{
  "apiVersion": "1",
  "version": "2026-09-30-arc-predictive-puzzles",
  "summary": "Composite IQ uses seven equally weighted capability dimensions mapped through IQ 70-160 expected-score ladders. Abstract Reasoning equally averages available source-backed ARC and puzzle projections. Without measured ARC or puzzle results, eligible models use a 50/50 blend of the complete legacy Abstract estimate and a predecessor-plus-trend predictor (lab-trend regression without a suitable predecessor). Explicit same-day sources are eligible only within the same family, from a strictly lower provider tier, with measured ARC1/2 and compatible reasoning mode. The blend is applied after legacy scoring, preserving the other six dimension scores. Predicted ARC coverage remains zero. D2–D7 use Bayesian posterior means from direct projected scores, benchmark offsets, and one uncertain compatible predecessor relationship. Inherited estimates retain zero direct coverage. Conditional 95% ranges are separate and not empirically calibrated; cold-start benchmark prediction remains a known limitation. Unsupported dimensions retain conservative fallbacks. Knowledge Work contributes one seventh of Composite IQ. Its six benchmarks and MirrorCode use provisional site-relative ladders. Chess and Mystery Game use provider-held-out predictive calibrations targeting measured ARC-only IQ, retaining equal D1 benchmark weights; posterior ranges omit ladder and source uncertainty. D7 with no direct observations retains zero direct coverage even when estimated. Emotional Reasoning (EQ) is a separate diagnostic domain excluded from Composite IQ.",
  "derivedRankings": [
    {
      "id": "composite-iq",
      "rankingName": "Composite IQ",
      "direction": "higher_is_better",
      "scoreField": "iq",
      "dimensions": [
        {
          "id": "D1",
          "slug": "abstract-reasoning",
          "name": "Abstract Reasoning",
          "minBenchmarks": 1,
          "benchmarks": [
            {
              "field": "arcAgi3",
              "name": "ARC-AGI-3"
            },
            {
              "field": "arcAgi2",
              "name": "ARC-AGI-2"
            },
            {
              "field": "arcAgi1",
              "name": "ARC-AGI-1"
            },
            {
              "field": "chessPuzzles",
              "name": "Chess Puzzles"
            },
            {
              "field": "mysteryGamePuzzles",
              "name": "Mystery Game Puzzles"
            }
          ]
        },
        {
          "id": "D2",
          "slug": "mathematical-reasoning",
          "name": "Mathematical Reasoning",
          "minBenchmarks": 1,
          "benchmarks": [
            {
              "field": "fmT4Acc",
              "name": "FrontierMath Tier 4"
            },
            {
              "field": "fmT13Acc",
              "name": "FrontierMath Tier 1-3"
            },
            {
              "field": "proofbench11",
              "name": "ProofBench 1.1"
            },
            {
              "field": "mathArena",
              "name": "MathArena"
            },
            {
              "field": "aime",
              "name": "AIME"
            }
          ]
        },
        {
          "id": "D3",
          "slug": "academic-reasoning",
          "name": "Academic Reasoning",
          "minBenchmarks": 1,
          "benchmarks": [
            {
              "field": "hle",
              "name": "Humanity's Last Exam"
            },
            {
              "field": "gpqa",
              "name": "GPQA Diamond"
            },
            {
              "field": "critPt",
              "name": "CritPt"
            },
            {
              "field": "sciCode",
              "name": "SciCode"
            },
            {
              "field": "mmlupro",
              "name": "MMLU-Pro"
            },
            {
              "field": "mmmuPro",
              "name": "MMMU-Pro"
            }
          ]
        },
        {
          "id": "D4",
          "slug": "programmatic-reasoning",
          "name": "Programmatic Reasoning",
          "minBenchmarks": 1,
          "benchmarks": [
            {
              "field": "livecodebench",
              "name": "LiveCodeBench"
            },
            {
              "field": "ioiV2",
              "name": "IOI V2"
            },
            {
              "field": "terminalbench40",
              "name": "Terminal-Bench 4.0"
            },
            {
              "field": "programBenchAlmostResolved",
              "name": "ProgramBench"
            },
            {
              "field": "sweRebench",
              "name": "SWE-rebench"
            },
            {
              "field": "mirrorCode",
              "name": "MirrorCode"
            }
          ]
        },
        {
          "id": "D5",
          "slug": "computer-use",
          "name": "Computer Use",
          "minBenchmarks": 1,
          "benchmarks": [
            {
              "field": "browseComp",
              "name": "BrowseComp"
            },
            {
              "field": "osworldVerified",
              "name": "OSWorld-Verified"
            },
            {
              "field": "mcpAtlas",
              "name": "MCP Atlas"
