@cassiomc1/forgeloop 1.7.0 → 1.8.0

package/benchmarks/execution-profiles/results/aggregate/codex-tail-repeat20-20260831/novatask-saas-landing-page.json

@@ -0,0 +1,1634 @@
+{
+  "schemaVersion": 1,
+  "benchmarkVersion": "2",
+  "runSetId": "codex-tail-repeat20-20260831",
+  "scenarioId": "novatask-saas-landing-page",
+  "expectedProfile": "light",
+  "modeAggregates": {
+    "direct": {
+      "runCount": 20,
+      "verificationSuccessRate": 0.95,
+      "usageSources": [
+        "HOST_REPORTED"
+      ],
+      "totalTokens": {
+        "count": 20,
+        "average": 152671.15,
+        "minimum": 72198,
+        "p25": 150716.25,
+        "p50": 153845,
+        "p75": 174624.25,
+        "p90": 177974.9,
+        "p95": 181062.65,
+        "maximum": 220766,
+        "iqr": 23908,
+        "mad": 15286.5,
+        "outlierCount": 1
+      },
+      "wallClockMs": {
+        "count": 20,
+        "average": 80965.1747,
+        "minimum": 65172.3162,
+        "p25": 77886.9855,
+        "p50": 80735.4863,
+        "p75": 83638.2964,
+        "p90": 90836.9517,
+        "p95": 91245.5039,
+        "maximum": 91753.5692,
+        "iqr": 5751.3109,
+        "mad": 3129.6361,
+        "outlierCount": 0
+      },
+      "verificationCycles": {
+        "count": 20,
+        "average": 1,
+        "minimum": 1,
+        "p25": 1,
+        "p50": 1,
+        "p75": 1,
+        "p90": 1,
+        "p95": 1,
+        "maximum": 1,
+        "iqr": 0,
+        "mad": 0,
+        "outlierCount": 0
+      },
+      "comparableSteps": {
+        "count": 20,
+        "average": 11,
+        "minimum": 11,
+        "p25": 11,
+        "p50": 11,
+        "p75": 11,
+        "p90": 11,
+        "p95": 11,
+        "maximum": 11,
+        "iqr": 0,
+        "mad": 0,
+        "outlierCount": 0
+      },
+      "tokensPerComparableStep": {
+        "count": 20,
+        "average": 13879.1955,
+        "minimum": 6563.454545454545,
+        "p25": 13701.4773,
+        "p50": 13985.9091,
+        "p75": 15874.9318,
+        "p90": 16179.5364,
+        "p95": 16460.2409,
+        "maximum": 20069.636363636364,
+        "iqr": 2173.4545,
+        "mad": 1389.6818,
+        "outlierCount": 1
+      },
+      "diagnostics": {
+        "verificationCycles": {
+          "count": 20,
+          "average": 1,
+          "minimum": 1,
+          "p25": 1,
+          "p50": 1,
+          "p75": 1,
+          "p90": 1,
+          "p95": 1,
+          "maximum": 1,
+          "iqr": 0,
+          "mad": 0,
+          "outlierCount": 0
+        },
+        "modelTurns": {
+          "count": 20,
+          "average": 1,
+          "minimum": 1,
+          "p25": 1,
+          "p50": 1,
+          "p75": 1,
+          "p90": 1,
+          "p95": 1,
+          "maximum": 1,
+          "iqr": 0,
+          "mad": 0,
+          "outlierCount": 0
+        },
+        "toolCalls": {
+          "count": 20,
+          "average": 12.35,
+          "minimum": 6,
+          "p25": 11,
+          "p50": 12.5,
+          "p75": 14,
+          "p90": 14.3,
+          "p95": 17,
+          "maximum": 17,
+          "iqr": 3,
+          "mad": 1.5,
+          "outlierCount": 0
+        },
+        "retries": {
+          "count": 0,
+          "average": null,
+          "minimum": null,
+          "p25": null,
+          "p50": null,
+          "p75": null,
+          "p90": null,
+          "p95": null,
+          "maximum": null,
+          "iqr": null,
+          "mad": null,
+          "outlierCount": 0
+        },
+        "correctionCycles": {
+          "count": 0,
+          "average": null,
+          "minimum": null,
+          "p25": null,
+          "p50": null,
+          "p75": null,
+          "p90": null,
+          "p95": null,
+          "maximum": null,
+          "iqr": null,
+          "mad": null,
+          "outlierCount": 0
+        },
+        "filesRead": {
+          "count": 0,
+          "average": null,
+          "minimum": null,
+          "p25": null,
+          "p50": null,
+          "p75": null,
+          "p90": null,
+          "p95": null,
+          "maximum": null,
+          "iqr": null,
+          "mad": null,
+          "outlierCount": 0
+        },
+        "filesWritten": {
+          "count": 20,
+          "average": 2.85,
+          "minimum": 0,
+          "p25": 3,
+          "p50": 3,
+          "p75": 3,
+          "p90": 3,
+          "p95": 3,
+          "maximum": 3,
+          "iqr": 0,
+          "mad": 0,
+          "outlierCount": 0
+        },
+        "contextRefreshes": {
+          "count": 0,
