Universal Clock Regularization for Recurrent Dynamics / 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    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 0.0016091743600554764
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.0008843674877425656
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.01
 27        },
 28        "mean": 0.0004280267676222138
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.0003420520333747845,
 33      "std": 0.00016964803808050996,
 34      "per_seed": [
 35        0.0002832822792697698,
 36        0.00029989160248078406,
 37        0.0003870299260597676,
 38        0.0007419032626785338,
 39        0.0003700012748595327,
 40        0.00012643028458114713,
 41        0.0002213301049778238,
 42        0.00030654753209091723
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 0.00031757883516547736,
 49    "std": 0.00012142437873431676,
 50    "per_seed": [
 51      0.0002894827339332551,
 52      0.0003101625479757786,
 53      0.0004648705362342298,
 54      0.0005544392624869943,
 55      0.0002837668580468744,
 56      0.00024210276023950428,
 57      0.00024163190391846,
 58      0.00015417407848872244
 59    ],
 60    "n": 8
 61  },
 62  "comparison": {
 63    "delta_mean": -2.4473198209307157e-05,
 64    "idea_wins": 3,
 65    "n_pairs": 8,
 66    "per_seed_diffs": [
 67      6.200454663485289e-06,
 68      1.0270945494994521e-05,
 69      7.78406101744622e-05,
 70      -0.00018746400019153953,
 71      -8.623441681265831e-05,
 72      0.00011567247565835714,
 73      2.0301798940636218e-05,
 74      -0.00015237345360219479
 75    ],
 76    "p_value": 0.5604,
 77    "mde": 8.994591863352359e-05,
 78    "mde_rel_pct": 26.29597542399941,
 79    "verdict": "no significant win",
 80    "system_worked": false
 81  },
 82  "mechanism_signature": {
 83    "mechanism_signature": {
 84      "prediction": "clock regularization reduces phase-velocity variance on trained recurrent trajectories",
 85      "baseline_test_mean_phase_velocity_var": 9.041786380112171,
 86      "idea_test_mean_phase_velocity_var": 0.005592545494437218,
 87      "baseline_test_mean_normalized_clock_error": 13.826433002948761,
 88      "idea_test_mean_normalized_clock_error": 0.46902846172451973,
 89      "confirmed": true
 90    },
 91    "idea_lr_grid": {
 92      "0.001": {
 93        "mean": 0.001738299906719476,
 94        "std": 0.0006141943245580857,
 95        "per_seed": [
 96          0.0011724976357072592,
 97          0.0012341204565018415,
 98          0.002423458034172654,
 99          0.0024381368421018124,
100          0.001139834406785667,
101          0.0011474714847281575,
102          0.0017310797702521086,
103          0.0026198006235063076
104        ],
105        "n": 8
106      },
107      "0.003": {
108        "mean": 0.0008836812994559295,
109        "std": 0.00024104514786964793,
110        "per_seed": [
111          0.0007145126583054662,
112          0.0006794828223064542,
113          0.0011653573019430041,
114          0.0012316256761550903,
115          0.000606869871262461,
116          0.0006913324468769133,
117          0.0008201156160794199,
118          0.0011601540027186275
119        ],
120        "n": 8
121      },
122      "0.01": {
123        "mean": 0.00031757883516547736,
124        "std": 0.00012142437873431676,
125        "per_seed": [
126          0.0002894827339332551,
127          0.0003101625479757786,
128          0.0004648705362342298,
129          0.0005544392624869943,
130          0.0002837668580468744,
131          0.00024210276023950428,
132          0.00024163190391846,
133          0.00015417407848872244
134        ],
135        "n": 8
136      }
137    },
138    "track_rationale": "Dynamics contains recurrent controlled pendulum trajectories and is the mandated structural match."
139  }
140}