{
  "snapshot_date": "2026-11-01",
  "retrieved_at_utc": "2026-11-01T22:26:28Z",
  "time_basis": "Model announcement_date in LLM Stats; not verified score-publication dates",
  "t_definition": "Days since 2026-01-01",
  "selection": "Strict new maxima, one per announcement date",
  "weighting": "Equal weight per record; no repeated daily plateau observations",
  "manifest": {
    "swe-bench-pro": {
      "source": "https://llm-stats.com/benchmarks/swe-bench-pro",
      "rows": 63,
      "records": 6,
      "csv_sha256": "b45fcd266cc284b15f96360429f8a99d197cad29ffb359fb82656339c1f81915",
      "source_updated_at": "2026-11-01T18:34:07.096365+00:00",
      "verified_count": 0
    },
    "deepswe-1.1": {
      "source": "https://llm-stats.com/benchmarks/deepswe-1.1",
      "rows": 52,
      "records": 8,
      "csv_sha256": "89b21bcec0863307bc1c5ba9030550e22582f5deae4cd2cd4e44b103fd3dc03b",
      "source_updated_at": "2026-11-01T18:33:49.086143+00:00",
      "verified_count": 0
    }
  },
  "pro_linear": {
    "a": 0.5446466924695511,
    "b": 0.0014573330875915795,
    "r2": 0.8639678272228438,
    "rmse_pp": 4.918471004210039
  },
  "deep_shifted_log": {
    "a": 0.23683944364810058,
    "b": 0.09960376215690199,
    "r2": 0.997614402122208,
    "rmse_pp": 0.7635821112144981,
    "tau": 45.11844856333594,
    "sse": 0.00046644611245343217,
    "offset_days": 1.8815514366640602,
    "offset_search_days": [
      0.0001,
      36500
    ],
    "parameters": 3
  },
  "score_levels": [
    0.79,
    0.823
  ],
  "equivalent_days": 101.40724998968477,
  "equivalent_pro_pp": 14.778414073163846,
  "pro_pp_per_month": 4.435757585356869,
  "anchored_pro_pp": 13.674384849223792,
  "deep_fitted_crossings_days": [
    303.2865870570542,
    404.69383704673896
  ],
  "sensitivity": [
    {
      "method": "Full-history shifted log + linear",
      "days": 101.40724998968477,
      "equivalent_pp": 14.778414073163846,
      "explanation": "Primary model; fit the time origin and invert both score levels."
    },
    {
      "method": "Anchor DeepSWE curve to latest record",
      "days": 93.83156785263402,
      "equivalent_pp": 13.674384849223792,
      "explanation": "Same curvature, vertically shifted to pass through 79.0% on October 12."
    },
    {
      "method": "Fix logarithmic origin at January 1",
      "days": 39.05050395678046,
      "equivalent_pp": 5.690959150334206,
      "explanation": "An arbitrary fixed origin changes the fitted curvature; poorer fit (R\u00b2 0.898)."
    },
    {
      "method": "Omit earliest DeepSWE record",
      "days": 114.34171186782737,
      "equivalent_pp": 16.66339599968476,
      "explanation": "Seven records instead of eight."
    },
    {
      "method": "Use April\u2013October DeepSWE linear trend",
      "days": 48.588003056545915,
      "equivalent_pp": 7.080890451430515,
      "explanation": "Recent six DeepSWE records are also close to a line (R\u00b2 0.990)."
    },
    {
      "method": "Use April\u2013September Pro linear trend",
      "days": 101.40724998968477,
      "equivalent_pp": 7.534095492430895,
      "explanation": "Only three Pro frontier records; its slope is lower than the full-history slope."
    }
  ],
  "rolling_holdout": [
    {
      "training_records": 4,
      "held_out_records": 4,
      "shifted_log_rmse_pp": 0.6632025955772229,
      "linear_rmse_pp": 28.51006923309734
    },
    {
      "training_records": 5,
      "held_out_records": 3,
      "shifted_log_rmse_pp": 0.7446032223090002,
      "linear_rmse_pp": 20.809593484770026
    },
    {
      "training_records": 6,
      "held_out_records": 2,
      "shifted_log_rmse_pp": 1.0390838792829078,
      "linear_rmse_pp": 10.800336134026157
    },
    {
      "training_records": 7,
      "held_out_records": 1,
      "shifted_log_rmse_pp": 1.0542041614377773,
      "linear_rmse_pp": 6.546716372178563
    }
  ],
  "exact_arithmetic": {
    "absolute_lead_pp": 3.299999999999992,
    "relative_score_gain_pct": 4.177215189873418,
    "unsolved_share_reduction_pct": 15.714285714285678
  }
}