+          "average": null,
+          "minimum": null,
+          "p25": null,
+          "p50": null,
+          "p75": null,
+          "p90": null,
+          "p95": null,
+          "maximum": null,
+          "iqr": null,
+          "mad": null,
+          "outlierCount": 0
+        }
+      },
+      "contextUsage": {
+        "sources": [
+          "UNKNOWN"
+        ],
+        "profiles": [],
+        "measuredRuns": 0,
+        "totalTokens": {
+          "count": 0,
+          "average": null,
+          "minimum": null,
+          "p25": null,
+          "p50": null,
+          "p75": null,
+          "p90": null,
+          "p95": null,
+          "maximum": null,
+          "iqr": null,
+          "mad": null,
+          "outlierCount": 0
+        },
+        "items": {
+          "taskContext": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "guides": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "history": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "protocolInstructions": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "repositoryContext": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "other": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          }
+        }
+      },
+      "quality": {
+        "sources": [
+          "EXTERNAL_REPORTED"
+        ],
+        "scores": {
+          "visualQuality": {
+            "count": 20,
+            "average": 4.305,
+            "minimum": 4.1,
+            "p25": 4.2,
+            "p50": 4.3,
+            "p75": 4.4,
+            "p90": 4.4,
+            "p95": 4.41,
+            "maximum": 4.6,
+            "iqr": 0.2,
+            "mad": 0.1,
+            "outlierCount": 0
+          },
+          "responsiveQuality": {
+            "count": 20,
+            "average": 4.52,
+            "minimum": 4.3,
+            "p25": 4.475,
+            "p50": 4.5,
+            "p75": 4.6,
+            "p90": 4.7,
+            "p95": 4.7,
+            "maximum": 4.7,
+            "iqr": 0.125,
+            "mad": 0.1,
+            "outlierCount": 0
+          },
+          "accessibility": {
+            "count": 20,
+            "average": 4.63,
+            "minimum": 4.4,
+            "p25": 4.6,
+            "p50": 4.6,
+            "p75": 4.7,
+            "p90": 4.8,
+            "p95": 4.8,
+            "maximum": 4.8,
+            "iqr": 0.1,
+            "mad": 0.1,
+            "outlierCount": 0
+          },
+          "interactionPolish": {
+            "count": 20,
+            "average": 4.095,
+            "minimum": 3.9,
+            "p25": 4,
+            "p50": 4.1,
+            "p75": 4.2,
+            "p90": 4.21,
+            "p95": 4.3,
+            "maximum": 4.3,
+            "iqr": 0.2,
+            "mad": 0.1,
+            "outlierCount": 0
+          },
+          "requirementsCompleteness": {
+            "count": 20,
+            "average": 4.755,
+            "minimum": 4.5,
+            "p25": 4.7,
+            "p50": 4.8,
+            "p75": 4.8,
+            "p90": 4.8,
+            "p95": 4.81,
+            "maximum": 5,
+            "iqr": 0.1,
+            "mad": 0,
+            "outlierCount": 1
+          }
+        }
+      }
+    },
+    "forgeloopBalanced": {
+      "runCount": 20,
+      "verificationSuccessRate": 0.95,
+      "usageSources": [
+        "HOST_REPORTED"
+      ],
+      "totalTokens": {
+        "count": 20,
+        "average": 159427.6,
+        "minimum": 71975,
+        "p25": 152393,
+        "p50": 158196,
+        "p75": 173189.5,
+        "p90": 196239.5,
+        "p95": 216572.1,
+        "maximum": 230539,
+        "iqr": 20796.5,
+        "mad": 13059,
+        "outlierCount": 2
+      },
+      "wallClockMs": {
+        "count": 20,
+        "average": 84810.8939,
+        "minimum": 69466.9672,
+        "p25": 80083.0198,
+        "p50": 82750.7952,
+        "p75": 89225.8956,
+        "p90": 96414.4547,
+        "p95": 100892.4798,
+        "maximum": 102134.1942,
+        "iqr": 9142.8758,
+        "mad": 4782.406,
+        "outlierCount": 0
+      },
+      "verificationCycles": {
+        "count": 20,
+        "average": 1,
+        "minimum": 1,
+        "p25": 1,
+        "p50": 1,
+        "p75": 1,
+        "p90": 1,
+        "p95": 1,
+        "maximum": 1,
+        "iqr": 0,
+        "mad": 0,
+        "outlierCount": 0
+      },
+      "comparableSteps": {
+        "count": 20,
+        "average": 11,
+        "minimum": 11,
+        "p25": 11,
+        "p50": 11,
+        "p75": 11,
+        "p90": 11,
+        "p95": 11,
+        "maximum": 11,
