@cassiomc1/forgeloop 1.7.0 → 1.8.0

package/benchmarks/execution-profiles/results/aggregate/codex-tail-repeat20-20260831/authentication-change.json

@@ -0,0 +1,1619 @@
+{
+  "schemaVersion": 1,
+  "benchmarkVersion": "2",
+  "runSetId": "codex-tail-repeat20-20260831",
+  "scenarioId": "authentication-change",
+  "expectedProfile": "full",
+  "modeAggregates": {
+    "direct": {
+      "runCount": 20,
+      "verificationSuccessRate": 1,
+      "usageSources": [
+        "HOST_REPORTED"
+      ],
+      "totalTokens": {
+        "count": 20,
+        "average": 61573.85,
+        "minimum": 55563,
+        "p25": 55985.5,
+        "p50": 56205.5,
+        "p75": 69780,
+        "p90": 69908.6,
+        "p95": 70431.45,
+        "maximum": 70858,
+        "iqr": 13794.5,
+        "mad": 590,
+        "outlierCount": 0
+      },
+      "wallClockMs": {
+        "count": 20,
+        "average": 21397.2941,
+        "minimum": 14962.5706,
+        "p25": 18822.5008,
+        "p50": 20792.2665,
+        "p75": 23494.056,
+        "p90": 24373.0749,
+        "p95": 25191.1884,
+        "maximum": 33747.3273,
+        "iqr": 4671.5551,
+        "mad": 2283.9702,
+        "outlierCount": 1
+      },
+      "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": 5,
+        "minimum": 5,
+        "p25": 5,
+        "p50": 5,
+        "p75": 5,
+        "p90": 5,
+        "p95": 5,
+        "maximum": 5,
+        "iqr": 0,
+        "mad": 0,
+        "outlierCount": 0
+      },
+      "tokensPerComparableStep": {
+        "count": 20,
+        "average": 12314.77,
+        "minimum": 11112.6,
+        "p25": 11197.1,
+        "p50": 11241.1,
+        "p75": 13956,
+        "p90": 13981.72,
+        "p95": 14086.29,
+        "maximum": 14171.6,
+        "iqr": 2758.9,
+        "mad": 118,
+        "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": 4.05,
+          "minimum": 3,
+          "p25": 4,
+          "p50": 4,
+          "p75": 4,
+          "p90": 5,
+          "p95": 5,
+          "maximum": 5,
+          "iqr": 0,
+          "mad": 0,
+          "outlierCount": 3
+        },
+        "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": 1,
+          "minimum": 1,
+          "p25": 1,
+          "p50": 1,
+          "p75": 1,
+          "p90": 1,
+          "p95": 1,
+          "maximum": 1,
+          "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": [
+          "UNKNOWN"
+        ],
+        "scores": {
+          "visualQuality": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "responsiveQuality": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "accessibility": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "interactionPolish": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "requirementsCompleteness": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          }
+        }
+      }
+    },
+    "forgeloopBalanced": {
+      "runCount": 20,
+      "verificationSuccessRate": 1,
+      "usageSources": [
+        "HOST_REPORTED"
+      ],
+      "totalTokens": {
+        "count": 20,
+        "average": 64157.5,
+        "minimum": 57272,
+        "p25": 57666,
+        "p50": 58088.5,
+        "p75": 71947.75,
+        "p90": 72173.4,
+        "p95": 72300.5,
+        "maximum": 73108,
+        "iqr": 14281.75,
+        "mad": 765.5,
+        "outlierCount": 0
+      },
+      "wallClockMs": {
+        "count": 20,
+        "average": 21290.9673,
+        "minimum": 17932.7938,
+        "p25": 19203.7865,
+        "p50": 21205.5051,
+        "p75": 22989.8312,
+        "p90": 23988.4525,
+        "p95": 25181.4752,
+        "maximum": 26322.821,
+        "iqr": 3786.0447,
+        "mad": 2014.7612,
+        "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": 5,
+        "minimum": 5,
+        "p25": 5,
+        "p50": 5,
+        "p75": 5,
+        "p90": 5,
+        "p95": 5,
+        "maximum": 5,
+        "iqr": 0,
+        "mad": 0,
+        "outlierCount": 0
+      },
+      "tokensPerComparableStep": {
+        "count": 20,
+        "average": 12831.5,
+        "minimum": 11454.4,
+        "p25": 11533.2,
+        "p50": 11617.7,
+        "p75": 14389.55,
