{
  "seed": 783,
  "target": "eventual winner from candidate afterstate and known rack leave; tied games excluded",
  "split": "80/20 by whole game; no hyperparameter selection on test games",
  "train_games": 2044,
  "test_games": 511,
  "train_turns": 45768,
  "test_turns": 11423,
  "majority_accuracy": 0.5029326796813447,
  "dataset": {
    "games": 2566,
    "samples": 57191,
    "records": 60623,
    "excluded": {
      "pass_exchange_or_adjustment": 3181,
      "tied_games": 11
    }
  },
  "circuit": {
    "pn": 146,
    "kc": 2580,
    "mbon": 48,
    "apl": 1,
    "pn_kc_edges": 11004,
    "pn_kc_synapses": 173587,
    "unknown_or_modulatory_edge_rows": 28,
    "nt_counts": {
      "ACH": 46220,
      "GABA": 13307,
      "GLUT": 126,
      "": 28
    },
    "sign_source": "neurons.csv.gz nt_type of presynaptic root_id; no edge-level sign overrides"
  },
  "metrics": {
    "real": {
      "accuracy": 0.5015319968484636,
      "balanced_accuracy": 0.5012961023123466,
      "auc": 0.5064647084267957,
      "game_bootstrap_95ci": [
        0.4932773107993575,
        0.5096999427566126
      ],
      "iterations": 5
    },
    "shuffled": {
      "accuracy": 0.5067845574717675,
      "balanced_accuracy": 0.506623812733924,
      "auc": 0.5089716588938541,
      "game_bootstrap_95ci": [
        0.4991261303748353,
        0.5143782469937473
      ],
      "iterations": 5
    },
    "logistic": {
      "accuracy": 0.5473168169482623,
      "balanced_accuracy": 0.5477505440662217,
      "auc": 0.5616725847950851,
      "game_bootstrap_95ci": [
        0.5374347002729563,
        0.55693126770767
      ],
      "iterations": 6
    }
  },
  "swaps": 110040,
  "degree_preserved": true,
  "limitations": [
    "Prediction accuracy on expert-played afterstates is not head-to-head playing strength.",
    "Single seed and one held-out split; no evidence that real wiring must outperform shuffled.",
    "ACh positive, GABA negative, glutamate negative are modeling assumptions; transmitter identity alone does not establish receptor-specific effect. Other or unknown transmitter edges receive zero effective weight.",
    "Input projection, row normalization, one-step APL feedback gain 0.5, top-5% threshold and output pooling are model choices, not measured physiology.",
    "The board encoding omits letter identity and score difference; winner prediction is deliberately a limited baseline.",
    "ENABLE is an open word-game lexicon, not an official tournament Scrabble dictionary.",
    "Recorded passes, exchanges and terminal adjustments do not become candidate training examples; terminal adjustments still determine winner labels."
  ]
}