+        "iqr": 0,
+        "mad": 0,
+        "outlierCount": 0
+      },
+      "tokensPerComparableStep": {
+        "count": 20,
+        "average": 14493.4182,
+        "minimum": 6543.181818181818,
+        "p25": 13853.9091,
+        "p50": 14381.4545,
+        "p75": 15744.5,
+        "p90": 17839.9545,
+        "p95": 19688.3727,
+        "maximum": 20958.090909090908,
+        "iqr": 1890.5909,
+        "mad": 1187.1818,
+        "outlierCount": 2
+      },
+      "diagnostics": {
+        "verificationCycles": {
+          "count": 20,
+          "average": 1,
+          "minimum": 1,
+          "p25": 1,
+          "p50": 1,
+          "p75": 1,
+          "p90": 1,
+          "p95": 1,
+          "maximum": 1,
+          "iqr": 0,
+          "mad": 0,
+          "outlierCount": 0
+        },
+        "modelTurns": {
+          "count": 20,
+          "average": 1,
+          "minimum": 1,
+          "p25": 1,
+          "p50": 1,
+          "p75": 1,
+          "p90": 1,
+          "p95": 1,
+          "maximum": 1,
+          "iqr": 0,
+          "mad": 0,
+          "outlierCount": 0
+        },
+        "toolCalls": {
+          "count": 20,
+          "average": 12.35,
+          "minimum": 8,
+          "p25": 10,
+          "p50": 12.5,
+          "p75": 14,
+          "p90": 16,
+          "p95": 16.05,
+          "maximum": 17,
+          "iqr": 4,
+          "mad": 2,
+          "outlierCount": 0
+        },
+        "retries": {
+          "count": 0,
+          "average": null,
+          "minimum": null,
+          "p25": null,
+          "p50": null,
+          "p75": null,
+          "p90": null,
+          "p95": null,
+          "maximum": null,
+          "iqr": null,
+          "mad": null,
+          "outlierCount": 0
+        },
+        "correctionCycles": {
+          "count": 0,
+          "average": null,
+          "minimum": null,
+          "p25": null,
+          "p50": null,
+          "p75": null,
+          "p90": null,
+          "p95": null,
+          "maximum": null,
+          "iqr": null,
+          "mad": null,
+          "outlierCount": 0
+        },
+        "filesRead": {
+          "count": 0,
+          "average": null,
+          "minimum": null,
+          "p25": null,
+          "p50": null,
+          "p75": null,
+          "p90": null,
+          "p95": null,
+          "maximum": null,
+          "iqr": null,
+          "mad": null,
+          "outlierCount": 0
+        },
+        "filesWritten": {
+          "count": 20,
+          "average": 2.85,
+          "minimum": 0,
+          "p25": 3,
+          "p50": 3,
+          "p75": 3,
+          "p90": 3,
+          "p95": 3,
+          "maximum": 3,
+          "iqr": 0,
+          "mad": 0,
+          "outlierCount": 0
+        },
+        "contextRefreshes": {
+          "count": 0,
+          "average": null,
+          "minimum": null,
+          "p25": null,
+          "p50": null,
+          "p75": null,
+          "p90": null,
+          "p95": null,
+          "maximum": null,
+          "iqr": null,
+          "mad": null,
+          "outlierCount": 0
+        }
+      },
+      "contextUsage": {
+        "sources": [
+          "UNKNOWN"
+        ],
+        "profiles": [
+          "balanced"
+        ],
+        "measuredRuns": 0,
+        "totalTokens": {
+          "count": 0,
+          "average": null,
+          "minimum": null,
+          "p25": null,
+          "p50": null,
+          "p75": null,
+          "p90": null,
+          "p95": null,
+          "maximum": null,
+          "iqr": null,
+          "mad": null,
+          "outlierCount": 0
+        },
+        "items": {
+          "taskContext": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "guides": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "history": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "protocolInstructions": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "repositoryContext": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "other": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          }
+        }
+      },
+      "quality": {
+        "sources": [
+          "EXTERNAL_REPORTED"
+        ],
+        "scores": {
+          "visualQuality": {
+            "count": 20,
+            "average": 4.455,
+            "minimum": 4.1,
+            "p25": 4.3,
+            "p50": 4.5,
+            "p75": 4.6,
+            "p90": 4.61,
+            "p95": 4.7,
+            "maximum": 4.7,
+            "iqr": 0.3,
+            "mad": 0.1,
+            "outlierCount": 0
+          },
