{
  "started_at": "2026-09-29T08:04:54.528484+00:00",
  "finished_at": "2026-09-29T08:37:39.305923+00:00",
  "model": "gpt-oss:20b",
  "runs_per_workload": 10,
  "environment": {
    "VIBEY_OLLAMA_URL": null,
    "VIBEY_OLLAMA_MODEL": null,
    "VIBEY_OLLAMA_CONTEXT": null,
    "VIBEY_OLLAMA_OUTPUT": null,
    "VIBEY_OLLAMA_FIT": null,
    "VIBEY_OLLAMA_RETRY_THINK": null
  },
  "summary": {
    "design.interview": {
      "n": 10,
      "outcomes": {
        "ok": 10
      },
      "ok_rate": 1.0,
      "empty_content_rate": 0.0,
      "schema_violation_rate": 0.0,
      "empty_content_done_reasons": {},
      "first_call_empty_rate": 0.0,
      "first_call_done_reasons": {
        "stop": 10
      },
      "mean_seconds": 25.1,
      "mean_calls": 1.0
    },
    "design.interview (production replay)": {
      "n": 10,
      "outcomes": {
        "ok": 10
      },
      "ok_rate": 1.0,
      "empty_content_rate": 0.0,
      "schema_violation_rate": 0.0,
      "empty_content_done_reasons": {},
      "first_call_empty_rate": 0.0,
      "first_call_done_reasons": {
        "stop": 10
      },
      "mean_seconds": 77.8,
      "mean_calls": 1.0
    },
    "build.decompose": {
      "n": 10,
      "outcomes": {
        "ok": 10
      },
      "ok_rate": 1.0,
      "empty_content_rate": 0.0,
      "schema_violation_rate": 0.0,
      "empty_content_done_reasons": {},
      "first_call_empty_rate": 0.4,
      "first_call_done_reasons": {
        "length": 10
      },
      "mean_seconds": 93.5,
      "mean_calls": 2.0
    }
  },
  "runs": [
    {
      "workload": "design.interview",
      "label": "context_free/brief-0",
      "seconds": 34.32,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 4096,
          "num_predict": 2048,
          "seconds": 34.32,
          "done_reason": "stop",
          "eval_count": 992,
          "prompt_eval_count": 419,
          "content_chars": 620,
          "thinking_chars": 3983
        }
      ]
    },
    {
      "workload": "design.interview (production replay)",
      "label": "walking_skeleton",
      "seconds": 81.43,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 8192,
          "num_predict": 2048,
          "seconds": 81.43,
          "done_reason": "stop",
          "eval_count": 1977,
          "prompt_eval_count": 5627,
          "content_chars": 779,
          "thinking_chars": 8837
        }
      ]
    },
    {
      "workload": "build.decompose",
      "label": "spec-0",
      "seconds": 109.87,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 4096,
          "num_predict": 2048,
          "seconds": 88.26,
          "done_reason": "length",
          "eval_count": 2048,
          "prompt_eval_count": 887,
          "content_chars": 0,
          "thinking_chars": 8204
        },
        {
          "format": "json",
          "think": "low",
          "num_ctx": 4983,
          "num_predict": 4096,
          "seconds": 21.59,
          "done_reason": "stop",
          "eval_count": 446,
          "prompt_eval_count": 887,
          "content_chars": 1001,
          "thinking_chars": 705
        }
      ]
    },
    {
      "workload": "design.interview",
      "label": "job_story/brief-1",
      "seconds": 20.53,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 4096,
          "num_predict": 2048,
          "seconds": 20.53,
          "done_reason": "stop",
          "eval_count": 427,
          "prompt_eval_count": 424,
          "content_chars": 497,
          "thinking_chars": 1437
        }
      ]
    },
    {
      "workload": "design.interview (production replay)",
      "label": "walking_skeleton",
      "seconds": 84.24,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 8192,
          "num_predict": 2048,
          "seconds": 84.22,
          "done_reason": "stop",
          "eval_count": 1977,
          "prompt_eval_count": 5627,
          "content_chars": 779,
          "thinking_chars": 8837
        }
      ]
    },
    {
      "workload": "build.decompose",
      "label": "spec-1",
      "seconds": 99.64,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 4096,
          "num_predict": 2048,
          "seconds": 66.39,
          "done_reason": "length",
          "eval_count": 2048,
          "prompt_eval_count": 835,
          "content_chars": 316,
          "thinking_chars": 8435
