[
  {
    "title": "Fold 1 \u2014 38,283 held-out parcels",
    "groups": [
      [
        "",
        [
          {
            "label": "Validation-selected ensemble (9 members)",
            "values": [
              54.93308503001464,
              95.02912519917457,
              61.207594983632276
            ],
            "source": "results/ensemble_library/Fold_1/report.json"
          },
          {
            "label": "Eight-model control: six + baseline checkpoints 27 / 78",
            "values": [
              54.15609283788674,
              94.96643418749838,
              59.63298862355673
            ],
            "source": "results/ensemble_library/Fold_1/report.json"
          },
          {
            "label": "Six-model ensemble: previous five + CNN \u2192 group TAE",
            "values": [
              54.235271797228854,
              94.97165843847138,
              60.06296844653283
            ],
            "source": "results/requested_comparison/six_model_ensemble/report.json"
          },
          {
            "label": "Five-model ensemble: four neural models + XGBoost",
            "values": [
              53.72404814348768,
              94.88023404644359,
              59.57984810873117
            ],
            "source": "results/requested_comparison/xgboost_ensemble/report.json"
          },
          {
            "label": "Four-model ensemble: baseline + capacity + groups + dropout",
            "values": [
              52.61947706839384,
              94.79664603087532,
              58.0201965919677
            ],
            "source": "results/requested_comparison/Fold_1/report.json"
          },
          {
            "label": "Three-model ensemble: baseline + capacity + groups",
            "values": [
              52.2793345220272,
              94.7574641485777,
              57.79896008395639
            ],
            "source": "results/requested_comparison/Fold_1/report.json"
          },
          {
            "label": "XGBoost: frozen embeddings + spectral summaries",
            "values": [
              52.2254181167317,
              93.5062560405402,
              61.61889361802421
            ],
            "source": "results/multiclass_probes/Fold_1/results.json"
          },
          {
            "label": "Date-dropout fine-tuning",
            "values": [
              51.67126199738541,
              94.4152757098451,
              58.78236262519556
            ],
            "source": "results/requested_comparison/Fold_1/report.json"
          },
          {
            "label": "Baseline checkpoint ensemble (epochs 27 / 38 / 78)",
            "values": [
              51.60732823966951,
              94.67387613300943,
              57.47223618501322
            ],
            "source": "results/requested_comparison/Fold_1/report.json"
          },
          {
            "label": "Original PSE+TAE baseline (cached evaluation)",
            "values": [
              51.31040863685266,
              94.23765117676253,
              58.35255834670093
            ],
            "source": "results/requested_comparison/Fold_1/report.json"
          },
          {
            "label": "CNN \u2192 group TAE correction (snapshot epoch 4)",
            "values": [
              51.65454151863263,
              94.37086957657445,
              58.66253159399146
            ],
            "source": "results/requested_comparison/six_model_ensemble/report.json"
          },
          {
            "label": "Baseline-size PSE+TAE distilled from scratch (epoch 86)",
            "values": [
              53.35062972779878,
              94.66081550557689,
              60.69560342779622
            ],
            "source": "results/distillation/Fold_1/report.json"
          },
          {
            "label": "Matched baseline-size CE-only control (epoch 80)",
            "values": [
              51.67979313433331,
              94.43356058825067,
              58.466877634443435
            ],
            "source": "results/distillation/Fold_1/report.json"
          }
        ]
      ],
      [
        "Earlier variants \u2014 no established improvement",
        [
          {
            "label": "CNN spatial correction",
            "values": [
              51.279162452007895,
              94.33952407073636,
              58.411938090161286
            ],
            "source": "results/requested_comparison/Fold_1/report.json"
          },
          {
            "label": "Patch baseline, validation-best epoch 80",
            "values": [
              49.93888753785837,
              94.17234803959983,
              57.56955796452944
            ],
            "source": "results/requested_comparison/Fold_1/report.json"
          },
          {
            "label": "Patch baseline, ensemble checkpoint 38",
            "values": [
              50.48081293679929,
              94.37870595303399,
              56.721247944936536
            ],
            "source": "results/requested_comparison/Fold_1/report.json"
          },
          {
            "label": "Group trajectories (epoch 38 ensemble snapshot)",
            "values": [
              49.437840416342176,
              93.98166287908471,
              56.53221654694958
            ],
            "source": "results/complementarity/snapshot_20261018_1605/test_evaluation/report.json",
            "group": "earlier"
          },
          {
            "label": "Capacity control (epoch 67 ensemble snapshot)",
            "values": [
              50.160251455192686,
              94.1044327769506,
              58.00589302861668
            ],
            "source": "results/complementarity/snapshot_20261018_1605/test_evaluation/report.json",
            "group": "earlier"
          },
          {
            "label": "Embedding moments (mean / std / skew / kurtosis)",
            "values": [
              46.15944847885274,
              93.9790507535982,
              51.16271573930846
            ],
            "source": "results/moments/Fold_1/conf_mat.pkl",
            "group": "earlier"
          },
          {
            "label": "Frozen encoder + date kurtosis",
            "values": [
              51.468444652295965,
              94.30556643941175,
              59.405290343896496
            ],
            "source": "results/date_kurtosis/Fold_1/conf_mat.pkl",
            "group": "earlier"
          },
          {
            "label": "Scaled embedding kurtosis, gradients on",
            "values": [
              47.98,
              94.17,
              null
            ],
            "source": "notes.md",
            "group": "earlier"
          },
          {
            "label": "Parcel mean / std in the pixel encoder",
            "values": [
              50.11,
              null,
              null
            ],