+          "responsiveQuality": {
+            "count": 20,
+            "average": 4.47,
+            "minimum": 3.4,
+            "p25": 4.4,
+            "p50": 4.5,
+            "p75": 4.7,
+            "p90": 4.71,
+            "p95": 4.8,
+            "maximum": 4.8,
+            "iqr": 0.3,
+            "mad": 0.1,
+            "outlierCount": 0
+          },
+          "accessibility": {
+            "count": 20,
+            "average": 4.67,
+            "minimum": 4.4,
+            "p25": 4.6,
+            "p50": 4.7,
+            "p75": 4.8,
+            "p90": 4.8,
+            "p95": 4.805,
+            "maximum": 4.9,
+            "iqr": 0.2,
+            "mad": 0.1,
+            "outlierCount": 0
+          },
+          "interactionPolish": {
+            "count": 20,
+            "average": 4.26,
+            "minimum": 4,
+            "p25": 4.1,
+            "p50": 4.3,
+            "p75": 4.4,
+            "p90": 4.41,
+            "p95": 4.5,
+            "maximum": 4.5,
+            "iqr": 0.3,
+            "mad": 0.1,
+            "outlierCount": 0
+          },
+          "requirementsCompleteness": {
+            "count": 20,
+            "average": 4.785,
+            "minimum": 4.5,
+            "p25": 4.775,
+            "p50": 4.8,
+            "p75": 4.9,
+            "p90": 4.9,
+            "p95": 4.9,
+            "maximum": 4.9,
+            "iqr": 0.125,
+            "mad": 0.1,
+            "outlierCount": 0
+          }
+        }
+      }
+    },
+    "forgeloopAdaptive": {
+      "runCount": 20,
+      "verificationSuccessRate": 0.95,
+      "usageSources": [
+        "HOST_REPORTED"
+      ],
+      "totalTokens": {
+        "count": 20,
+        "average": 147581.45,
+        "minimum": 68057,
+        "p25": 133976,
+        "p50": 157093.5,
+        "p75": 180443.5,
+        "p90": 199372.1,
+        "p95": 201132.1,
+        "maximum": 203604,
+        "iqr": 46467.5,
+        "mad": 26583.5,
+        "outlierCount": 0
+      },
+      "wallClockMs": {
+        "count": 20,
+        "average": 83390.6459,
+        "minimum": 66162.1262,
+        "p25": 77304.3993,
+        "p50": 82684.9781,
+        "p75": 87741.7117,
+        "p90": 94573.711,
+        "p95": 97980.7131,
+        "maximum": 98330.924,
+        "iqr": 10437.3124,
+        "mad": 5930.316,
+        "outlierCount": 0
+      },
+      "verificationCycles": {
+        "count": 20,
+        "average": 1,
+        "minimum": 1,
+        "p25": 1,
+        "p50": 1,
+        "p75": 1,
+        "p90": 1,
+        "p95": 1,
+        "maximum": 1,
+        "iqr": 0,
+        "mad": 0,
+        "outlierCount": 0
+      },
+      "comparableSteps": {
+        "count": 20,
+        "average": 11,
+        "minimum": 11,
+        "p25": 11,
+        "p50": 11,
+        "p75": 11,
+        "p90": 11,
+        "p95": 11,
+        "maximum": 11,
+        "iqr": 0,
+        "mad": 0,
+        "outlierCount": 0
+      },
+      "tokensPerComparableStep": {
+        "count": 20,
+        "average": 13416.4955,
+        "minimum": 6187,
+        "p25": 12179.6364,
+        "p50": 14281.2273,
+        "p75": 16403.9545,
+        "p90": 18124.7364,
+        "p95": 18284.7364,
+        "maximum": 18509.454545454544,
+        "iqr": 4224.3182,
+        "mad": 2416.6818,
+        "outlierCount": 0
+      },
+      "diagnostics": {
+        "verificationCycles": {
+          "count": 20,
+          "average": 1,
+          "minimum": 1,
+          "p25": 1,
+          "p50": 1,
+          "p75": 1,
+          "p90": 1,
+          "p95": 1,
+          "maximum": 1,
+          "iqr": 0,
+          "mad": 0,
+          "outlierCount": 0
+        },
+        "modelTurns": {
+          "count": 20,
+          "average": 1,
+          "minimum": 1,
+          "p25": 1,
+          "p50": 1,
+          "p75": 1,
+          "p90": 1,
+          "p95": 1,
+          "maximum": 1,
+          "iqr": 0,
+          "mad": 0,
+          "outlierCount": 0
+        },
+        "toolCalls": {
+          "count": 20,
+          "average": 11.95,
+          "minimum": 7,
+          "p25": 10,
+          "p50": 12.5,
+          "p75": 14,
+          "p90": 15.1,
+          "p95": 16,
+          "maximum": 16,
+          "iqr": 4,
+          "mad": 2.5,
+          "outlierCount": 0
+        },
+        "retries": {
+          "count": 0,
+          "average": null,
+          "minimum": null,
+          "p25": null,
+          "p50": null,
+          "p75": null,
+          "p90": null,
+          "p95": null,
+          "maximum": null,
+          "iqr": null,
