Equivariant Shared-Mechanism World Model / bench_report.json

Mechanism confirmed, baseline not beaten

Raw ⬇ ZIP
  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      "epochs": 10,
 11      "lam": 0.0
 12    },
 13    "sweep": [
 14      {
 15        "cfg": {
 16          "lr": 0.001,
 17          "epochs": 10,
 18          "lam": 0.0
 19        },
 20        "mean": 0.022081697126850486
 21      },
 22      {
 23        "cfg": {
 24          "lr": 0.003,
 25          "epochs": 10,
 26          "lam": 0.0
 27        },
 28        "mean": 0.0036966767656849697
 29      },
 30      {
 31        "cfg": {
 32          "lr": 0.01,
 33          "epochs": 10,
 34          "lam": 0.0
 35        },
 36        "mean": 0.0033956777187995613
 37      }
 38    ],
 39    "full": {
 40      "mean": 0.0028021317266393453,
 41      "std": 0.001671312456779879,
 42      "per_seed": [
 43        0.001397879095748067,
 44        0.004815699998289347,
 45        0.0011151251383125782,
 46        0.006254006642848253,
 47        0.001966016599908471,
 48        0.002434998285025358,
 49        0.0020769357215613127,
 50        0.0023563923314213753
 51      ],
 52      "n": 8
 53    }
 54  },
 55  "idea": {
 56    "mean": 0.002574730635387823,
 57    "std": 0.0013220860264022514,
 58    "per_seed": [
 59      0.0030565059278160334,
 60      0.0013654178474098444,
 61      0.003834570525214076,
 62      0.005198074039071798,
 63      0.0012422980507835746,
 64      0.0024880883283913136,
 65      0.002220690716058016,
 66      0.0011921996483579278
 67    ],
 68    "n": 8
 69  },
 70  "comparison": {
 71    "delta_mean": -0.0002274010912515223,
 72    "idea_wins": 4,
 73    "n_pairs": 8,
 74    "per_seed_diffs": [
 75      0.0016586268320679665,
 76      -0.0034502821508795023,
 77      0.002719445386901498,
 78      -0.001055932603776455,
 79      -0.0007237185491248965,
 80      5.309004336595535e-05,
 81      0.00014375499449670315,
 82      -0.0011641926830634475
 83    ],
 84    "p_value": 0.72035,
 85    "mde": 0.0015692337631269659,
 86    "mde_rel_pct": 56.00142734934808,
 87    "verdict": "no significant win",
 88    "system_worked": false
 89  },
 90  "math_check": {
 91    "c2_input_composition_max_abs": 0.0,
 92    "c2_output_composition_max_abs": 0.0,
 93    "exact_transform_identity": true
 94  },
 95  "idea_sweep": [
 96    {
 97      "cfg": {
 98        "lr": 0.001,
 99        "epochs": 10,
100        "lam": 0.05
101      },
102      "mean": 0.022260770318098366
103    },
104    {
105      "cfg": {
106        "lr": 0.001,
107        "epochs": 10,
108        "lam": 0.2
109      },
110      "mean": 0.022491522133350372
111    },
112    {
113      "cfg": {
114        "lr": 0.001,
115        "epochs": 10,
116        "lam": 1.0
117      },
118      "mean": 0.020668090670369565
119    },
120    {
121      "cfg": {
122        "lr": 0.003,
123        "epochs": 10,
124        "lam": 0.05
125      },
126      "mean": 0.0037502831255551428
127    },
128    {
129      "cfg": {
130        "lr": 0.003,
131        "epochs": 10,
132        "lam": 0.2
133      },
134      "mean": 0.0036739293136633933
135    },
136    {
137      "cfg": {
138        "lr": 0.003,
139        "epochs": 10,
140        "lam": 1.0
141      },
142      "mean": 0.00422825881105382
143    },
144    {
145      "cfg": {
146        "lr": 0.01,
147        "epochs": 10,
148        "lam": 0.05
149      },
150      "mean": 0.0037559740303549916
151    },
152    {
153      "cfg": {
154        "lr": 0.01,
155        "epochs": 10,
156        "lam": 0.2
157      },
158      "mean": 0.003363642084877938
159    },
160    {
161      "cfg": {
162        "lr": 0.01,
163        "epochs": 10,
164        "lam": 1.0
165      },
166      "mean": 0.0034596045152284205
167    }
168  ],
169  "structural_match": "controlled pendulum rollout has sign-equivariant local dynamics",
170  "mechanism_signature": {
171    "mean_prediction_abs": 0.7621783018112183,
172    "trained_model_c2_penalty": 0.0005555228563025594,
173    "prediction_metric_seed0": 0.0030565059278160334,
174    "expected_c2_penalty_direction": "lower is more equivariant",
175    "confirmed": true
176  }
177}