1{
  2  "bench_version": 1,
  3  "track": "tangent_evasion_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": 0.2184290774166584
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.20245468616485596
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.01
 27        },
 28        "mean": 0.14241471700370312
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.1416995758190751,
 33      "std": 0.033042414878946644,
 34      "per_seed": [
 35        0.17624007165431976,
 36        0.18836471438407898,
 37        0.0758553221821785,
 38        0.12919875979423523,
 39        0.13717937469482422,
 40        0.14442023634910583,
 41        0.12063135206699371,
 42        0.16170677542686462
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 0.14497848600149155,
 49    "std": 0.025671386028026995,
 50    "per_seed": [
 51      0.17028073966503143,
 52      0.15589244663715363,
 53      0.09865665435791016,
 54      0.1305036097764969,
 55      0.12954431772232056,
 56      0.18738092482089996,
 57      0.15148448944091797,
 58      0.13608470559120178
 59    ],
 60    "n": 8
 61  },
 62  "comparison": {
 63    "delta_mean": 0.003278910182416439,
 64    "idea_wins": 4,
 65    "n_pairs": 8,
 66    "per_seed_diffs": [
 67      -0.00595933198928833,
 68      -0.032472267746925354,
 69      0.02280133217573166,
 70      0.0013048499822616577,
 71      -0.007635056972503662,
 72      0.04296068847179413,
 73      0.030853137373924255,
 74      -0.025622069835662842
 75    ],
 76    "p_value": 0.74865,
 77    "mde": 0.022431633117906365,
 78    "mde_rel_pct": 15.83041656140702,
 79    "verdict": "no significant win",
 80    "system_worked": false
 81  },
 82  "mechanism_signature": {
 83    "predicted": "projection keeps one-step disk penetration at numerical zero",
 84    "observed_baseline_max_penetration": 0.05145234614610672,
 85    "observed_idea_max_penetration": 0.0,
 86    "confirmed": true,
 87    "custom_track": {
 88      "name": "tangent_evasion_dynamics",
 89      "file": "bench_tangent_stage2.py",
 90      "domain": "dynamics"
 91    },
 92    "idea_sweep": [
 93      {
 94        "cfg": {
 95          "lr": 0.001
 96        },
 97        "mean": 0.22645806334912777
 98      },
 99      {
100        "cfg": {
101          "lr": 0.003
102        },
103        "mean": 0.2063459139317274
104      },
105      {
106        "cfg": {
107          "lr": 0.01
108        },
109        "mean": 0.14497848600149155
110      }
111    ]
112  }
113}