Differentiable Physics-Equilibrium Projection / bench_report.json

✓✓ Beats tuned baseline

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": 0.012,
 10      "epochs": 25,
 11      "steps": 0
 12    },
 13    "sweep": [
 14      {
 15        "cfg": {
 16          "lr": 0.003,
 17          "epochs": 25,
 18          "steps": 0
 19        },
 20        "mean": 0.0008186178019968793
 21      },
 22      {
 23        "cfg": {
 24          "lr": 0.006,
 25          "epochs": 25,
 26          "steps": 0
 27        },
 28        "mean": 0.0005262107733869925
 29      },
 30      {
 31        "cfg": {
 32          "lr": 0.012,
 33          "epochs": 25,
 34          "steps": 0
 35        },
 36        "mean": 0.0003371372222318314
 37      }
 38    ],
 39    "full": {
 40      "mean": 0.00027095018594991416,
 41      "std": 0.00010329659269599283,
 42      "per_seed": [
 43        0.0003306560392957181,
 44        0.0003460621228441596,
 45        0.00033211964182555676,
 46        0.0003397110849618912,
 47        0.0003853111120406538,
 48        0.00018891783838625997,
 49        7.910988642834127e-05,
 50        0.00016571376181673259
 51      ],
 52      "n": 8
 53    }
 54  },
 55  "idea": {
 56    "mean": 8.354870317361929e-06,
 57    "std": 8.429703290269694e-07,
 58    "per_seed": [
 59      9.509363735560328e-06,
 60      7.033356723695761e-06,
 61      8.578040251450147e-06,
 62      7.191956683527678e-06,
 63      9.002241313282866e-06,
 64      9.134431820712052e-06,
 65      8.445092134934384e-06,
 66      7.944479875732213e-06
 67    ],
 68    "n": 8
 69  },
 70  "comparison": {
 71    "delta_mean": -0.00026259531563255223,
 72    "idea_wins": 8,
 73    "n_pairs": 8,
 74    "per_seed_diffs": [
 75      -0.00032114667556015775,
 76      -0.00033902876612046384,
 77      -0.0003235416015741066,
 78      -0.0003325191282783635,
 79      -0.00037630887072737096,
 80      -0.00017978340656554792,
 81      -7.066479429340689e-05,
 82      -0.00015776928194100037
 83    ],
 84    "p_value": 0.0081,
 85    "mde": 9.240406097577938e-05,
 86    "mde_rel_pct": 34.10370827088508,
 87    "verdict": "idea better (significant)",
 88    "system_worked": true
 89  },
 90  "mechanism_signature": {
 91    "track_match": "dynamics: controlled pendulum rollout",
 92    "idea_config_sweep": [
 93      {
 94        "cfg": {
 95          "lr": 0.003,
 96          "epochs": 25,
 97          "steps": 1
 98        },
 99        "mean": 8.078179348558479e-06
100      },
101      {
102        "cfg": {
103          "lr": 0.006,
104          "epochs": 25,
105          "steps": 2
106        },
107        "mean": 8.078180940174207e-06
108      },
109      {
110        "cfg": {
111          "lr": 0.012,
112          "epochs": 25,
113          "steps": 3
114        },
115        "mean": 8.078180940174207e-06
116      }
117    ],
118    "best_idea_cfg": {
119      "lr": 0.003,
120      "epochs": 25,
121      "steps": 1
122    },
123    "mechanism": {
124      "claim": "implicit equality projection removes the measured equilibrium residual",
125      "trained_model_raw_residual_mean": 0.7149586826562881,
126      "trained_model_projected_residual_mean": 0.0,
127      "residual_ratio": 0.0,
128      "confirmed": true
129    }
130  },
131  "protocol_notes": "8 paired seeds; baseline and idea share rnn_small, dataset sizes, lr/epoch union, and train_model."
132}