{
  "schema_version": 1,
  "historical_run_date": "2026-10-20",
  "evidence_review_date": "2026-10-22",
  "status": "Historical published observations; not independently reproduced. Original clients recovered, not reconstructed.",
  "provenance": {
    "scripts": "Byte-identical copies from the October 20 staged release bench directory and contemporaneous scratchpad. Commands recovered from that session record.",
    "results": "Unchanged published JSONL, byte-identical to the staged release; run labels were added when concatenating the three original output files.",
    "launch": "Explicit arguments and GPU environment recovered from historical command record; paths below use portable placeholders."
  },
  "model": {
    "name": "qwen3:4b",
    "blob_sha256_identifier": "3e4cb14174460404e7a233e531675303b2fbf7749c02f91864fe311ab6344e4f",
    "identifier_evidence": "Model blob filename in historical launch command and server logs. Not independently rehashed; this is not a fresh checksum verification.",
    "size": "2.5 GB, as reported in original article"
  },
  "versions_reported": {
    "ollama": "0.40.1",
    "llama_server": "0.5.0-dev / 631109b34 (bundled with Ollama)",
    "nvidia_driver": "595.99",
    "gpu": "GTX 1650, 4 GB"
  },
  "prompt": {
    "base_utf8": "Explain in plain words how a key-value cache speeds up text generation in a transformer. Explain in plain words how a key-value cache speeds up text generation in a transformer. Explain in plain words how a key-value cache speeds up text generation in a transformer. Explain in plain words how a key-value cache speeds up text generation in a transformer. ",
    "llama_request": "zero-based decimal request index + one ASCII space + base_utf8",
    "ollama_request": "base_utf8",
    "warmup": "one request with prompt hi per client invocation, same generation options, excluded from rows",
    "tokens": "Not a universal 80-token prompt. GPU server log reports 70 prompt tokens for the first numbered request; tokenizer/template dependent."
  },
  "clients": {
    "llama_server": {
      "file": "llama_server.rb",
      "endpoint": "http://127.0.0.1:8195/completion",
      "read_timeout_seconds": 600,
      "request": {
        "n_predict": 200,
        "temperature": 0,
        "seed": 1,
        "ignore_eos": true,
        "cache_prompt": false
      },
      "gpu_invocation": "SLOTS=<1|4|8> ruby llama_server.rb 1 4 8 16",
      "cpu_invocation": "SLOTS=<1|4|8> ruby llama_server.rb 1 4 8"
    },
    "ollama": {
      "file": "ollama.rb",
      "endpoint": "http://127.0.0.1:11434/api/generate",
      "read_timeout_seconds": 600,
      "request": {
        "model": "qwen3:4b",
        "stream": false,
        "think": false,
        "options": {
          "num_predict": 200,
          "temperature": 0,
          "seed": 1
        }
      },
      "invocation": "ruby ollama.rb 1 1 2 4 8",
      "cache": "No explicit cache-disabling option; identical prompts may reuse prompt cache.",
      "server_config": "Existing service defaults; complete historical environment/model template not archived in this bundle."
    }
  },
  "llama_launch": {
    "variables": {
      "SERVER": "path to historical Ollama-bundled llama-server",
      "MODEL": "path to blob identified above",
      "D": "Ollama library directory",
      "SLOTS": "1, 4 or 8"
    },
    "argv": [
      "$SERVER",
      "-m",
      "$MODEL",
      "-ngl",
      "99",
      "-np",
      "$SLOTS",
      "-c",
      "1024 * SLOTS",
      "--host",
      "127.0.0.1",
      "--port",
      "8195",
      "--metrics"
    ],
    "gpu_environment": {
      "LD_LIBRARY_PATH": "$D/cuda_v13:$D",
      "GGML_BACKEND_PATH": "$D/cuda_v13/libggml-cuda.so"
    },
    "cpu_environment": "No explicit CUDA environment overrides in recorded command; same -ngl 99 request, backend fell back to CPU.",
    "readiness": "Historical command polled /health before each client invocation.",
    "slots_note": "SLOTS labels client output only. It does not change server slots.",
    "unknowns": [
      "Full inherited environment and engine defaults",
      "Binary and CUDA library checksums",
      "Contemporaneous model bytes independently hashed",
      "Full Ollama service configuration",
      "Per-request raw responses and repeated trial variance"
    ]
  },
  "metrics": {
    "aggregate_tps": "sum returned generated tokens / monotonic batch wall time (includes queueing and HTTP)",
    "per_req_decode_tps": "Backend-reported decode rate, excludes full client queueing/latency",
    "max_latency_s": "Maximum per-request client wall time",
    "precision": "wall/latency rounded to 0.01 seconds; token rates to 0.1; rounding tolerance used by checker",
    "repetitions": "One row per setting except Ollama concurrency 1 appears twice. No repeated GPU trials."
  },
  "sha256": {
    "llama_server.rb": "3597602d824578867f273fe3e090a6ebc05d1b0bef252366e2c9176a9fae043f",
    "ollama.rb": "1ac305ddc28778f72b7b53b405b948a55a422e8132700c334a0e5bc3d4957168",
    "../inference-roadmap-bench.jsonl": "97069560d9ae83d0fb9d75e9966e92122e462924187ef4f6d813fb8c18e14ed6",
    "README.md": "ad6802d975447d46314b108990c1de48b8027f477005501fdeca9cdc2f2db3e0",
    "check.rb": "48e51680936faaf6fe0d4d875a918870604f50020a1a0b4feec4a745d555ab0c",
    "test.rb": "8291cf13efefb0651c670464717dd344a95f370b1b5d9218c1c46244fac564f6",
    "fixture.rb": "7959c6f5ce0e44c968523515eccab39db0631026148aea64d92454280649d144"
  }
}