{
  "bench_version": 1,
  "track": "sequence",
  "model": "transformer_tiny",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.001
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001
        },
        "mean": 0.4451246112585068
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 0.4927415996789932
      },
      {
        "cfg": {
          "lr": 0.01
        },
        "mean": 1.6619523614645004
      }
    ],
    "full": {
      "mean": 0.45439932495355606,
      "std": 0.04081052029431072,
      "per_seed": [
        0.4278562068939209,
        0.3956301808357239,
        0.46248915791511536,
        0.49452289938926697,
        0.44014036655426025,
        0.4744090139865875,
        0.5266208648681641,
        0.41352590918540955
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 1.0940600484609604,
    "std": 0.0844539494160516,
    "per_seed": [
      1.086983323097229,
      0.9213972091674805,
      1.21669602394104,
      1.1440339088439941,
      1.0787184238433838,
      1.0460021495819092,
      1.177429437637329,
      1.0812199115753174
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": 0.6396607235074043,
    "idea_wins": 0,
    "n_pairs": 8,
    "per_seed_diffs": [
      0.6591271162033081,
      0.5257670283317566,
      0.7542068660259247,
      0.6495110094547272,
      0.6385780572891235,
      0.5715931355953217,
      0.650808572769165,
      0.6676940023899078
    ],
    "p_value": 0.0081,
    "mde": 0.056634104586998066,
    "mde_rel_pct": 12.463509841874568,
    "verdict": "idea worse (significant)",
    "system_worked": false
  },
  "mechanism_signature": {
    "track_match": "multi-token sequence forecast",
    "mean_events_per_sample": 1536.0,
    "mean_shared_node_visits": 32.0,
    "predicted_shared_visits": 32.0,
    "observed_to_predicted_ratio": 1.0,
    "source_message_abs_mean": 0.6570842266082764,
    "prediction": "each sequence token visits every shared hidden node once",
    "confirmed": true
  },
  "idea_sweep": [
    {
      "cfg": {
        "lr": 0.001
      },
      "mean": 1.0940600484609604
    },
    {
      "cfg": {
        "lr": 0.003
      },
      "mean": 1.0993548929691315
    },
    {
      "cfg": {
        "lr": 0.01
      },
      "mean": 1.121301993727684
    }
  ],
  "protocol_notes": {
    "epochs": 12,
    "n_train": 400,
    "n_test": 200,
    "baseline_grid": [
      {
        "lr": 0.001
      },
      {
        "lr": 0.003
      },
      {
        "lr": 0.01
      }
    ],
    "idea_grid": [
      {
        "lr": 0.001
      },
      {
        "lr": 0.003
      },
      {
        "lr": 0.01
      }
    ],
    "architecture": "transformer baseline vs trainable delayed shared-accumulator graph"
  }
}