        },
        {
          "format": "json",
          "think": "low",
          "num_ctx": 4931,
          "num_predict": 4096,
          "seconds": 33.25,
          "done_reason": "stop",
          "eval_count": 907,
          "prompt_eval_count": 835,
          "content_chars": 3005,
          "thinking_chars": 413
        }
      ]
    },
    {
      "workload": "design.interview",
      "label": "laddering/brief-2",
      "seconds": 58.44,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 4096,
          "num_predict": 2048,
          "seconds": 58.44,
          "done_reason": "stop",
          "eval_count": 1804,
          "prompt_eval_count": 415,
          "content_chars": 622,
          "thinking_chars": 8481
        }
      ]
    },
    {
      "workload": "design.interview (production replay)",
      "label": "walking_skeleton",
      "seconds": 76.54,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 8192,
          "num_predict": 2048,
          "seconds": 76.54,
          "done_reason": "stop",
          "eval_count": 1977,
          "prompt_eval_count": 5627,
          "content_chars": 779,
          "thinking_chars": 8837
        }
      ]
    },
    {
      "workload": "build.decompose",
      "label": "spec-2",
      "seconds": 87.85,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 4096,
          "num_predict": 2048,
          "seconds": 66.55,
          "done_reason": "length",
          "eval_count": 2048,
          "prompt_eval_count": 932,
          "content_chars": 286,
          "thinking_chars": 8360
        },
        {
          "format": "json",
          "think": "low",
          "num_ctx": 5028,
          "num_predict": 4096,
          "seconds": 21.3,
          "done_reason": "stop",
          "eval_count": 481,
          "prompt_eval_count": 932,
          "content_chars": 1434,
          "thinking_chars": 470
        }
      ]
    },
    {
      "workload": "design.interview",
      "label": "example_mapping/brief-0",
      "seconds": 19.18,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 4096,
          "num_predict": 2048,
          "seconds": 19.18,
          "done_reason": "stop",
          "eval_count": 481,
          "prompt_eval_count": 419,
          "content_chars": 644,
          "thinking_chars": 1258
        }
      ]
    },
    {
      "workload": "design.interview (production replay)",
      "label": "walking_skeleton",
      "seconds": 76.47,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 8192,
          "num_predict": 2048,
          "seconds": 76.47,
          "done_reason": "stop",
          "eval_count": 1977,
          "prompt_eval_count": 5627,
          "content_chars": 779,
          "thinking_chars": 8837
        }
      ]
    },
    {
      "workload": "build.decompose",
      "label": "spec-0",
      "seconds": 87.44,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 4096,
          "num_predict": 2048,
          "seconds": 67.59,
          "done_reason": "length",
          "eval_count": 2048,
          "prompt_eval_count": 887,
          "content_chars": 0,
          "thinking_chars": 8204
        },
        {
          "format": "json",
          "think": "low",
          "num_ctx": 4983,
          "num_predict": 4096,
          "seconds": 19.84,
          "done_reason": "stop",
          "eval_count": 446,
          "prompt_eval_count": 887,
          "content_chars": 1001,
          "thinking_chars": 705
        }
      ]
    },
    {
      "workload": "design.interview",
      "label": "walking_skeleton/brief-1",
      "seconds": 23.98,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 4096,
          "num_predict": 2048,
          "seconds": 23.98,
          "done_reason": "stop",
          "eval_count": 669,
          "prompt_eval_count": 425,
          "content_chars": 634,
          "thinking_chars": 2518
        }
      ]
    },
    {
      "workload": "design.interview (production replay)",
      "label": "walking_skeleton",
      "seconds": 76.82,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 8192,
          "num_predict": 2048,
          "seconds": 76.82,
          "done_reason": "stop",
          "eval_count": 1977,
          "prompt_eval_count": 5627,
          "content_chars": 779,
          "thinking_chars": 8837
        }