+        "p90": 14434.68,
+        "p95": 14460.1,
+        "maximum": 14621.6,
+        "iqr": 2856.35,
+        "mad": 153.1,
+        "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": 3.85,
+          "minimum": 3,
+          "p25": 4,
+          "p50": 4,
+          "p75": 4,
+          "p90": 4,
+          "p95": 4.05,
+          "maximum": 5,
+          "iqr": 0,
+          "mad": 0,
+          "outlierCount": 1
+        },
+        "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": 1,
+          "minimum": 1,
+          "p25": 1,
+          "p50": 1,
+          "p75": 1,
+          "p90": 1,
+          "p95": 1,
+          "maximum": 1,
+          "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": [
+          "full"
+        ],
+        "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": [
+          "UNKNOWN"
+        ],
+        "scores": {
+          "visualQuality": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "responsiveQuality": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "accessibility": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "interactionPolish": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "requirementsCompleteness": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          }
+        }
+      }
+    },
+    "forgeloopAdaptive": {
+      "runCount": 20,
+      "verificationSuccessRate": 1,
+      "usageSources": [
+        "HOST_REPORTED"
+      ],
+      "totalTokens": {
+        "count": 20,
+        "average": 62722.7,
+        "minimum": 56828,
+        "p25": 57368.75,
+        "p50": 57585.5,
+        "p75": 71479.75,
+        "p90": 72466.4,
+        "p95": 73952.8,
+        "maximum": 89814,
+        "iqr": 14111,
+        "mad": 259.5,
+        "outlierCount": 0
+      },
+      "wallClockMs": {
+        "count": 20,
+        "average": 23157.1048,
+        "minimum": 18186.0438,
+        "p25": 20962.052,
+        "p50": 21827.999,
+        "p75": 23087.6345,
+        "p90": 28743.3015,
+        "p95": 29895.9979,
+        "maximum": 40222.0097,
+        "iqr": 2125.5825,
+        "mad": 1229.5444,
+        "outlierCount": 3
+      },
+      "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": 5,
+        "minimum": 5,
+        "p25": 5,
+        "p50": 5,
+        "p75": 5,
+        "p90": 5,
+        "p95": 5,
+        "maximum": 5,
+        "iqr": 0,
+        "mad": 0,
+        "outlierCount": 0
+      },
+      "tokensPerComparableStep": {
+        "count": 20,
+        "average": 12544.54,
+        "minimum": 11365.6,
+        "p25": 11473.75,
+        "p50": 11517.1,
+        "p75": 14295.95,
+        "p90": 14493.28,
+        "p95": 14790.56,
+        "maximum": 17962.8,
+        "iqr": 2822.2,
+        "mad": 51.9,
+        "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": 4.3,
+          "minimum": 3,
+          "p25": 4,
+          "p50": 4,
+          "p75": 4,
+          "p90": 5,
+          "p95": 5.15,
+          "maximum": 8,
+          "iqr": 0,
+          "mad": 0,
+          "outlierCount": 4
+        },
+        "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": 1,
+          "minimum": 1,
+          "p25": 1,
+          "p50": 1,
+          "p75": 1,
+          "p90": 1,
+          "p95": 1,
+          "maximum": 1,
+          "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": [
+          "full"
+        ],
+        "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": [
+          "UNKNOWN"
+        ],
+        "scores": {
+          "visualQuality": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "responsiveQuality": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "accessibility": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "interactionPolish": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          },
+          "requirementsCompleteness": {
+            "count": 0,
+            "average": null,
+            "minimum": null,
+            "p25": null,
+            "p50": null,
+            "p75": null,
+            "p90": null,
+            "p95": null,
+            "maximum": null,
+            "iqr": null,
+            "mad": null,
+            "outlierCount": 0