+          "mad": null,
+          "outlierCount": 0
+        },
+        "correctionCycles": {
+          "count": 0,
+          "average": null,
+          "minimum": null,
+          "p25": null,
+          "p50": null,
+          "p75": null,
+          "p90": null,
+          "p95": null,
+          "maximum": null,
+          "iqr": null,
+          "mad": null,
+          "outlierCount": 0
+        },
+        "filesRead": {
+          "count": 0,
+          "average": null,
+          "minimum": null,
+          "p25": null,
+          "p50": null,
+          "p75": null,
+          "p90": null,
+          "p95": null,
+          "maximum": null,
+          "iqr": null,
+          "mad": null,
+          "outlierCount": 0
+        },
+        "filesWritten": {
+          "count": 20,
+          "average": 2.85,
+          "minimum": 0,
+          "p25": 3,
+          "p50": 3,
+          "p75": 3,
+          "p90": 3,
+          "p95": 3,
+          "maximum": 3,
+          "iqr": 0,
+          "mad": 0,
+          "outlierCount": 0
+        },
+        "contextRefreshes": {
+          "count": 0,
+          "average": null,
+          "minimum": null,
+          "p25": null,
+          "p50": null,
+          "p75": null,
+          "p90": null,
+          "p95": null,
+          "maximum": null,
+          "iqr": null,
+          "mad": null,
+          "outlierCount": 0
+        }
+      },
+      "contextUsage": {
+        "sources": [
+          "UNKNOWN"
+        ],
+        "profiles": [
+          "light"
+        ],
+        "measuredRuns": 0,
+        "totalTokens": {
+          "count": 0,
+          "average": null,
+          "minimum": null,
+          "p25": null,
+          "p50": null,
+          "p75": null,
+          "p90": null,
+          "p95": null,
+          "maximum": null,
+          "iqr": null,
+          "mad": null,
+          "outlierCount": 0
+        },
+        "items": {
+          "taskContext": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "guides": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "history": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "protocolInstructions": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "repositoryContext": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "other": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          }
+        }
+      },
+      "quality": {
+        "sources": [
+          "EXTERNAL_REPORTED"
+        ],
+        "scores": {
+          "visualQuality": {
+            "count": 20,
+            "average": 4.52,
+            "minimum": 4,
+            "p25": 4.4,
+            "p50": 4.55,
+            "p75": 4.7,
+            "p90": 4.8,
+            "p95": 4.8,
+            "maximum": 4.8,
+            "iqr": 0.3,
+            "mad": 0.15,
+            "outlierCount": 0
+          },
+          "responsiveQuality": {
+            "count": 20,
+            "average": 4.455,
+            "minimum": 2.6,
+            "p25": 4.5,
+            "p50": 4.6,
+            "p75": 4.7,
+            "p90": 4.8,
+            "p95": 4.8,
+            "maximum": 4.8,
+            "iqr": 0.2,
+            "mad": 0.1,
+            "outlierCount": 0
+          },
+          "accessibility": {
+            "count": 20,
+            "average": 4.54,
+            "minimum": 4.1,
+            "p25": 4.5,
+            "p50": 4.6,
+            "p75": 4.625,
+            "p90": 4.7,
+            "p95": 4.705,
+            "maximum": 4.8,
+            "iqr": 0.125,
+            "mad": 0.1,
+            "outlierCount": 0
+          },
+          "interactionPolish": {
+            "count": 20,
+            "average": 4.17,
+            "minimum": 3.2,
+            "p25": 4.075,
+            "p50": 4.25,
+            "p75": 4.4,
+            "p90": 4.41,
+            "p95": 4.5,
+            "maximum": 4.5,
+            "iqr": 0.325,
+            "mad": 0.15,
+            "outlierCount": 0
+          },
+          "requirementsCompleteness": {
+            "count": 20,
+            "average": 4.65,
+            "minimum": 3.9,
+            "p25": 4.675,
+            "p50": 4.7,
+            "p75": 4.8,
+            "p90": 4.9,
+            "p95": 4.9,
+            "maximum": 4.9,
+            "iqr": 0.125,
+            "mad": 0.1,
+            "outlierCount": 0
+          }
+        }
+      }
+    }
+  },
+  "comparisons": {
+    "direct": null,
+    "forgeloopBalanced": {
+      "comparablePairs": 18,
+      "tokenComparablePairs": 18,
+      "timeComparablePairs": 18,
+      "tokenOverheadPercent": {
