{
  "title": "Custom LLM Harness: paired Terminal-Bench study",
  "recordType": "Curated summary of existing experiment records; not a new benchmark run",
  "experimentDate": "2026-10-01",
  "method": {
    "benchmark": "Terminal-Bench 2.0 through Harbor",
    "modelConfiguration": "Qwen3.8-27B abliterated checkpoint served through vLLM",
    "taskCount": 30,
    "attemptsPerTaskPerArm": 3,
    "arms": ["Baseline", "Optional delegation-oriented workflow"],
    "successSignal": "Each task's executable tests",
    "comparison": "Paired difference in average task success"
  },
  "results": {
    "baseline": {"scoredTrials": 90, "successes": 50, "successRate": 0.5555555556, "infrastructureInvalidTrials": 25, "trialsInvokingSecondaryWorkerCLIs": 0},
    "optionalWorkflow": {"scoredTrials": 90, "successes": 52, "successRate": 0.5777777778, "infrastructureInvalidTrials": 28, "trialsInvokingSecondaryWorkerCLIs": 0},
    "differencePercentagePoints": 2.2222222222,
    "paired95PercentIntervalPercentagePoints": [-7.7777777778, 12.2222222222],
    "interpretation": "This experiment did not establish an improvement. Available secondary worker commands were not observed being invoked."
  },
  "limitations": [
    "Selected 30-task subset, one model configuration, three attempts per task and arm.",
    "Infrastructure-invalid attempts were recorded separately and rerun; the 180 scored trials exclude 53 invalid attempts.",
    "Task success is a benchmark outcome, not evidence of general safety or honesty.",
    "The model weights, inference engine, base agent, benchmark environments, and task checkers are upstream components.",
    "Underlying source and run logs remain private; this file is a manually curated results summary."
  ]
}