      ]
    },
    {
      "workload": "build.decompose",
      "label": "spec-1",
      "seconds": 98.69,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 4096,
          "num_predict": 2048,
          "seconds": 65.75,
          "done_reason": "length",
          "eval_count": 2048,
          "prompt_eval_count": 835,
          "content_chars": 316,
          "thinking_chars": 8435
        },
        {
          "format": "json",
          "think": "low",
          "num_ctx": 4931,
          "num_predict": 4096,
          "seconds": 32.94,
          "done_reason": "stop",
          "eval_count": 907,
          "prompt_eval_count": 835,
          "content_chars": 3005,
          "thinking_chars": 413
        }
      ]
    },
    {
      "workload": "design.interview",
      "label": "nfr_planguage/brief-2",
      "seconds": 15.94,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 4096,
          "num_predict": 2048,
          "seconds": 15.94,
          "done_reason": "stop",
          "eval_count": 406,
          "prompt_eval_count": 417,
          "content_chars": 627,
          "thinking_chars": 1231
        }
      ]
    },
    {
      "workload": "design.interview (production replay)",
      "label": "walking_skeleton",
      "seconds": 76.26,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 8192,
          "num_predict": 2048,
          "seconds": 76.26,
          "done_reason": "stop",
          "eval_count": 1977,
          "prompt_eval_count": 5627,
          "content_chars": 779,
          "thinking_chars": 8837
        }
      ]
    },
    {
      "workload": "build.decompose",
      "label": "spec-2",
      "seconds": 86.22,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 4096,
          "num_predict": 2048,
          "seconds": 65.92,
          "done_reason": "length",
          "eval_count": 2048,
          "prompt_eval_count": 932,
          "content_chars": 286,
          "thinking_chars": 8360
        },
        {
          "format": "json",
          "think": "low",
          "num_ctx": 5028,
          "num_predict": 4096,
          "seconds": 20.3,
          "done_reason": "stop",
          "eval_count": 481,
          "prompt_eval_count": 932,
          "content_chars": 1434,
          "thinking_chars": 470
        }
      ]
    },
    {
      "workload": "design.interview",
      "label": "premortem/brief-0",
      "seconds": 28.36,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 4096,
          "num_predict": 2048,
          "seconds": 28.36,
          "done_reason": "stop",
          "eval_count": 822,
          "prompt_eval_count": 420,
          "content_chars": 564,
          "thinking_chars": 3239
        }
      ]
    },
    {
      "workload": "design.interview (production replay)",
      "label": "walking_skeleton",
      "seconds": 75.81,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 8192,
          "num_predict": 2048,
          "seconds": 75.81,
          "done_reason": "stop",
          "eval_count": 1977,
          "prompt_eval_count": 5627,
          "content_chars": 779,
          "thinking_chars": 8837
        }
      ]
    },
    {
      "workload": "build.decompose",
      "label": "spec-0",
      "seconds": 91.27,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 4096,
          "num_predict": 2048,
          "seconds": 70.93,
          "done_reason": "length",
          "eval_count": 2048,
          "prompt_eval_count": 887,
          "content_chars": 0,
          "thinking_chars": 8204
        },
        {
          "format": "json",
          "think": "low",
          "num_ctx": 4983,
          "num_predict": 4096,
          "seconds": 20.33,
          "done_reason": "stop",
          "eval_count": 446,
          "prompt_eval_count": 887,
          "content_chars": 1001,
          "thinking_chars": 705
        }
      ]
    },
    {
      "workload": "design.interview",
      "label": "context_free/brief-1",
      "seconds": 14.52,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 4096,
          "num_predict": 2048,
          "seconds": 14.52,
          "done_reason": "stop",
          "eval_count": 354,
          "prompt_eval_count": 424,
          "content_chars": 430,
          "thinking_chars": 1125
        }
      ]
    },
    {
      "workload": "design.interview (production replay)",
