{
  "study": "Plain-prose storyteller comparison",
  "date": "2026-09-27",
  "training_examples": 1250,
  "base_model": "Qwen3.8-27B",
  "hardware": "Two RTX 3090 GPUs, 24 GB each; 128 GB host RAM",
  "method": "Fresh QLoRA SFT; completion-only loss; source-separated training and evaluation",
  "training_seconds": 16715.592765569687,
  "optimizer_steps": 313,
  "reported_epochs": 1.9984,
  "splits": {
    "validation": {
      "cases": 40,
      "preference": {
        "left": "base",
        "right": "s1250",
        "left_preferred": 13,
        "right_preferred": 24,
        "ties": 3
      },
      "arms": {
        "base": {
          "outcomes": {
            "pass": 25,
            "fail": 12,
            "uncertain": 3
          },
          "median_seconds": 5.83235,
          "median_words": 136.0
        },
        "s1250": {
          "outcomes": {
            "fail": 6,
            "pass": 33,
            "uncertain": 1
          },
          "median_seconds": 1.8433000000000002,
          "median_words": 16.0
        }
      }
    },
    "test": {
      "cases": 40,
      "preference": {
        "left": "base",
        "right": "s1250",
        "left_preferred": 19,
        "right_preferred": 19,
        "ties": 2
      },
      "arms": {
        "base": {
          "outcomes": {
            "uncertain": 5,
            "pass": 27,
            "fail": 8
          },
          "median_seconds": 7.4651,
          "median_words": 159.5
        },
        "s1250": {
          "outcomes": {
            "fail": 9,
            "pass": 30,
            "uncertain": 1
          },
          "median_seconds": 1.90015,
          "median_words": 18.5
        }
      }
    }
  },
  "review": "80 scenarios audited and independently cross-reviewed by Astra High. Answers judged by the primary assistant with model identity hidden until each split was graded; no human calibration and not a fresh independent reference-blind judge.",
  "limits": [
    "One generation per case, one sampling seed.",
    "Four held-out source families; cases are correlated.",
    "Writer-only evaluation on curated context, not live audio or full copilot quality.",
    "Shorter outputs explain much of the latency change; no throughput improvement established."
  ],
  "decision": "No clear held-out storytelling improvement. Retained as experimental; subsequently enabled for an explicit user-requested private interactive trial."
}