+        "p50": 5.0916,
+        "p95": 90.0598
+      },
+      "pairedOverheadPercent": {
+        "p50": 5.0916,
+        "p95": 90.0598
+      },
+      "distributionDeltaPercent": {
+        "p50": 2.3522,
+        "p95": 17.7068
+      },
+      "timeOverheadPercent": {
+        "p50": 0.7028,
+        "p95": 28.6809
+      },
+      "pairedRatioDiagnostics": {
+        "pairCount": 18,
+        "baselineMinimum": 72198,
+        "baselineP25": 152608,
+        "baselineP50": 154560.5,
+        "lowBaselineThreshold": 92736.3,
+        "lowBaselinePairCount": 2
+      },
+      "pairedRuns": [
+        {
+          "runIndex": 1,
+          "directTokens": 175336,
+          "candidateTokens": 71975,
+          "absoluteDeltaTokens": -103361,
+          "pairedOverheadPercent": -58.9502,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 2,
+          "directTokens": 152599,
+          "candidateTokens": 163374,
+          "absoluteDeltaTokens": 10775,
+          "pairedOverheadPercent": 7.061,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 3,
+          "directTokens": 177864,
+          "candidateTokens": 151988,
+          "absoluteDeltaTokens": -25876,
+          "pairedOverheadPercent": -14.5482,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 4,
+          "directTokens": 80306,
+          "candidateTokens": 176533,
+          "absoluteDeltaTokens": 96227,
+          "pairedOverheadPercent": 119.8254,
+          "baselineRegime": "LOW_BASELINE_TOKEN_REGIME"
+        },
+        {
+          "runIndex": 5,
+          "directTokens": 153385,
+          "candidateTokens": 194062,
+          "absoluteDeltaTokens": 40677,
+          "pairedOverheadPercent": 26.5195,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 6,
+          "directTokens": 154816,
+          "candidateTokens": 170435,
+          "absoluteDeltaTokens": 15619,
+          "pairedOverheadPercent": 10.0888,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 7,
+          "directTokens": 152635,
+          "candidateTokens": 157336,
+          "absoluteDeltaTokens": 4701,
+          "pairedOverheadPercent": 3.0799,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 8,
+          "directTokens": 169212,
+          "candidateTokens": 152528,
+          "absoluteDeltaTokens": -16684,
+          "pairedOverheadPercent": -9.8598,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 9,
+          "directTokens": 93065,
+          "candidateTokens": 123620,
+          "absoluteDeltaTokens": 30555,
+          "pairedOverheadPercent": 32.8319,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 11,
+          "directTokens": 154305,
+          "candidateTokens": 180814,
+          "absoluteDeltaTokens": 26509,
+          "pairedOverheadPercent": 17.1796,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 12,
+          "directTokens": 72198,
+          "candidateTokens": 133427,
+          "absoluteDeltaTokens": 61229,
+          "pairedOverheadPercent": 84.8071,
+          "baselineRegime": "LOW_BASELINE_TOKEN_REGIME"
+        },
+        {
+          "runIndex": 13,
+          "directTokens": 146715,
+          "candidateTokens": 101102,
+          "absoluteDeltaTokens": -45613,
+          "pairedOverheadPercent": -31.0895,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 14,
+          "directTokens": 177643,
+          "candidateTokens": 215837,
+          "absoluteDeltaTokens": 38194,
+          "pairedOverheadPercent": 21.5004,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 15,
+          "directTokens": 220766,
+          "candidateTokens": 158285,
+          "absoluteDeltaTokens": -62481,
+          "pairedOverheadPercent": -28.3019,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 17,
+          "directTokens": 169051,
+          "candidateTokens": 154299,
+          "absoluteDeltaTokens": -14752,
+          "pairedOverheadPercent": -8.7264,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 18,
+          "directTokens": 153320,
+          "candidateTokens": 158107,
+          "absoluteDeltaTokens": 4787,
+          "pairedOverheadPercent": 3.1222,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 19,
+          "directTokens": 178973,
+          "candidateTokens": 166334,
+          "absoluteDeltaTokens": -12639,
+          "pairedOverheadPercent": -7.062,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 20,
+          "directTokens": 174387,
+          "candidateTokens": 230539,
+          "absoluteDeltaTokens": 56152,
+          "pairedOverheadPercent": 32.1996,
+          "baselineRegime": "NORMAL"
