Detailed-Balance Graph Transport Layer / bench_report.json

Failed on benchmark

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    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 0.001443413842935115
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.0009000279824249446
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.01
 27        },
 28        "mean": 0.0005028522355132736
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.00042912505523418076,
 33      "std": 0.00017915832886079294,
 34      "per_seed": [
 35        0.000361639482434839,
 36        0.00038413508445955813,
 37        0.00039487716276198626,
 38        0.0008707572123967111,
 39        0.0004429189139045775,
 40        0.0002050218463409692,
 41        0.00038642011350020766,
 42        0.00038723062607459724
 43      ],
 44      "n": 8
 45    },
 46    "union_extra": [
 47      {
 48        "cfg": {
 49          "lr": 0.0033333333333333335
 50        },
 51        "full": {
 52          "mean": 0.0008873757251421921,
 53          "std": 0.00023680384896603653,
 54          "per_seed": [
 55            0.0007913073641248047,
 56            0.000795992324128747,
 57            0.0006051670643500984,
 58            0.001341381692327559,
 59            0.0006803745054639876,
 60            0.0007620747201144695,
 61            0.0009469457436352968,
 62            0.0011757623869925737
 63          ],
 64          "n": 8
 65        }
 66      },
 67      {
 68        "cfg": {
 69          "lr": 0.03
 70        },
 71        "full": {
 72          "mean": 0.00019077555771218613,
 73          "std": 0.00010217559367124779,
 74          "per_seed": [
 75            0.00023975277144927531,
 76            0.00015833131328690797,
 77            0.00018476269906386733,
 78            0.00011282971536275,
 79            0.0004178391245659441,
 80            7.224319415399805e-05,
 81            0.00023147270258050412,
 82            0.00010897294123424217
 83          ],
 84          "n": 8
 85        }
 86      }
 87    ]
 88  },
 89  "idea": {
 90    "mean": 0.0009516679710941389,
 91    "std": 0.0002400579622568702,
 92    "per_seed": [
 93      0.0009455473045818508,
 94      0.0009278051438741386,
 95      0.0008952125790528953,
 96      0.0009500241139903665,
 97      0.0007523722597397864,
 98      0.0006179050542414188,
 99      0.0015001472784206271,
100      0.001024330034852028
101    ],
102    "n": 8,
103    "sweep": [
104      {
105        "cfg": {
106          "lr": 0.0033333333333333335,
107          "dt": 0.05
108        },
109        "mean": 0.0009296472853748128
110      },
111      {
112        "cfg": {
113          "lr": 0.01,
114          "dt": 0.05
115        },
116        "mean": NaN
117      },
118      {
119        "cfg": {
120          "lr": 0.03,
121          "dt": 0.05
122        },
123        "mean": NaN
124      }
125    ],
126    "best_cfg": {
127      "lr": 0.0033333333333333335,
128      "dt": 0.05
129    }
130  },
131  "comparison": {
132    "delta_mean": 0.0005225429158599582,
133    "idea_wins": 0,
134    "n_pairs": 8,
135    "per_seed_diffs": [
136      0.0005839078221470118,
137      0.0005436700594145805,
138      0.000500335416290909,
139      7.926690159365535e-05,
140      0.0003094533458352089,
141      0.00041288320790044963,
142      0.0011137271649204195,
143      0.0006370994087774307
144    ],
145    "p_value": 0.0081,
146    "mde": 0.00024909869994053333,
147    "mde_rel_pct": 58.04804378168876,
148    "verdict": "idea worse (significant)",
149    "system_worked": false
150  },
151  "mechanism_signature": {
152    "mass_conservation_rel_error": 3.493384204134031e-08,
153    "observed_energy_change": -0.003466067530098371,
154    "predicted_energy_derivative": -0.04637642973102629,
155    "derivative_relative_error": 0.5008485137480757,
156    "energy_dissipates": 1.0,
157    "confirmed": false
158  },
159  "protocol_notes": {
160    "structural_match": "controlled pendulum rollout tests stability of recurrent dynamical states",
161    "paired_seeds": 8,
162    "epochs": 20,
163    "samples_per_seed": [
164      400,
165      200
166    ],
167    "device": "cuda"
168  }
169}