1{
  2  "bench_version": 1,
  3  "track": "dynamics",
  4  "model": "rnn_small",
  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    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "weight_decay": 0.0
 17        },
 18        "mean": 0.00222656229743734
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.001,
 23          "weight_decay": 0.0001
 24        },
 25        "mean": 0.002201732073444873
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.003,
 30          "weight_decay": 0.0
 31        },
 32        "mean": 0.0012092416291125119
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.003,
 37          "weight_decay": 0.0001
 38        },
 39        "mean": 0.0011863615218317136
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.01,
 44          "weight_decay": 0.0
 45        },
 46        "mean": 0.0006035602564224973
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.01,
 51          "weight_decay": 0.0001
 52        },
 53        "mean": 0.0006214260865817778
 54      }
 55    ],
 56    "full": {
 57      "mean": 0.000540327702765353,
 58      "std": 0.00010833688702413184,
 59      "per_seed": [
 60        0.0005864710547029972,
 61        0.0004609658499248326,
 62        0.000695098948199302,
 63        0.0006717051728628576,
 64        0.0005945063894614577,
 65        0.00035726596252061427,
 66        0.0004858067259192467,
 67        0.0004708015185315162
 68      ],
 69      "n": 8
 70    }
 71  },
 72  "idea": {
 73    "mean": 0.000540327702765353,
 74    "std": 0.00010833688702413184,
 75    "per_seed": [
 76      0.0005864710547029972,
 77      0.0004609658499248326,
 78      0.000695098948199302,
 79      0.0006717051728628576,
 80      0.0005945063894614577,
 81      0.00035726596252061427,
 82      0.0004858067259192467,
 83      0.0004708015185315162
 84    ],
 85    "n": 8
 86  },
 87  "comparison": {
 88    "delta_mean": 0.0,
 89    "idea_wins": 0,
 90    "n_pairs": 8,
 91    "per_seed_diffs": [
 92      0.0,
 93      0.0,
 94      0.0,
 95      0.0,
 96      0.0,
 97      0.0,
 98      0.0,
 99      0.0
100    ],
101    "p_value": 1.0,
102    "mde": 0.0,
103    "mde_rel_pct": 0.0,
104    "verdict": "no measurable effect",
105    "system_worked": false
106  },
107  "mechanism_signature": {
108    "prediction": "training enforces nonpositive expected barrier drift on continuation states; held-out drift should be lower than zero",
109    "predicted_mean_drift": 0.00018591515208754572,
110    "heldout_mean_drift": 0.00018507556751501397,
111    "margin": 0.02,
112    "confirmed": false
113  },
114  "idea_sweep": [
115    {
116      "cfg": {
117        "lr": 0.001,
118        "lam": 0.5
119      },
120      "result": {
121        "mean": 0.0021250569116091356,
122        "std": 0.000452414474028617,
123        "per_seed": [
124          0.0021977173164486885,
125          0.0023012966848909855,
126          0.0026522083207964897,
127          0.0017550268676131964,
128          0.002219437388703227,
129          0.0011543183354660869,
130          0.002609655261039734,
131          0.0021107951179146767
132        ],
133        "n": 8
134      }
135    },
136    {
137      "cfg": {
138        "lr": 0.003,
139        "lam": 1.0
140      },
141      "result": {
142        "mean": 0.0011597898919717409,
143        "std": 0.0003122565544309852,
144        "per_seed": [
145          0.0009158510947600007,
146          0.001064310665242374,
147          0.0017766037490218878,
148          0.001080201007425785,
149          0.0010183502454310656,
150          0.0006965000065974891,
151          0.0013033286668360233,
152          0.0014231737004593015
153        ],
154        "n": 8
155      }
156    },
157    {
158      "cfg": {
159        "lr": 0.01,
160        "lam": 2.0
161      },
162      "result": {
163        "mean": 0.000540327702765353,
164        "std": 0.00010833688702413184,
165        "per_seed": [
166          0.0005864710547029972,
167          0.0004609658499248326,
168          0.000695098948199302,
169          0.0006717051728628576,
170          0.0005945063894614577,
171          0.00035726596252061427,
172          0.0004858067259192467,
173          0.0004708015185315162
174        ],
175        "n": 8
176      }
177    }
178  ],
179  "math_sanity": {
180    "claim": "E[B_next]-B <= 0 implies failure probability <= B",
181    "expected_B_next_state0": 0.69,
182    "B_state0": 0.7,
183    "drift_state0": -0.010000000000000009,
184    "failure_probability": 0.6666666666666666,
185    "bound_holds": true
186  },
187  "protocol_note": "Matched dynamics task and shared GRU predictor/barrier-head architecture; only auxiliary uniform stochastic barrier loss differs. All idea learning rates are included in baseline sweep."
188}