+        }
+      ],
+      "claimStatus": "OBSERVATIONAL",
+      "claimAllowed": true,
+      "reason": "Trusted usage, actual timing, verification, and matching comparability metadata are present.",
+      "tail": {
+        "sampleMinimum": 20,
+        "sampleCount": 18,
+        "p95TokenOverheadPercent": 90.0598,
+        "pairedOverheadP95Percent": 90.0598,
+        "distributionP95DeltaPercent": 17.7068,
+        "outlierCount": 2,
+        "status": "NOT_ENOUGH_SAMPLES",
+        "pairedStatus": "NOT_ENOUGH_SAMPLES",
+        "distributionStatus": "NOT_ENOUGH_SAMPLES",
+        "combinedInterpretation": "TAIL_UNRESOLVED"
+      }
+    },
+    "forgeloopAdaptive": {
+      "comparablePairs": 18,
+      "tokenComparablePairs": 18,
+      "timeComparablePairs": 18,
+      "tokenOverheadPercent": {
+        "p50": 0.6198,
+        "p95": 83.6626
+      },
+      "pairedOverheadPercent": {
+        "p50": 0.6198,
+        "p95": 83.6626
+      },
+      "distributionDeltaPercent": {
+        "p50": 1.6388,
+        "p95": 8.7185
+      },
+      "timeOverheadPercent": {
+        "p50": 4.9779,
+        "p95": 18.859
+      },
+      "pairedRatioDiagnostics": {
+        "pairCount": 18,
+        "baselineMinimum": 72198,
+        "baselineP25": 148186,
+        "baselineP50": 154560.5,
+        "lowBaselineThreshold": 92736.3,
+        "lowBaselinePairCount": 2
+      },
+      "pairedRuns": [
+        {
+          "runIndex": 1,
+          "directTokens": 175336,
+          "candidateTokens": 156914,
+          "absoluteDeltaTokens": -18422,
+          "pairedOverheadPercent": -10.5067,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 2,
+          "directTokens": 152599,
+          "candidateTokens": 190144,
+          "absoluteDeltaTokens": 37545,
+          "pairedOverheadPercent": 24.6037,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 3,
+          "directTokens": 177864,
+          "candidateTokens": 157273,
+          "absoluteDeltaTokens": -20591,
+          "pairedOverheadPercent": -11.5768,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 4,
+          "directTokens": 80306,
+          "candidateTokens": 150522,
+          "absoluteDeltaTokens": 70216,
+          "pairedOverheadPercent": 87.4356,
+          "baselineRegime": "LOW_BASELINE_TOKEN_REGIME"
+        },
+        {
+          "runIndex": 5,
+          "directTokens": 153385,
+          "candidateTokens": 203604,
+          "absoluteDeltaTokens": 50219,
+          "pairedOverheadPercent": 32.7405,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 6,
+          "directTokens": 154816,
+          "candidateTokens": 153525,
+          "absoluteDeltaTokens": -1291,
+          "pairedOverheadPercent": -0.8339,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 7,
+          "directTokens": 152635,
+          "candidateTokens": 68057,
+          "absoluteDeltaTokens": -84578,
+          "pairedOverheadPercent": -55.4119,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 8,
+          "directTokens": 169212,
+          "candidateTokens": 157625,
+          "absoluteDeltaTokens": -11587,
+          "pairedOverheadPercent": -6.8476,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 9,
+          "directTokens": 93065,
+          "candidateTokens": 170306,
+          "absoluteDeltaTokens": 77241,
+          "pairedOverheadPercent": 82.9968,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 10,
+          "directTokens": 144797,
+          "candidateTokens": 73892,
+          "absoluteDeltaTokens": -70905,
+          "pairedOverheadPercent": -48.9686,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 11,
+          "directTokens": 154305,
+          "candidateTokens": 199191,
+          "absoluteDeltaTokens": 44886,
+          "pairedOverheadPercent": 29.0891,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 12,
+          "directTokens": 72198,
+          "candidateTokens": 73695,
+          "absoluteDeltaTokens": 1497,
+          "pairedOverheadPercent": 2.0735,
+          "baselineRegime": "LOW_BASELINE_TOKEN_REGIME"
+        },
+        {
+          "runIndex": 13,
+          "directTokens": 146715,
+          "candidateTokens": 154966,
+          "absoluteDeltaTokens": 8251,
+          "pairedOverheadPercent": 5.6238,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 14,
+          "directTokens": 177643,
+          "candidateTokens": 149413,
+          "absoluteDeltaTokens": -28230,
+          "pairedOverheadPercent": -15.8914,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 15,
+          "directTokens": 220766,
