1{
  2  "bench_version": 1,
  3  "track": "singular_value_constitutive_energy",
  4  "model": "shared_invariant_mlp_width16",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.006
 10    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 0.022746754344552755
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.01957778725773096
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.006
 27        },
 28        "mean": 0.018540473887696862
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.017490515718236566,
 33      "std": 0.003886538162934768,
 34      "per_seed": [
 35        0.015612582676112652,
 36        0.01946304365992546,
 37        0.019211620092391968,
 38        0.01987464912235737,
 39        0.022817162796854973,
 40        0.015229465439915657,
 41        0.009119716472923756,
 42        0.01859588548541069
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 0.48515734821558,
 49    "std": 0.02347761538492244,
 50    "per_seed": [
 51      0.4996644854545593,
 52      0.48556891083717346,
 53      0.44942617416381836,
 54      0.46209976077079773,
 55      0.4696982502937317,
 56      0.5224891304969788,
 57      0.5122184753417969,
 58      0.4800935983657837
 59    ],
 60    "n": 8
 61  },
 62  "comparison": {
 63    "delta_mean": 0.4676668324973434,
 64    "idea_wins": 0,
 65    "n_pairs": 8,
 66    "per_seed_diffs": [
 67      0.4840519027784467,
 68      0.466105867177248,
 69      0.4302145540714264,
 70      0.44222511164844036,
 71      0.4468810874968767,
 72      0.5072596650570631,
 73      0.5030987588688731,
 74      0.461497712880373
 75    ],
 76    "p_value": 0.0081,
 77    "mde": 0.02367950289717046,
 78    "mde_rel_pct": 135.38481814164533,
 79    "verdict": "idea worse (significant)",
 80    "system_worked": false
 81  },
 82  "custom_track": {
 83    "name": "singular_value_constitutive_energy",
 84    "file": "stage2_bench.py",
 85    "domain": "constitutive mechanics"
 86  },
 87  "idea_sweep": [
 88    {
 89      "cfg": {
 90        "lr": 0.001
 91      },
 92      "mean": 7.160933494567871
 93    },
 94    {
 95      "cfg": {
 96        "lr": 0.003
 97      },
 98      "mean": 2.1708776354789734
 99    },
100    {
101      "cfg": {
102        "lr": 0.006
103      },
104      "mean": 0.4741898328065872
105    }
106  ],
107  "mechanism_signature": {
108    "baseline_jensen_violation": 0.4740000367164612,
109    "idea_jensen_violation": 0.0,
110    "baseline_monotonicity_violation": 0.0,
111    "idea_monotonicity_violation": 0.0,
112    "baseline_lower_violation": 0.18800000846385956,
113    "idea_lower_violation": 1.0,
114    "prediction": "trained nonnegative ICNN should eliminate Jensen and coordinatewise monotonicity violations",
115    "confirmed": true
116  }
117}