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": 0.006,
 10      "score_weight": 0.0
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "score_weight": 0.0
 17        },
 18        "mean": 0.004586851398926228
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "score_weight": 0.0
 24        },
 25        "mean": 0.0030439750116784126
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.006,
 30          "score_weight": 0.0
 31        },
 32        "mean": 0.0019343624007888138
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.0018362020200584084,
 37      "std": 0.0005092446787647265,
 38      "per_seed": [
 39        0.0017339912010356784,
 40        0.0023557941894978285,
 41        0.0013064503436908126,
 42        0.002341213868930936,
 43        0.001163388486020267,
 44        0.001496884971857071,
 45        0.0026453696191310883,
 46        0.0016465234803035855
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.0018397023231955245,
 53    "std": 0.0005083828899465245,
 54    "per_seed": [
 55      0.0017356511671096087,
 56      0.0023582950234413147,
 57      0.001307925907894969,
 58      0.0023415545001626015,
 59      0.0011693481355905533,
 60      0.0015009039780125022,
 61      0.0026489796582609415,
 62      0.0016549602150917053
 63    ],
 64    "n": 8
 65  },
 66  "comparison": {
 67    "delta_mean": 3.500303137116134e-06,
 68    "idea_wins": 0,
 69    "n_pairs": 8,
 70    "per_seed_diffs": [
 71      1.6599660739302635e-06,
 72      2.5008339434862137e-06,
 73      1.4755642041563988e-06,
 74      3.4063123166561127e-07,
 75      5.959649570286274e-06,
 76      4.019006155431271e-06,
 77      3.6100391298532486e-06,
 78      8.436734788119793e-06
 79    ],
 80    "p_value": 0.0081,
 81    "mde": 2.2160150645813298e-06,
 82    "mde_rel_pct": 0.12068470900118276,
 83    "verdict": "idea worse (significant)",
 84    "system_worked": false
 85  },
 86  "mechanism_signature": {
 87    "idea_cfg": {
 88      "lr": 0.006,
 89      "score_weight": 0.03
 90    },
 91    "signature": {
 92      "prediction": "probability-flow consistency reduces prediction sensitivity to score transport",
 93      "baseline_abs_sensitivity_mean": 0.0339321275241673,
 94      "idea_abs_sensitivity_mean": 0.033821433782577515,
 95      "ratio_idea_over_baseline": 0.9967377895032464,
 96      "n_models": 8,
 97      "confirmed": true
 98    }
 99  },
100  "runtime_seconds": 87.39407682418823,
101  "idea_sweep": [
102    {
103      "cfg": {
104        "lr": 0.006,
105        "score_weight": 0.03
106      },
107      "result": {
108        "mean": 0.0018397023231955245,
109        "std": 0.0005083828899465245,
110        "per_seed": [
111          0.0017356511671096087,
112          0.0023582950234413147,
113          0.001307925907894969,
114          0.0023415545001626015,
115          0.0011693481355905533,
116          0.0015009039780125022,
117          0.0026489796582609415,
118          0.0016549602150917053
119        ],
120        "n": 8
121      }
122    },
123    {
124      "cfg": {
125        "lr": 0.006,
126        "score_weight": 0.1
127      },
128      "result": {
129        "mean": 0.0018478815909475088,
130        "std": 0.0005064124207020823,
131        "per_seed": [
132          0.0017395142931491137,
133          0.002364138374105096,
134          0.001311446656472981,
135          0.0023424141108989716,
136          0.001183195854537189,
137          0.001510326866991818,
138          0.002657373435795307,
139          0.0016746431356295943
140        ],
141        "n": 8
142      }
143    },
144    {
145      "cfg": {
146        "lr": 0.001,
147        "score_weight": 0.03
148      },
149      "result": {
150        "mean": 0.005814504460431635,
151        "std": 0.0022258700624946614,
152        "per_seed": [
153          0.0073998901061713696,
154          0.0020209946669638157,
155          0.003580172546207905,
156          0.005346550140529871,
157          0.005660634022206068,
158          0.006507787387818098,
159          0.0060682338662445545,
160          0.009931772947311401
161        ],
162        "n": 8
163      }
164    },
165    {
166      "cfg": {
167        "lr": 0.003,
168        "score_weight": 0.03
169      },
170      "result": {
171        "mean": 0.0029023701208643615,
172        "std": 0.0012422328212200683,
173        "per_seed": [
174          0.0010790829546749592,
175          0.003418009728193283,
176          0.0023018938954919577,
177          0.005378886125981808,
178          0.002690472174435854,
179          0.0023395069874823093,
180          0.003974617458879948,
181          0.0020364916417747736
182        ],
183        "n": 8
184      }
185    }
186  ],
187  "protocol_notes": "Dynamics chosen because controlled pendulum rollout is explicitly a stability/control task. Both systems use rnn_small, same datasets, epochs and Adam; only the self-supervised score-transport consistency loss differs."
188}