            "source": "notes.md",
            "group": "earlier"
          },
          {
            "label": "Parcel mean / std / kurtosis in the pixel encoder",
            "values": [
              50.05,
              null,
              null
            ],
            "source": "notes.md",
            "group": "earlier"
          },
          {
            "label": "Inverse-square-root class-weighted loss",
            "values": [
              49.07,
              92.14,
              null
            ],
            "source": "notes.md",
            "group": "earlier"
          },
          {
            "label": "Small per-pixel temporal attention, from scratch",
            "values": [
              48.10858560576429,
              93.73351095786641,
              55.818663310797625
            ],
            "source": "results/pixel_temporal_scratch/Fold_1/report.json",
            "group": "earlier"
          },
          {
            "label": "Small temporal + spatial attention, from scratch",
            "values": [
              49.200045423941965,
              94.02868113784186,
              55.018680447837966
            ],
            "source": "results/pixel_temporal_scratch/Fold_1/report.json",
            "group": "earlier"
          },
          {
            "label": "Full-width embedding-only correction",
            "values": [
              51.56003908043143,
              94.34474832170937,
              58.974577977199935
            ],
            "source": "results/grouped_corrections/Fold_1/report.json",
            "group": "earlier"
          },
          {
            "label": "Full-width group temporal correction",
            "values": [
              51.244643853192585,
              94.23765117676253,
              58.4145327419458
            ],
            "source": "results/grouped_corrections/Fold_1/report.json",
            "group": "earlier"
          },
          {
            "label": "Full-width group + spatial attention correction",
            "values": [
              51.46421578108142,
              94.36303320011493,
              58.366341094545234
            ],
            "source": "results/grouped_corrections/Fold_1/report.json",
            "group": "earlier"
          },
          {
            "label": "Per-pixel temporal-distribution correction",
            "values": [
              51.14892057151782,
              94.3525846981689,
              57.665216576517174
            ],
            "source": "results/frozen_corrections/Fold_1/results.json",
            "group": "earlier"
          },
          {
            "label": "Small patch-pixel temporal correction",
            "values": [
              51.779382450863565,
              94.38915445498002,
              58.71511913744347
            ],
            "source": "results/frozen_corrections/Fold_1/results.json",
            "group": "earlier"
          },
          {
            "label": "Small temporal + spatial correction, no coordinates",
            "values": [
              51.79733905260681,
              94.39960295692606,
              58.70125937430109
            ],
            "source": "results/frozen_corrections/Fold_1/results.json",
            "group": "earlier"
          },
          {
            "label": "Small temporal + spatial correction, with coordinates",
            "values": [
              51.69318989046536,
              94.35780894914191,
              58.58341478211055
            ],
            "source": "results/frozen_corrections/Fold_1/results.json",
            "group": "earlier"
          },
          {
            "label": "Small replacement classifier (best new variant by validation)",
            "values": [
              50.72609030301305,
              93.99994775749026,
              59.075365749764764
            ],
            "source": "results/frozen_trajectories/Fold_1/results.json",
            "group": "earlier"
          }
        ]
      ]
    ],
    "caption": [
      "All values are percentages. One fold and one seed; these comparisons are exploratory.",
      "Earlier variants include ties and inconclusive gains. Missing values are unmeasured, not zero.",
      "Library additions and mixture weights are chosen on validation; test scores do not select the ensemble."
    ]
  },
  {
    "title": "Five-fold results \u2014 mean and sample standard deviation",
    "groups": [
      [
        "",
        [
          {
            "label": "PSE+TAE reproduction",
            "values": [
              50.96253822685437,
              94.22293256405372,
              58.504182483372226
            ],
            "sd": [
              0.6019289857741531,
              0.09171186030656048,
              0.9397230573626493
            ],
            "source": "results/requested_comparison/baseline_five_folds.json"
          }
        ]
      ],
      [
        "Earlier alternatives",
        [
          {
            "label": "L-TAE reproduction",
            "values": [
              50.67,
              94.33,
              null
            ],
            "sd": [
              0.95,
              0.04,
              null
            ],
            "source": "notes.md"
          },
          {
            "label": "RoPE",
            "values": [
              50.1261056275956,
              94.20308047859001,
              57.772785466570326
            ],
            "sd": [
              0.42281397969611906,
              0.10411895189135764,
              0.5186037310544795
            ],
            "source": "results/rope/Fold_*/conf_mat.pkl"
          },
          {
            "label": "Scaled kurtosis, stopped gradient",
            "values": [
              50.79,
              94.17,
              null
            ],
            "sd": [
              0.75,
              0.21,
              null
            ],
            "source": "notes.md"
          },
          {
            "label": "XGBoost on band mean / std",
            "values": [
              42.16,
              93.27,
              null
            ],
            "sd": [
              0.41,
              0.1,
              null
            ],
            "source": "notes.md"
          },
          {
            "label": "XGBoost on band mean / std + kurtosis",
            "values": [
              41.47,
              null,
              null
            ],
            "sd": [
              null,
              null,
              null
            ],
            "source": "notes.md"
          }
        ]
      ]
    ],
    "caption": [
      "Standard deviations are shown in smaller type; they are not confidence intervals.",
      "Training sets overlap across folds. Do not compare these means as if they were fold-1 results.",
      "Distillation has one completed test fold only; the further-fold queue is paused."
    ]
  }
]