+          "candidateTokens": 177210,
+          "absoluteDeltaTokens": -43556,
+          "pairedOverheadPercent": -19.7295,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 17,
+          "directTokens": 169051,
+          "candidateTokens": 195147,
+          "absoluteDeltaTokens": 26096,
+          "pairedOverheadPercent": 15.4368,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 19,
+          "directTokens": 178973,
+          "candidateTokens": 72492,
+          "absoluteDeltaTokens": -106481,
+          "pairedOverheadPercent": -59.4956,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 20,
+          "directTokens": 174387,
+          "candidateTokens": 201002,
+          "absoluteDeltaTokens": 26615,
+          "pairedOverheadPercent": 15.262,
+          "baselineRegime": "NORMAL"
+        }
+      ],
+      "claimStatus": "OBSERVATIONAL",
+      "claimAllowed": true,
+      "reason": "Trusted usage, actual timing, verification, and matching comparability metadata are present.",
+      "tail": {
+        "sampleMinimum": 20,
+        "sampleCount": 18,
+        "p95TokenOverheadPercent": 83.6626,
+        "pairedOverheadP95Percent": 83.6626,
+        "distributionP95DeltaPercent": 8.7185,
+        "outlierCount": 0,
+        "status": "NOT_ENOUGH_SAMPLES",
+        "pairedStatus": "NOT_ENOUGH_SAMPLES",
+        "distributionStatus": "NOT_ENOUGH_SAMPLES",
+        "combinedInterpretation": "TAIL_UNRESOLVED"
+      }
+    }
+  },
+  "lightObjectives": {
+    "p50TokenOverheadPercent": 35,
+    "p95TokenOverheadPercent": 60,
+    "status": "OBSERVATIONAL",
+    "p50Pass": true,
+    "p95Pass": false
+  },
+  "sourcePolicy": "PROVIDER_REPORTED_OR_HOST_REPORTED_ONLY",
+  "claimsAllowed": true,
+  "generatedFromRunCount": 60,
+  "contextInflation": {
+    "status": "NOT_COMPARABLE",
+    "comparablePairs": 0,
+    "inflatedPairs": 0,
+    "lightContextTokens": {
+      "count": 0,
+      "average": null,
+      "p50": null,
+      "p95": null,
+      "minimum": null,
+      "maximum": null
+    },
+    "balancedContextTokens": {
+      "count": 0,
+      "average": null,
+      "p50": null,
+      "p95": null,
+      "minimum": null,
+      "maximum": null
+    },
+    "blocking": false,
+    "reason": "Context inflation requires complete HOST_REPORTED context items for matching LIGHT and BALANCED runs."
+  },
+  "outlierAnalysis": {
+    "policy": "TOKEN_IQR_1_5",
+    "minimumSamples": 4,
+    "modes": {
+      "direct": {
+        "sampleCount": 20,
+        "measuredCount": 20,
+        "status": "MEASURED",
+        "q1": 150716.25,
+        "q3": 174624.25,
+        "iqr": 23908,
+        "median": 153845,
+        "upperFence": 210486.25,
+        "outliers": [
+          {
+            "runId": "run-novatask-saas-landing-page-direct-015",
+            "runIndex": 15,
+            "totalTokens": 220766,
+            "scenarioMedianTokens": 153845,
+            "ratioToMedian": 1.435,
+            "reasons": [
+              "TOKEN_IQR_OUTLIER"
+            ],
+            "diagnosticSignals": [
+              "UNKNOWN_TOKEN_SPIKE"
+            ]
+          }
+        ]
+      },
+      "forgeloopBalanced": {
+        "sampleCount": 20,
+        "measuredCount": 20,
+        "status": "MEASURED",
+        "q1": 152393,
+        "q3": 173189.5,
+        "iqr": 20796.5,
+        "median": 158196,
+        "upperFence": 204384.25,
+        "outliers": [
+          {
+            "runId": "run-novatask-saas-landing-page-forgeloopBalanced-014",
+            "runIndex": 14,
+            "totalTokens": 215837,
+            "scenarioMedianTokens": 158196,
+            "ratioToMedian": 1.3644,
+            "reasons": [
+              "TOKEN_IQR_OUTLIER"
+            ],
+            "diagnosticSignals": [
+              "UNKNOWN_TOKEN_SPIKE"
+            ]
+          },
+          {
+            "runId": "run-novatask-saas-landing-page-forgeloopBalanced-020",
+            "runIndex": 20,
+            "totalTokens": 230539,
+            "scenarioMedianTokens": 158196,
+            "ratioToMedian": 1.4573,
+            "reasons": [
+              "TOKEN_IQR_OUTLIER"
+            ],
+            "diagnosticSignals": [
+              "UNKNOWN_TOKEN_SPIKE"
+            ]
+          }
+        ]
+      },
+      "forgeloopAdaptive": {
+        "sampleCount": 20,
+        "measuredCount": 20,
+        "status": "MEASURED",
+        "q1": 133976,
+        "q3": 180443.5,
+        "iqr": 46467.5,
+        "median": 157093.5,
+        "upperFence": 250144.75,
+        "outliers": []
+      }
+    }
+  }
+}