488090516
    ],
    "pairs": 1020,
    "bootstrapSeed": 637110
  },
  "byRepetition": {
    "0": {
      "baselineP95Ms": 5167.947264600004,
      "candidateP95Ms": 3911.64142295,
      "reduction": 0.24309571621513848,
      "reduction95CI": [
        0.18945642770216203,
        0.3616147400014487
      ],
      "pairs": 340,
      "bootstrapSeed": 637110
    },
    "1": {
      "baselineP95Ms": 4383.2566458500005,
      "candidateP95Ms": 1566.1346518000003,
      "reduction": 0.6427006725050444,
      "reduction95CI": [
        0.6334316831628702,
        0.6512910940306428
      ],
      "pairs": 340,
      "bootstrapSeed": 637110
    },
    "2": {
      "baselineP95Ms": 5088.330983649999,
      "candidateP95Ms": 1700.05406315,
      "reduction": 0.6658916118836073,
      "reduction95CI": [
        0.6503110614285991,
        0.6778175921330353
      ],
      "pairs": 340,
      "bootstrapSeed": 637110
    }
  },
  "strata": {
    "coldRuntime": {
      "baselineP95Ms": 4604.827,
      "candidateP95Ms": 4263.810625,
      "reduction": 0.07405628376484075,
      "reduction95CI": [
        -0.0047493123902795364,
        0.37202528239079147
      ],
      "pairs": 41,
      "bootstrapSeed": 637110
    },
    "warmRuntime": {
      "baselineP95Ms": 4983.494704699999,
      "candidateP95Ms": 1698.2349583999996,
      "reduction": 0.6592281001526153,
      "reduction95CI": [
        0.6498577871790978,
        0.666146672240713
      ],
      "pairs": 979,
      "bootstrapSeed": 637110
    },
    "documents": {
      "baselineP95Ms": 5017.788868749999,
      "candidateP95Ms": 1710.2450253,
      "reduction": 0.6591636136882647,
      "reduction95CI": [
        0.6472144065437833,
        0.6660209816113408
      ],
      "pairs": 900,
      "bootstrapSeed": 637110
    },
    "memory": {
      "baselineP95Ms": 4467.85896875,
      "candidateP95Ms": 3911.92532915,
      "reduction": 0.12442954074612089,
      "reduction95CI": [
        0.07130586943769031,
        0.16250207151570015
      ],
      "pairs": 60,
      "bootstrapSeed": 637110
    },
    "shared": {
      "baselineP95Ms": 4502.1631479,
      "candidateP95Ms": 4058.717695499999,
      "reduction": 0.09849608684368594,
      "reduction95CI": [
        -0.006591325070172349,
        0.18037928525702096
      ],
      "pairs": 60,
      "bootstrapSeed": 637110
    }
  },
  "fixedQuality": {
    "exactAnswerAccuracy": 0.6411764705882353,
    "unanswerableAbstention": 0.961038961038961,
    "expectedSourceCitationRate": 0.0,
    "scope": "Exact synthetic fixture scoring; expected-source citation matching is not semantic support verification"
  },
  "compiledQuality": {
    "exactAnswerAccuracy": 0.6411764705882353,
    "unanswerableAbstention": 0.961038961038961,
    "expectedSourceCitationRate": 0.0,
    "scope": "Exact synthetic fixture scoring; expected-source citation matching is not semantic support verification"
  },
  "qualityLoss": 0.0,
  "identicalOutputs": true,
  "compilerGatePassed": true,
  "limits": [
    "Original procedural synthetic data; not independent production-domain validation.",
    "Queries predominantly require short factual answers; long conversations, software generation and broad agent tasks are outside this measured scope.",
    "Headless benchmark harness on an interactive desktop, not an isolated performance lab; UI checks and ordinary application activity occurred during the run, with all observations retained.",
    "Research scope is this Gemma/M2/8 GiB/headless development/debug profile; no general multi-platform speed claim.",
    "Production-optimized builds need separate performance validation; avoiding redundant integrity work is not a new GPU kernel or a model-quality improvement.",
    "Model selection/activation preflight excluded identically; every timed generation includes complete Service context resolution, planning, worker admission, integrity checks and streamed Metal output.",
    "Expected-source citation quality is not scored as a support guarantee. No model training or base-weight changes occurred.",
    "Confidence intervals bootstrap paired measurements on this fixed procedural benchmark; repeated queries are related, and the intervals do not estimate between-customer or between-domain uncertainty."
  ],
  "benchmarkBinarySha256": "0c5924ff6ee1576941ed4a37fc06a4ee51551637a63981c84405f487c07019c4",
  "sourceSnapshot": "source-snapshot/pins.json"
}