      "label": "walking_skeleton",
      "seconds": 76.72,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 8192,
          "num_predict": 2048,
          "seconds": 76.72,
          "done_reason": "stop",
          "eval_count": 1977,
          "prompt_eval_count": 5627,
          "content_chars": 779,
          "thinking_chars": 8837
        }
      ]
    },
    {
      "workload": "build.decompose",
      "label": "spec-1",
      "seconds": 100.38,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 4096,
          "num_predict": 2048,
          "seconds": 66.78,
          "done_reason": "length",
          "eval_count": 2048,
          "prompt_eval_count": 835,
          "content_chars": 316,
          "thinking_chars": 8435
        },
        {
          "format": "json",
          "think": "low",
          "num_ctx": 4931,
          "num_predict": 4096,
          "seconds": 33.6,
          "done_reason": "stop",
          "eval_count": 907,
          "prompt_eval_count": 835,
          "content_chars": 3005,
          "thinking_chars": 413
        }
      ]
    },
    {
      "workload": "design.interview",
      "label": "job_story/brief-2",
      "seconds": 19.63,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 4096,
          "num_predict": 2048,
          "seconds": 19.63,
          "done_reason": "stop",
          "eval_count": 524,
          "prompt_eval_count": 415,
          "content_chars": 371,
          "thinking_chars": 2000
        }
      ]
    },
    {
      "workload": "design.interview (production replay)",
      "label": "walking_skeleton",
      "seconds": 76.81,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 8192,
          "num_predict": 2048,
          "seconds": 76.81,
          "done_reason": "stop",
          "eval_count": 1977,
          "prompt_eval_count": 5627,
          "content_chars": 779,
          "thinking_chars": 8837
        }
      ]
    },
    {
      "workload": "build.decompose",
      "label": "spec-2",
      "seconds": 88.11,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 4096,
          "num_predict": 2048,
          "seconds": 67.66,
          "done_reason": "length",
          "eval_count": 2048,
          "prompt_eval_count": 932,
          "content_chars": 286,
          "thinking_chars": 8360
        },
        {
          "format": "json",
          "think": "low",
          "num_ctx": 5028,
          "num_predict": 4096,
          "seconds": 20.45,
          "done_reason": "stop",
          "eval_count": 481,
          "prompt_eval_count": 932,
          "content_chars": 1434,
          "thinking_chars": 470
        }
      ]
    },
    {
      "workload": "design.interview",
      "label": "laddering/brief-0",
      "seconds": 16.2,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 4096,
          "num_predict": 2048,
          "seconds": 16.2,
          "done_reason": "stop",
          "eval_count": 407,
          "prompt_eval_count": 419,
          "content_chars": 460,
          "thinking_chars": 1320
        }
      ]
    },
    {
      "workload": "design.interview (production replay)",
      "label": "walking_skeleton",
      "seconds": 77.14,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 8192,
          "num_predict": 2048,
          "seconds": 77.14,
          "done_reason": "stop",
          "eval_count": 1977,
          "prompt_eval_count": 5627,
          "content_chars": 779,
          "thinking_chars": 8837
        }
      ]
    },
    {
      "workload": "build.decompose",
      "label": "spec-0",
      "seconds": 85.86,
      "outcome": "ok",
      "error": null,
      "exchanges": [
        {
          "format": "json",
          "think": null,
          "num_ctx": 4096,
          "num_predict": 2048,
          "seconds": 66.84,
          "done_reason": "length",
          "eval_count": 2048,
          "prompt_eval_count": 887,
          "content_chars": 0,
          "thinking_chars": 8204
        },
        {
          "format": "json",
          "think": "low",
          "num_ctx": 4983,
          "num_predict": 4096,
          "seconds": 19.02,
          "done_reason": "stop",
          "eval_count": 446,
          "prompt_eval_count": 887,
          "content_chars": 1001,
          "thinking_chars": 705
        }
      ]
    }
  ]
}