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      "M": 1
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "M": 1
 17        },
 18        "mean": 0.001921573159052059
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "M": 1
 24        },
 25        "mean": 0.0011179699067724869
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.01,
 30          "M": 1
 31        },
 32        "mean": 0.0006299250526353717
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.0005546553857129766,
 37      "std": 0.00021396768993230646,
 38      "per_seed": [
 39        0.0004971042508259416,
 40        0.0004999673110432923,
 41        0.00046853424282744527,
 42        0.0010540944058448076,
 43        0.0005468997405841947,
 44        0.00023900401720311493,
 45        0.0006055250996723771,
 46        0.0005261140177026391
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.0005546553857129766,
 53    "std": 0.00021396768993230646,
 54    "per_seed": [
 55      0.0004971042508259416,
 56      0.0004999673110432923,
 57      0.00046853424282744527,
 58      0.0010540944058448076,
 59      0.0005468997405841947,
 60      0.00023900401720311493,
 61      0.0006055250996723771,
 62      0.0005261140177026391
 63    ],
 64    "n": 8,
 65    "sweep": [
 66      {
 67        "cfg": {
 68          "lr": 0.001,
 69          "M": 1
 70        },
 71        "mean": 0.001921573159052059
 72      },
 73      {
 74        "cfg": {
 75          "lr": 0.003,
 76          "M": 1
 77        },
 78        "mean": 0.0011179699067724869
 79      },
 80      {
 81        "cfg": {
 82          "lr": 0.01,
 83          "M": 1
 84        },
 85        "mean": 0.0006299250526353717
 86      }
 87    ]
 88  },
 89  "comparison": {
 90    "delta_mean": 0.0,
 91    "idea_wins": 0,
 92    "n_pairs": 8,
 93    "per_seed_diffs": [
 94      0.0,
 95      0.0,
 96      0.0,
 97      0.0,
 98      0.0,
 99      0.0,
100      0.0,
101      0.0
102    ],
103    "p_value": 1.0,
104    "mde": 0.0,
105    "mde_rel_pct": 0.0,
106    "verdict": "no measurable effect",
107    "system_worked": false
108  },
109  "mechanism_signature": {
110    "prediction": "Lyapunov certificate filters sampled updates; rejected proposals should have positive certificate and accepted proposals nonpositive.",
111    "trained_model_measurements": [
112      {
113        "accepted": 72,
114        "rejected": 0,
115        "checks": 72,
116        "violations": 0,
117        "violation_rate": 0.0,
118        "mean_certificate": -0.49112362704343265
119      },
120      {
121        "accepted": 72,
122        "rejected": 0,
123        "checks": 72,
124        "violations": 0,
125        "violation_rate": 0.0,
126        "mean_certificate": -0.4969692817992634
127      },
128      {
129        "accepted": 72,
130        "rejected": 0,
131        "checks": 72,
132        "violations": 0,
133        "violation_rate": 0.0,
134        "mean_certificate": -0.5120618219176928
135      },
136      {
137        "accepted": 72,
138        "rejected": 0,
139        "checks": 72,
140        "violations": 0,
141        "violation_rate": 0.0,
142        "mean_certificate": -0.5117074669235282
143      },
144      {
145        "accepted": 72,
146        "rejected": 0,
147        "checks": 72,
148        "violations": 0,
149        "violation_rate": 0.0,
150        "mean_certificate": -0.5226765928996934
151      },
152      {
153        "accepted": 72,
154        "rejected": 0,
155        "checks": 72,
156        "violations": 0,
157        "violation_rate": 0.0,
158        "mean_certificate": -0.48182762745353913
159      },
160      {
161        "accepted": 72,
162        "rejected": 0,
163        "checks": 72,
164        "violations": 0,
165        "violation_rate": 0.0,
166        "mean_certificate": -0.4946398569477929
167      },
168      {
169        "accepted": 72,
170        "rejected": 0,
171        "checks": 72,
172        "violations": 0,
173        "violation_rate": 0.0,
174        "mean_certificate": -0.508300371054146
175      }
176    ],
177    "confirmed": false
178  },
179  "protocol_notes": "Baseline and idea share rnn_small, data, epochs, batch, Adam, and lr/M grid; only certificate rejection differs."
180}