+          }
+        }
+      }
+    }
+  },
+  "comparisons": {
+    "direct": null,
+    "forgeloopBalanced": {
+      "comparablePairs": 20,
+      "tokenComparablePairs": 20,
+      "timeComparablePairs": 20,
+      "tokenOverheadPercent": {
+        "p50": 3.4338,
+        "p95": 28.1138
+      },
+      "pairedOverheadPercent": {
+        "p50": 3.4338,
+        "p95": 28.1138
+      },
+      "distributionDeltaPercent": {
+        "p50": 3.3502,
+        "p95": 2.6537
+      },
+      "timeOverheadPercent": {
+        "p50": 0.7717,
+        "p95": 26.7166
+      },
+      "pairedRatioDiagnostics": {
+        "pairCount": 20,
+        "baselineMinimum": 55563,
+        "baselineP25": 55985.5,
+        "baselineP50": 56205.5,
+        "lowBaselineThreshold": 33723.3,
+        "lowBaselinePairCount": 0
+      },
+      "pairedRuns": [
+        {
+          "runIndex": 1,
+          "directTokens": 55563,
+          "candidateTokens": 57423,
+          "absoluteDeltaTokens": 1860,
+          "pairedOverheadPercent": 3.3476,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 2,
+          "directTokens": 56183,
+          "candidateTokens": 72058,
+          "absoluteDeltaTokens": 15875,
+          "pairedOverheadPercent": 28.2559,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 3,
+          "directTokens": 55615,
+          "candidateTokens": 57839,
+          "absoluteDeltaTokens": 2224,
+          "pairedOverheadPercent": 3.9989,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 4,
+          "directTokens": 69851,
+          "candidateTokens": 71694,
+          "absoluteDeltaTokens": 1843,
+          "pairedOverheadPercent": 2.6385,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 5,
+          "directTokens": 56066,
+          "candidateTokens": 57924,
+          "absoluteDeltaTokens": 1858,
+          "pairedOverheadPercent": 3.314,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 6,
+          "directTokens": 56113,
+          "candidateTokens": 57476,
+          "absoluteDeltaTokens": 1363,
+          "pairedOverheadPercent": 2.429,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 7,
+          "directTokens": 69670,
+          "candidateTokens": 72164,
+          "absoluteDeltaTokens": 2494,
+          "pairedOverheadPercent": 3.5797,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 8,
+          "directTokens": 70409,
+          "candidateTokens": 57573,
+          "absoluteDeltaTokens": -12836,
+          "pairedOverheadPercent": -18.2306,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 9,
+          "directTokens": 69853,
+          "candidateTokens": 57697,
+          "absoluteDeltaTokens": -12156,
+          "pairedOverheadPercent": -17.4023,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 10,
+          "directTokens": 56068,
+          "candidateTokens": 71825,
+          "absoluteDeltaTokens": 15757,
+          "pairedOverheadPercent": 28.1034,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 11,
+          "directTokens": 55616,
+          "candidateTokens": 58253,
+          "absoluteDeltaTokens": 2637,
+          "pairedOverheadPercent": 4.7414,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 12,
+          "directTokens": 69773,
+          "candidateTokens": 73108,
+          "absoluteDeltaTokens": 3335,
+          "pairedOverheadPercent": 4.7798,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 13,
+          "directTokens": 56228,
+          "candidateTokens": 57724,
+          "absoluteDeltaTokens": 1496,
+          "pairedOverheadPercent": 2.6606,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 14,
+          "directTokens": 69620,
+          "candidateTokens": 57272,
+          "absoluteDeltaTokens": -12348,
+          "pairedOverheadPercent": -17.7363,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 15,
+          "directTokens": 55619,
+          "candidateTokens": 57832,
+          "absoluteDeltaTokens": 2213,
+          "pairedOverheadPercent": 3.9789,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 16,
+          "directTokens": 55948,
+          "candidateTokens": 57374,
+          "absoluteDeltaTokens": 1426,
