1{
  2  "bench_version": 1,
  3  "track": "dynamics",
  4  "model": "matched_ternary_rnn",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.01,
 10      "weight_decay": 0.0,
 11      "epochs": 12
 12    },
 13    "sweep": [
 14      {
 15        "cfg": {
 16          "lr": 0.001,
 17          "weight_decay": 0.0,
 18          "epochs": 12
 19        },
 20        "mean": 0.020193250384181738
 21      },
 22      {
 23        "cfg": {
 24          "lr": 0.003,
 25          "weight_decay": 0.0,
 26          "epochs": 12
 27        },
 28        "mean": 0.00683436612598598
 29      },
 30      {
 31        "cfg": {
 32          "lr": 0.01,
 33          "weight_decay": 0.0,
 34          "epochs": 12
 35        },
 36        "mean": 0.002462160831782967
 37      },
 38      {
 39        "cfg": {
 40          "lr": 0.003,
 41          "weight_decay": 0.0001,
 42          "epochs": 12
 43        },
 44        "mean": 0.006820366135798395
 45      },
 46      {
 47        "cfg": {
 48          "lr": 0.003,
 49          "weight_decay": 0.001,
 50          "epochs": 12
 51        },
 52        "mean": 0.0067008426412940025
 53      }
 54    ],
 55    "full": {
 56      "mean": 0.0032130961772054434,
 57      "std": 0.0010584401251843786,
 58      "per_seed": [
 59        0.0025007869116961956,
 60        0.0035816810559481382,
 61        0.0025283501017838717,
 62        0.001237825257703662,
 63        0.0037454410921782255,
 64        0.0029973499476909637,
 65        0.004620881285518408,
 66        0.0044924537651240826
 67      ],
 68      "n": 8
 69    }
 70  },
 71  "idea": {
 72    "mean": 0.041399288922548294,
 73    "std": 0.005352299829694784,
 74    "per_seed": [
 75      0.043414220213890076,
 76      0.03751295432448387,
 77      0.051180846989154816,
 78      0.048405274748802185,
 79      0.037778589874506,
 80      0.035160426050424576,
 81      0.03840555250644684,
 82      0.039336446672677994
 83    ],
 84    "n": 8,
 85    "best_cfg": {
 86      "lr": 0.01,
 87      "weight_decay": 0.0,
 88      "epochs": 12,
 89      "R": 1.2
 90    },
 91    "sweep": [
 92      {
 93        "cfg": {
 94          "lr": 0.001,
 95          "weight_decay": 0.0,
 96          "epochs": 12,
 97          "R": 0.6
 98        },
 99        "mean": 0.04771255608648062
100      },
101      {
102        "cfg": {
103          "lr": 0.003,
104          "weight_decay": 0.0,
105          "epochs": 12,
106          "R": 0.9
107        },
108        "mean": 0.05168276559561491
109      },
110      {
111        "cfg": {
112          "lr": 0.01,
113          "weight_decay": 0.0,
114          "epochs": 12,
115          "R": 1.2
116        },
117        "mean": 0.04512832406908274
118      }
119    ]
120  },
121  "comparison": {
122    "delta_mean": 0.03818619274534285,
123    "idea_wins": 0,
124    "n_pairs": 8,
125    "per_seed_diffs": [
126      0.04091343330219388,
127      0.03393127326853573,
128      0.048652496887370944,
129      0.04716744949109852,
130      0.03403314878232777,
131      0.03216307610273361,
132      0.03378467122092843,
133      0.03484399290755391
134    ],
135    "p_value": 0.0081,
136    "mde": 0.005472251358790593,
137    "mde_rel_pct": 170.3108483839418,
138    "verdict": "idea worse (significant)",
139    "system_worked": false
140  },
141  "mechanism_signature": {
142    "prediction": "disorder above Rc suppresses persistent mean-field basin magnetization",
143    "predicted_Rc": 0.7978845608028654,
144    "tested_R": 1.2,
145    "observed_mean_abs_late_m": 0.010480209394700069,
146    "observed_per_seed": [
147      0.016682207817211747,
148      0.012151359231211245,
149      0.008320131455548108,
150      0.01342325797304511,
151      0.006388145731762052,
152      0.0017870731899165548,
153      0.01985514839179814,
154      0.0052343513671075925
155    ],
156    "confirmed": true
157  },
158  "math_sanity": {
159    "Rc_predicted": 0.7978845608028654,
160    "rows": [
161      {
162        "R": 0.6,
163        "observed_slope": 1.3298076013374933,
164        "predicted_slope": 1.329807601338109,
165        "abs_error": 6.157296894571118e-13
166      },
167      {
168        "R": 0.7978845608028654,
169        "observed_slope": 0.9999999999997379,
170        "predicted_slope": 1.0,
171        "abs_error": 2.6212365611399946e-13
172      },
173      {
174        "R": 1.0,
175        "observed_slope": 0.7978845608027323,
176        "predicted_slope": 0.7978845608028654,
177        "abs_error": 1.3311574065255627e-13
178      }
179    ],
180    "max_abs_error": 6.157296894571118e-13
181  },
182  "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."
183}