{
  "schemaVersion": "independent-evidence-review/1.0.0",
  "status": "pass",
  "reviewedAt": "2026-08-23T07:30:00Z",
  "reviewer": "Codex separate-path evidence recomputation",
  "independenceBoundary": "Displayed claims were recomputed directly from retained raw JSON, command JSON, telemetry CSV, artifact hashes, and the hardware/runtime manifest; candidate receipt values were not used as the measurement source.",
  "resultId": "gemma-4-31b-it/ryzen-7950x-rtx-4090-24gb-128gb",
  "receiptHashes": [
    {
      "file": "gemma4-31b-it-q4km-rtx4090-physical-context-f16-16384-v2.receipt.json",
      "runId": "gemma4-31b-it-q4km-rtx4090-physical-context-f16-16384-v2",
      "receiptHash": "sha256:08e51fc613a310b278f642d2d367d18a4a661939c209baa14e03478c77dbbfe2"
    },
    {
      "file": "gemma4-31b-it-q4km-rtx4090-physical-context-f16-32768-v2.receipt.json",
      "runId": "gemma4-31b-it-q4km-rtx4090-physical-context-f16-32768-v2",
      "receiptHash": "sha256:2d0fbb4ab97a80af985f10f2d04c498aadd2c44604c26a4590c3d3831fa78da4"
    },
    {
      "file": "gemma4-31b-it-q4km-rtx4090-physical-context-f16-8192-v2.receipt.json",
      "runId": "gemma4-31b-it-q4km-rtx4090-physical-context-f16-8192-v2",
      "receiptHash": "sha256:237d2c8b0688930d49e07cf91f91b3cc5f485dcb9fa1804defe7771322a0c3c7"
    },
    {
      "file": "gemma4-31b-it-q4km-rtx4090-physical-full-text-quality-v2.receipt.json",
      "runId": "gemma4-31b-it-q4km-rtx4090-physical-full-text-quality-v2",
      "receiptHash": "sha256:792709080c17508dcc5a3e4960af12d57b02f737fdd3554c8f0f56a10b53eb22"
    },
    {
      "file": "gemma4-31b-it-q4km-rtx4090-physical-llama-bench-f16-v2.receipt.json",
      "runId": "gemma4-31b-it-q4km-rtx4090-physical-llama-bench-f16-v2",
      "receiptHash": "sha256:dc04d73645036565a6a38cfb7f359ec5d00aabf77cb6d5b0a083d71a66340613"
    },
    {
      "file": "gemma4-31b-it-q4km-rtx4090-physical-streaming-8192-v2.receipt.json",
      "runId": "gemma4-31b-it-q4km-rtx4090-physical-streaming-8192-v2",
      "receiptHash": "sha256:c7108f68c5d92ffed68efeaa0735ef9300b3dbb2693d19801500975cca9dcd63"
    }
  ],
  "decisiveSources": [
    {
      "path": "work/evidence/results/gemma4/performance/q8_0-q8_0/context-q8_0-65536.json",
      "sha256": "b34760bd5e03af0681e8258994ccfd42a5247f6ca32e6713d61e5ac8911b5df0",
      "sizeBytes": 2339
    },
    {
      "path": "work/evidence/results/gemma4/performance/q8_0-q8_0/context-q8_0-65536-command.json",
      "sha256": "61b0bd406d854ac55a40c89a5462e2a980260782f8d6d862a70056f186b1e1e4",
      "sizeBytes": 811
    },
    {
      "path": "work/evidence/results/gemma4/stream-timing/summary.json",
      "sha256": "3af590cda2fe8b4160fb45b1cd6435ba74062b72cbac3f36c879ee4b61119e3a",
      "sizeBytes": 2294
    },
    {
      "path": "work/evidence/results/gemma4/stream-timing/stream-timing-gpu.csv",
      "sha256": "01500accc381edc0115c77eac92e9c4b98c0119602f7e0318124bb93aab63c6f",
      "sizeBytes": 6413
    },
    {
      "path": "work/evidence/results/gemma4/performance/f16-f16/llama-bench.stdout.json",
      "sha256": "71b7b85f7942a18678e041b991e16a3b9bd913446f0535727668650df72db6cd",
      "sizeBytes": 4994
    },
    {
      "path": "work/evidence/sources/artifacts.sha256",
      "sha256": "9e87d3e7067ec5764b7484d9e20ad8371c6cee886c8598da1d4b0efee492d764",
      "sizeBytes": 865
    },
    {
      "path": "work/hardware-runtime-manifest.md",
      "sha256": "b679a92925c9d626c6a2d66d9dea8c11a2d0181cb8a4241b83e2cbf05a827b44",
      "sizeBytes": 1940
    },
    {
      "path": "work/evidence/results/gemma4/performance/f16-f16/context-f16-65536.json",
      "sha256": "c55364e3b0e5d06bf74083ad5cdb2f6f7b59015b1769d369b912868e0344b21b",
      "sizeBytes": 2202
    },
    {
      "path": "work/evidence/results/gemma4/performance/f16-f16/context-f16-65536.log",
      "sha256": "cdbf56bd2e0420c43a0e24dbf8f245e72806b6706eb2abb9c5e44fabdf7e29c3",
      "sizeBytes": 2029
    }
  ],
  "findings": {
    "runtime": "llama.cpp b10453, commit 3cb7ffb1a1f612d5e4a46244ae5a3c77ad934a70, CUDA compute architecture 89",
    "artifact": {
      "filename": "gemma-4-31B-it-Q4_K_M.gguf",
      "sha256": "38bd64c852c4b460434cc7162fa9bdcf242faf86502581a754cb72956bb17f84",
      "quantization": "Q4_K_M"
    },
    "context": {
      "allocatedTokens": 65536,
      "kvCache": "Q8_0",
      "passed": true,
      "inputTokens": 63482,
      "peakGpuMiB": 21956
    },
    "throughput": {
      "profile": "llama-bench 256-token synthetic decode, five repetitions",
      "medianTokensPerSecond": 44.9998,
      "samplesTokensPerSecond": [
        44.9688,
        45.0015,
        44.9998,
        44.9991,
        45.0024
      ]
    },
    "power": {
      "profile": "8K streaming request series after one excluded warm-up",
      "medianWatts": 63.79,
      "peakWatts": 333.83,
      "samples": 160
    },
    "failureCondition": "64K with FP16 KV failed during context allocation: cudaMalloc could not allocate a 1,200 MiB KV-cache buffer. The same 64K workload passed after changing only KV cache to Q8_0; 32K FP16 KV also passed."
  }
}