+          "pairedOverheadPercent": 2.5488,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 17,
+          "directTokens": 69801,
+          "candidateTokens": 72258,
+          "absoluteDeltaTokens": 2457,
+          "pairedOverheadPercent": 3.52,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 18,
+          "directTokens": 70858,
+          "candidateTokens": 71936,
+          "absoluteDeltaTokens": 1078,
+          "pairedOverheadPercent": 1.5214,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 19,
+          "directTokens": 56625,
+          "candidateTokens": 71983,
+          "absoluteDeltaTokens": 15358,
+          "pairedOverheadPercent": 27.1223,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 20,
+          "directTokens": 55998,
+          "candidateTokens": 71737,
+          "absoluteDeltaTokens": 15739,
+          "pairedOverheadPercent": 28.1064,
+          "baselineRegime": "NORMAL"
+        }
+      ],
+      "claimStatus": "OBSERVATIONAL",
+      "claimAllowed": true,
+      "reason": "Trusted usage, actual timing, verification, and matching comparability metadata are present.",
+      "tail": {
+        "sampleMinimum": 20,
+        "sampleCount": 20,
+        "p95TokenOverheadPercent": 28.1138,
+        "pairedOverheadP95Percent": 28.1138,
+        "distributionP95DeltaPercent": 2.6537,
+        "outlierCount": 0,
+        "status": "TAIL_STABLE",
+        "pairedStatus": "TAIL_STABLE",
+        "distributionStatus": "TAIL_ACCEPTABLE",
+        "combinedInterpretation": "TAIL_CONSISTENT"
+      }
+    },
+    "forgeloopAdaptive": {
+      "comparablePairs": 20,
+      "tokenComparablePairs": 20,
+      "timeComparablePairs": 20,
+      "tokenOverheadPercent": {
+        "p50": 2.5493,
+        "p95": 29.6296
+      },
+      "pairedOverheadPercent": {
+        "p50": 2.5493,
+        "p95": 29.6296
+      },
+      "distributionDeltaPercent": {
+        "p50": 2.4553,
+        "p95": 4.9997
+      },
+      "timeOverheadPercent": {
+        "p50": 5.0181,
+        "p95": 68.479
+      },
+      "pairedRatioDiagnostics": {
+        "pairCount": 20,
+        "baselineMinimum": 55563,
+        "baselineP25": 55985.5,
+        "baselineP50": 56205.5,
+        "lowBaselineThreshold": 33723.3,
+        "lowBaselinePairCount": 0
+      },
+      "pairedRuns": [
+        {
+          "runIndex": 1,
+          "directTokens": 55563,
+          "candidateTokens": 57356,
+          "absoluteDeltaTokens": 1793,
+          "pairedOverheadPercent": 3.227,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 2,
+          "directTokens": 56183,
+          "candidateTokens": 57300,
+          "absoluteDeltaTokens": 1117,
+          "pairedOverheadPercent": 1.9881,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 3,
+          "directTokens": 55615,
+          "candidateTokens": 57816,
+          "absoluteDeltaTokens": 2201,
+          "pairedOverheadPercent": 3.9576,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 4,
+          "directTokens": 69851,
+          "candidateTokens": 57388,
+          "absoluteDeltaTokens": -12463,
+          "pairedOverheadPercent": -17.8423,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 5,
+          "directTokens": 56066,
+          "candidateTokens": 57344,
+          "absoluteDeltaTokens": 1278,
+          "pairedOverheadPercent": 2.2795,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 6,
+          "directTokens": 56113,
+          "candidateTokens": 56828,
+          "absoluteDeltaTokens": 715,
+          "pairedOverheadPercent": 1.2742,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 7,
+          "directTokens": 69670,
+          "candidateTokens": 57511,
+          "absoluteDeltaTokens": -12159,
+          "pairedOverheadPercent": -17.4523,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 8,
+          "directTokens": 70409,
+          "candidateTokens": 72394,
+          "absoluteDeltaTokens": 1985,
+          "pairedOverheadPercent": 2.8192,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 9,
+          "directTokens": 69853,
+          "candidateTokens": 57641,
+          "absoluteDeltaTokens": -12212,
