{
  "title": "Terse-Coder Q6_K with native MTP \u2014 FaithVoice parser benchmark",
  "testedBy": "Jason",
  "date": "2026-10-02",
  "startedAt": "2026-10-02T19:33:00.747336+00:00",
  "endedAt": "2026-10-02T19:44:23.879296+00:00",
  "artifact": {
    "repo": "Shockem/Qwen3.8-27b-Terse-Coder-GGUF",
    "revision": "7602eadf6a90853492e4e8a377c2a2f531bf21e2",
    "file": "Qwen3.8-27b-Terse-Coder.Q6_K.gguf",
    "size": 22430999904,
    "sha256": "3164a8d8cc7f5c669b7a6de57ce9e5db126d73b8d39a0f424a2b9c649f25df07"
  },
  "hardware": {
    "chip": "Apple M5 Pro",
    "memoryGiB": 48,
    "modelIdentifier": "Mac17,8",
    "provenance": "Same Mac local server/model path and post-run sysctl corroboration; run.json did not record a separate hardware snapshot."
  },
  "task": "Canonicalize accepted verse-range separators and surrounding whitespace to ASCII hyphen while preserving visible linked Markdown, URL round trips and ambiguity protection.",
  "baselineCommit": "2c7aebd67d44e4858f94dce73de5c4f99173fc57",
  "promptSHA256": "a9bd80dad2cf9792ebaacdde38e71a2a0b4c85113f578954c5ced789a9710a9d",
  "runtime": "LM Studio bundled llama.cpp 2.37.0, commit 8172e65; standalone server",
  "hermesCommit": "3aee290899e478c5fdfb6a241ef62758a49829b3",
  "template": {
    "repo": "Shockem/froggeric-terse-coder",
    "revision": "c557b014b6675a87683b45d0800a68842311ffe2",
    "sha256": "1ecd24354e9069f30c31627bf2f6c314f066ab633134edeb8ea68118075ab50e",
    "renderedUseVerified": true
  },
  "settings": {
    "reasoning": "medium",
    "contextTokens": 118016,
    "kvCache": "q8_0 K/V",
    "mtp": "native draft-mtp",
    "mtpMaxDraftTokens": 4,
    "backendDraftSampling": false,
    "visionProjector": false,
    "maxResponseTokens": 8192,
    "temperature": 0.6,
    "top_p": 0.95,
    "top_k": 20,
    "min_p": 0.05,
    "presence_penalty": 1.0
  },
  "result": {
    "outcome": "pass",
    "agentElapsedSeconds": 683.1314880419998,
    "outputTokens": 7427,
    "toolCalls": 19,
    "apiCalls": 19,
    "independentFocusedTests": 92,
    "testFailures": 0,
    "independentDebugBuild": "passed",
    "originalPatchSHA256": "b64feade024cf55fe556a295b800ba5d8cc4d55b6fa3396b481aac2e6034b700",
    "fixedFixtureSHA256": "f8c54b591bed62ee5fd7eec5e81e7c9e88b6481511b726369f9f31e768fc7f64",
    "independentTestCommandExit": 0,
    "independentBuildCommandExit": 0,
    "mtpAcceptedDrafts": 5332,
    "mtpDrafts": 8536,
    "mtpAcceptancePercent": 62.46,
    "costUSD": null,
    "separateReasoningTokens": null
  },
  "previousNoMTPRun": {
    "agentElapsedSeconds": 783.3782455839998,
    "outputTokens": 5972,
    "toolCalls": 22,
    "apiCalls": 21,
    "independentFocusedTests": 91,
    "testFailures": 0,
    "independentDebugBuild": "passed"
  },
  "caveats": [
    "Single complete-configuration run, not an isolated MTP causal-speed experiment or a model-quality ranking.",
    "Historical Signal used Xcode beta 6, buffered proxy, available graph and a vision projector; current Terse used released Xcode27.0, transparent streaming, native-search fallback and no vision projector.",
    "Model patch preserved before independent fixed evaluation; no evaluator-guided repair and no FaithVoice merge or commit.",
    "Agent recovered unaided from obsolete Xcode path; piped commands without pipefail and default DerivedData outside worktree remain instruction-following caveats.",
    "Focused parser/test and Debug-build qualification only; no release, physical-device, security or full-app certification.",
    "API reasoning-token zero and estimated-cost zero are not separate reasoning measurement or verified zero operating cost.",
    "Historical 14-setup companion scope and 7/14 reported passes are unchanged."
  ]
}