{
  "schema_version": "1.0",
  "research_date": "2026-10-22",
  "scope": "Public unit-price and inference-work evidence for Knauf Digital presentation; no claims about Knauf measured spending",
  "sources_file": "2026-10-22-source-manifest.json",
  "claims": [
    {
      "id": "UP01",
      "claim": "A non-reasoning model frontier meeting MMLU >= 64.8 fell from USD 20 to USD 0.07 per million blended API tokens from November 2022 to October 2024.",
      "verdict": "supported_under_checked_conditions",
      "priority": "main_chart",
      "source_ids": [
        "UP-S01",
        "UP-S02"
      ],
      "metric": "USD per 1M API tokens weighted 3:1 input/output",
      "threshold": {
        "benchmark": "MMLU",
        "minimum_percent": 64.8
      },
      "endpoints": [
        {
          "date": "2022-11-30",
          "model": "GPT-3.5",
          "usd_per_million_tokens": 20,
          "benchmark_score_percent": 64.8
        },
        {
          "date": "2024-10-03",
          "model": "Gemini-1.5-Flash-8B",
          "usd_per_million_tokens": 0.07,
          "benchmark_score_percent": 75
        }
      ],
      "derived": {
        "ratio_from_rounded_values": 285.71428571428567,
        "display_ratio": "more than 280x",
        "elapsed_days": 673,
        "method": "rounded endpoint division and Gregorian calendar subtraction"
      },
      "caveats": [
        "Changing-model frontier, not same-model price cut",
        "Benchmark threshold is not general capability equivalence",
        "Not current October 2026 prices",
        "Not task cost; no enterprise review/tools/infrastructure included",
        "Reasoning models excluded from upstream analysis"
      ],
      "chart_data": "2026-10-22-fixed-threshold-price-chart.csv"
    },
    {
      "id": "UP02",
      "claim": "The per-token price study excluded reasoning models because their longer generated traces can make token-price comparisons misleading for total evaluation cost.",
      "verdict": "supported_under_checked_conditions",
      "priority": "spoken_bridge",
      "source_ids": [
        "UP-S02"
      ],
      "locator": "Data methodology paragraph on reasoning models",
      "interpretation": "Measure tokens and complete workflow consumption per accepted result rather than inferring task savings from a token-price frontier."
    },
    {
      "id": "UP03",
      "claim": "In OpenAI's historical o1 AIME 2024 experiment, pass@1 was 74.4% and consensus over 64 samples was 83.3%.",
      "verdict": "supported_as_vendor_reported_experiment",
      "priority": "optional_main_chart",
      "source_ids": [
        "UP-S04"
      ],
      "denominator": "AIME 2024 problem accuracy; article expresses mean against 15 questions per exam",
      "configurations": [
        {
          "candidates_per_problem": 1,
          "selection": "single_sample",
          "accuracy_percent": 74.4
        },
        {
          "candidates_per_problem": 64,
          "selection": "consensus",
          "accuracy_percent": 83.3
        }
      ],
      "derived": {
        "accuracy_gain_percentage_points": 8.9,
        "method": "83.3-74.4"
      },
      "caveats": [
        "Historical research o1, not o1-preview/current models",
        "No task-dollar or exact FLOPs ratio follows from candidate count",
        "Per-sample reasoning length and selection overhead are not disclosed",
        "Competition mathematics does not establish product-data workflow benefit",
        "Vendor-reported; no independently reproduced outcome in this audit"
      ],
      "chart_data": "2026-10-22-o1-sampling-chart.csv"
    },
    {
      "id": "UP03-BACKUP",
      "claim": "The same OpenAI article reports rounded 93% AIME accuracy after learned reranking of 1000 candidates.",
      "verdict": "supported_as_vendor_reported_experiment",
      "priority": "backup_only",
      "source_ids": [
        "UP-S04"
      ],
      "caveats": [
        "Main-text rounded value rather than appendix precision",
        "Selection method changes in addition to sample count",
        "Do not label 1000 candidate solutions as 1000 times an API bill"
      ],
      "chart_data": "2026-10-22-o1-sampling-chart.csv"
    },
    {
      "id": "UP04",
      "claim": "Snell et al. report adaptive search/revision comparisons using about one quarter of a best-of-N baseline generation budget at similar or better MATH performance.",
      "verdict": "supported_under_research_accounting_boundary",
      "priority": "technical_backup",
      "source_ids": [
        "UP-S05"
      ],
      "pairs": [
        {
          "experiment": "PRM_search",
          "adaptive_generations": 16,
          "baseline_generations": 64,
          "comparison": "nearly matches or exceeds baseline in reported low-budget comparison"
        },
        {
          "experiment": "revision",
          "adaptive_generations": 64,
          "baseline_generations": 256,
          "comparison": "outperforms baseline in reported comparison"
        }
      ],
      "caveats": [
        "Difficulty-estimation overhead excluded; generating 2048 samples to estimate difficulty can dominate savings",
        "Specific fine-tuned models and 500 MATH test questions",
        "Research generation budgets, not production monetary savings",
        "No exact curve accuracy values digitised"
      ],
      "chart_data": "2026-10-22-adaptive-budget-comparisons.csv"
    },
    {
      "id": "UP05",
      "claim": "The checked official 2026 AI Index does not provide a refreshed fixed-capability token-price series; its inference section concerns energy/carbon per query.",
      "verdict": "supported_as_scoped_report_inspection",
      "priority": "provenance_note",
      "source_ids": [
        "UP-S03"
      ],
      "caveats": [
        "Absence statement applies to this report check only",
        "No denial that other newer price research exists"
      ]
    },
    {
      "id": "UP06",
      "claim": "A falling token price does not by itself imply a falling cost per accepted business task.",
      "verdict": "accounting_inference_not_a_measured_spend_trend",
      "priority": "main_message",
      "source_ids": [
        "UP-S01",
        "UP-S02",
        "UP-S04",
        "UP-S05"
      ],
      "formula": "Workflow cost per accepted result = total workflow cost / accepted results; token component = sum(input_tokens*input_rate + output_and_reasoning_tokens*output_rate) over calls; tool/infrastructure/review costs remain separate.",
      "caveats": [
        "No claim that every bill must rise",
        "No assertion that Knauf currently samples64answers",
        "Fixed workload can directly benefit from falling unit prices",
        "Model-specific billable reasoning and cache treatment must be measured rather than assumed"
      ]
    }
  ],
  "rejected_wordings": [
    {
      "wording": "280x cheaper in 18 months",
      "reason": "Endpoint dates imply 673 days, around 22.1 average months; Stanford's duration prose conflicts."
    },
    {
      "wording": "AI bills must keep rising despite cheaper tokens",
      "reason": "No universal enterprise spend claim established by these sources."
    },
    {
      "wording": "64 samples costs 64 times more",
      "reason": "No per-configuration token/latency/caching/aggregation ledger."
    },
    {
      "wording": "Adaptive reasoning cuts production AI cost 75%",
      "reason": "Research compute proxy excludes difficulty estimation and is workload-specific."
    }
  ]
}