            },
            {
              "field": "agentArena",
              "name": "Arena.ai Agent Arena"
            },
            {
              "field": "agentsLastExam",
              "name": "Agents' Last Exam"
            }
          ]
        },
        {
          "id": "D6",
          "slug": "reliability",
          "name": "Reliability",
          "minBenchmarks": 1,
          "benchmarks": [
            {
              "field": "simpleQAVerified",
              "name": "SimpleQA Verified"
            },
            {
              "field": "aaOmniscience",
              "name": "aaOmniscience"
            },
            {
              "field": "bullshitBench",
              "name": "BullshitBench v2"
            },
            {
              "field": "ifBench",
              "name": "IFBench"
            },
            {
              "field": "aaLCR11",
              "name": "AA Long Context Reasoning v1.1"
            },
            {
              "field": "factsGrounding",
              "name": "FACTS Grounding"
            }
          ]
        },
        {
          "id": "D7",
          "slug": "knowledge-work",
          "name": "Knowledge Work",
          "minBenchmarks": 1,
          "benchmarks": [
            {
              "field": "aaBriefcase",
              "name": "AA-Briefcase v1.1"
            },
            {
              "field": "gdpvalAAV2",
              "name": "GDPval-AA v2.1"
            },
            {
              "field": "automationBenchAA",
              "name": "AutomationBench-AA"
            },
            {
              "field": "gdpPdf",
              "name": "GDP.pdf"
            },
            {
              "field": "aaAnalystAgent",
              "name": "AA-AnalystAgent"
            },
            {
              "field": "enterpriseOpsGymAA",
              "name": "EnterpriseOps-Gym-AA"
            }
          ]
        }
      ]
    },
    {
      "id": "effective-cost",
      "rankingName": "Effective Cost",
      "direction": "lower_is_better",
      "scoreField": "effectiveCost",
      "unit": "USD per 1M I/O Tokens",
      "breakdown": [
        {
          "id": "published-pricing",
          "name": "Published Token Pricing",
          "inputs": [
            {
              "field": "inP",
              "name": "Input token price"
            },
            {
              "field": "outP",
              "name": "Output token price"
            },
            {
              "field": "cacheReadP",
              "name": "Cache read token price"
            },
            {
              "field": "cacheWriteP",
              "name": "Cache write token price"
            },
            {
              "field": "cacheWrite1hP",
              "name": "1-hour cache write token price"
            },
            {
              "field": "cacheStorageHourlyP",
              "name": "Cache storage hourly token price"
            }
          ],
          "summary": "Base effective cost still uses input price plus output price for 1M input tokens and 1M output tokens. Cache read/write/storage fields are exposed separately for cache-aware task-cost views and provider-specific cost decomposition."
        },
        {
          "id": "observed-token-usage",
          "name": "Observed Token Usage",
          "inputs": [
            {
              "field": "aaOutputTokensPerTask",
              "name": "AA v4.3.2 weighted output tokens per task"
            }
          ],
          "summary": "AA v4.3.2 weighted output tokens per task calibrate the median output-usage multiplier after paired configuration and cost validation. Historical suite totals are excluded; this is not input-token efficiency."
        },
        {
          "id": "task-cost-residuals",
          "name": "Task Cost Residuals",
          "benchmarks": [
            {
              "field": "arcCostPerTask",
              "name": "ARC-AGI-2 cost per task"
            },
            {
              "field": "valsCostPerTestV2",
              "name": "VALS cost per test"
            },
            {
              "field": "swebenchCostPerTask",
              "name": "SWE-Bench cost per task"
            },
            {
              "field": "hleCostPerQuestion",
              "name": "Humanity's Last Exam cost per question"
            }
          ],
          "summary": "Task-level cost benchmarks adjust for observed cost differences not explained by published token prices alone."
        },
        {
          "id": "usage-multiplier-waterfall",
          "name": "Usage Multiplier Waterfall",
          "steps": [
            "measured benchmark multiplier from validated token usage and price-adjusted task-cost residuals",
            "one-generation-back same-family same-lineage multiplier",
            "two-generations-back same-family same-lineage multiplier",
            "geometric average of the three closest measured peers",
            "assumed 1x fallback for positive-price models"
          ],
          "summary": "Effective cost is sticker price multiplied by the first available usage multiplier in this waterfall."
        }
      ]
    }
  ],
  "updatedAt": "2026-10-13T12:11:32.230Z",
  "url": "https://www.aiiq.org/methodology/"
}