+          "pairedOverheadPercent": -17.4824,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 10,
+          "directTokens": 56068,
+          "candidateTokens": 71438,
+          "absoluteDeltaTokens": 15370,
+          "pairedOverheadPercent": 27.4131,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 11,
+          "directTokens": 55616,
+          "candidateTokens": 89814,
+          "absoluteDeltaTokens": 34198,
+          "pairedOverheadPercent": 61.4895,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 12,
+          "directTokens": 69773,
+          "candidateTokens": 57373,
+          "absoluteDeltaTokens": -12400,
+          "pairedOverheadPercent": -17.7719,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 13,
+          "directTokens": 56228,
+          "candidateTokens": 57894,
+          "absoluteDeltaTokens": 1666,
+          "pairedOverheadPercent": 2.9629,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 14,
+          "directTokens": 69620,
+          "candidateTokens": 57504,
+          "absoluteDeltaTokens": -12116,
+          "pairedOverheadPercent": -17.403,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 15,
+          "directTokens": 55619,
+          "candidateTokens": 57308,
+          "absoluteDeltaTokens": 1689,
+          "pairedOverheadPercent": 3.0367,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 16,
+          "directTokens": 55948,
+          "candidateTokens": 57544,
+          "absoluteDeltaTokens": 1596,
+          "pairedOverheadPercent": 2.8526,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 17,
+          "directTokens": 69801,
+          "candidateTokens": 73118,
+          "absoluteDeltaTokens": 3317,
+          "pairedOverheadPercent": 4.7521,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 18,
+          "directTokens": 70858,
+          "candidateTokens": 71605,
+          "absoluteDeltaTokens": 747,
+          "pairedOverheadPercent": 1.0542,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 19,
+          "directTokens": 56625,
+          "candidateTokens": 57627,
+          "absoluteDeltaTokens": 1002,
+          "pairedOverheadPercent": 1.7695,
+          "baselineRegime": "NORMAL"
+        },
+        {
+          "runIndex": 20,
+          "directTokens": 55998,
+          "candidateTokens": 71651,
+          "absoluteDeltaTokens": 15653,
+          "pairedOverheadPercent": 27.9528,
+          "baselineRegime": "NORMAL"
+        }
+      ],
+      "claimStatus": "OBSERVATIONAL",
+      "claimAllowed": true,
+      "reason": "Trusted usage, actual timing, verification, and matching comparability metadata are present.",
+      "tail": {
+        "sampleMinimum": 20,
+        "sampleCount": 20,
+        "p95TokenOverheadPercent": 29.6296,
+        "pairedOverheadP95Percent": 29.6296,
+        "distributionP95DeltaPercent": 4.9997,
+        "outlierCount": 0,
+        "status": "TAIL_STABLE",
+        "pairedStatus": "TAIL_STABLE",
+        "distributionStatus": "TAIL_ACCEPTABLE",
+        "combinedInterpretation": "TAIL_CONSISTENT"
+      }
+    }
+  },
+  "lightObjectives": null,
+  "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": "The adaptive mode did not resolve to LIGHT for this scenario."
+  },
+  "outlierAnalysis": {
+    "policy": "TOKEN_IQR_1_5",
+    "minimumSamples": 4,
+    "modes": {
+      "direct": {
+        "sampleCount": 20,
+        "measuredCount": 20,
+        "status": "MEASURED",
+        "q1": 55985.5,
+        "q3": 69780,
+        "iqr": 13794.5,
+        "median": 56205.5,
+        "upperFence": 90471.75,
+        "outliers": []
+      },
+      "forgeloopBalanced": {
+        "sampleCount": 20,
+        "measuredCount": 20,
+        "status": "MEASURED",
+        "q1": 57666,
+        "q3": 71947.75,
+        "iqr": 14281.75,
+        "median": 58088.5,
+        "upperFence": 93370.375,
+        "outliers": []
+      },
+      "forgeloopAdaptive": {
+        "sampleCount": 20,
+        "measuredCount": 20,
+        "status": "MEASURED",
+        "q1": 57368.75,
+        "q3": 71479.75,
+        "iqr": 14111,
+        "median": 57585.5,
+        "upperFence": 92646.25,
+        "outliers": []
+      }
+    }
+  }
+}