{
  "bench_version": 1,
  "track": "dynamics",
  "model": "matched_ternary_rnn",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.01,
      "weight_decay": 0.0,
      "epochs": 12
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "weight_decay": 0.0,
          "epochs": 12
        },
        "mean": 0.020193250384181738
      },
      {
        "cfg": {
          "lr": 0.003,
          "weight_decay": 0.0,
          "epochs": 12
        },
        "mean": 0.00683436612598598
      },
      {
        "cfg": {
          "lr": 0.01,
          "weight_decay": 0.0,
          "epochs": 12
        },
        "mean": 0.002462160831782967
      },
      {
        "cfg": {
          "lr": 0.003,
          "weight_decay": 0.0001,
          "epochs": 12
        },
        "mean": 0.006820366135798395
      },
      {
        "cfg": {
          "lr": 0.003,
          "weight_decay": 0.001,
          "epochs": 12
        },
        "mean": 0.0067008426412940025
      }
    ],
    "full": {
      "mean": 0.0032130961772054434,
      "std": 0.0010584401251843786,
      "per_seed": [
        0.0025007869116961956,
        0.0035816810559481382,
        0.0025283501017838717,
        0.001237825257703662,
        0.0037454410921782255,
        0.0029973499476909637,
        0.004620881285518408,
        0.0044924537651240826
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.041399288922548294,
    "std": 0.005352299829694784,
    "per_seed": [
      0.043414220213890076,
      0.03751295432448387,
      0.051180846989154816,
      0.048405274748802185,
      0.037778589874506,
      0.035160426050424576,
      0.03840555250644684,
      0.039336446672677994
    ],
    "n": 8,
    "best_cfg": {
      "lr": 0.01,
      "weight_decay": 0.0,
      "epochs": 12,
      "R": 1.2
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "weight_decay": 0.0,
          "epochs": 12,
          "R": 0.6
        },
        "mean": 0.04771255608648062
      },
      {
        "cfg": {
          "lr": 0.003,
          "weight_decay": 0.0,
          "epochs": 12,
          "R": 0.9
        },
        "mean": 0.05168276559561491
      },
      {
        "cfg": {
          "lr": 0.01,
          "weight_decay": 0.0,
          "epochs": 12,
          "R": 1.2
        },
        "mean": 0.04512832406908274
      }
    ]
  },
  "comparison": {
    "delta_mean": 0.03818619274534285,
    "idea_wins": 0,
    "n_pairs": 8,
    "per_seed_diffs": [
      0.04091343330219388,
      0.03393127326853573,
      0.048652496887370944,
      0.04716744949109852,
      0.03403314878232777,
      0.03216307610273361,
      0.03378467122092843,
      0.03484399290755391
    ],
    "p_value": 0.0081,
    "mde": 0.005472251358790593,
    "mde_rel_pct": 170.3108483839418,
    "verdict": "idea worse (significant)",
    "system_worked": false
  },
  "mechanism_signature": {
    "prediction": "disorder above Rc suppresses persistent mean-field basin magnetization",
    "predicted_Rc": 0.7978845608028654,
    "tested_R": 1.2,
    "observed_mean_abs_late_m": 0.010480209394700069,
    "observed_per_seed": [
      0.016682207817211747,
      0.012151359231211245,
      0.008320131455548108,
      0.01342325797304511,
      0.006388145731762052,
      0.0017870731899165548,
      0.01985514839179814,
      0.0052343513671075925
    ],
    "confirmed": true
  },
  "math_sanity": {
    "Rc_predicted": 0.7978845608028654,
    "rows": [
      {
        "R": 0.6,
        "observed_slope": 1.3298076013374933,
        "predicted_slope": 1.329807601338109,
        "abs_error": 6.157296894571118e-13
      },
      {
        "R": 0.7978845608028654,
        "observed_slope": 0.9999999999997379,
        "predicted_slope": 1.0,
        "abs_error": 2.6212365611399946e-13
      },
      {
        "R": 1.0,
        "observed_slope": 0.7978845608027323,
        "predicted_slope": 0.7978845608028654,
        "abs_error": 1.3311574065255627e-13
      }
    ],
    "max_abs_error": 6.157296894571118e-13
  },
  "protocol_notes": "Dynamics selected because the idea targets recurrent stability/basin dependence. Baseline and idea share trainable input, recurrent, and head maps; only state update/disorder differs."
}