Collision-aware physical-support abstention / bench_report.json

Failed on benchmark

Raw ⬇ ZIP
  1{
  2  "bench_version": 1,
  3  "track": "collision_support",
  4  "model": "mlp_tiny",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.003,
 10      "epochs": 15,
 11      "batch": 128
 12    },
 13    "sweep": [
 14      {
 15        "cfg": {
 16          "lr": 0.001,
 17          "epochs": 15,
 18          "batch": 128
 19        },
 20        "mean": 0.05312499962747097
 21      },
 22      {
 23        "cfg": {
 24          "lr": 0.003,
 25          "epochs": 15,
 26          "batch": 128
 27        },
 28        "mean": 0.018124999478459358
 29      },
 30      {
 31        "cfg": {
 32          "lr": 0.01,
 33          "epochs": 15,
 34          "batch": 128
 35        },
 36        "mean": 0.019374999683350325
 37      }
 38    ],
 39    "full": {
 40      "mean": 0.01874999958090484,
 41      "std": 0.0045069391227432615,
 42      "per_seed": [
 43        0.014999999664723873,
 44        0.022499999031424522,
 45        0.019999999552965164,
 46        0.014999999664723873,
 47        0.027499999850988388,
 48        0.017500000074505806,
 49        0.01249999925494194,
 50        0.019999999552965164
 51      ],
 52      "n": 8
 53    }
 54  },
 55  "idea": {
 56    "mean": 0.017499999434221536,
 57    "std": 0.005590169558486883,
 58    "per_seed": [
 59      0.017500000074505806,
 60      0.019999999552965164,
 61      0.014999999664723873,
 62      0.02499999850988388,
 63      0.017500000074505806,
 64      0.02499999850988388,
 65      0.007499999832361937,
 66      0.01249999925494194
 67    ],
 68    "n": 8
 69  },
 70  "comparison": {
 71    "delta_mean": -0.0012500001466833055,
 72    "idea_wins": 5,
 73    "n_pairs": 8,
 74    "per_seed_diffs": [
 75      0.002500000409781933,
 76      -0.002499999478459358,
 77      -0.004999999888241291,
 78      0.009999998845160007,
 79      -0.009999999776482582,
 80      0.007499998435378075,
 81      -0.004999999422580004,
 82      -0.007500000298023224
 83    ],
 84    "p_value": 0.57515,
 85    "mde": 0.006017153669723515,
 86    "mde_rel_pct": 32.09148695582604,
 87    "verdict": "no significant win",
 88    "system_worked": false
 89  },
 90  "mechanism_signature": {
 91    "math_check": {
 92      "predicted_exponents": {
 93        "I_S_vs_s": 2.0,
 94        "I_D_vs_s": 6.0
 95      },
 96      "observed_exponents": {
 97        "I_S_vs_s": 2.000000000000001,
 98        "I_D_vs_s": 6.000000000000005
 99      },
100      "gate_at_bench": {
101        "I_S": 1.3061224489795922,
102        "I_D": 0.025600000000000008,
103        "prediction": "parent_block_with_child_ambiguity"
104      }
105    },
106    "trained_behavior": {
107      "baseline": {
108        "mean_prediction_change": 0.02209057251457125,
109        "mean_parent_accuracy": 0.9812499657273293
110      },
111      "idea": {
112        "mean_prediction_change": 0.0,
113        "mean_parent_accuracy": 0.9828124567866325
114      },
115      "prediction": "low I_D implies child identity is not exposed; sign/permutation invariant parent output",
116      "observed_child_identity_sensitivity": 0.0,
117      "confirmed": true
118    },
119    "custom_track": {
120      "name": "collision_support",
121      "file": "collision_track.py",
122      "domain": "embedding"
123    },
124    "idea_sweep": [
125      {
126        "cfg": {
127          "lr": 0.001,
128          "epochs": 15,
129          "batch": 128
130        },
131        "result": {
132          "mean": 0.02281249954830855,
133          "std": 0.005510997348379881,
134          "per_seed": [
135            0.027499999850988388,
136            0.029999999329447746,
137            0.017500000074505806,
138            0.029999999329447746,
139            0.022499999031424522,
140            0.017500000074505806,
141            0.014999999664723873,
142            0.022499999031424522
143          ],
144          "n": 8
145        }
146      },
147      {
148        "cfg": {
149          "lr": 0.003,
150          "epochs": 15,
151          "batch": 128
152        },
153        "result": {
154          "mean": 0.019062499632127583,
155          "std": 0.006724198598013715,
156          "per_seed": [
157            0.019999999552965164,
158            0.03500000014901161,
159            0.014999999664723873,
160            0.019999999552965164,
161            0.017500000074505806,
162            0.019999999552965164,
163            0.01249999925494194,
164            0.01249999925494194
165          ],
166          "n": 8
167        }
168      },
169      {
170        "cfg": {
171          "lr": 0.01,
172          "epochs": 15,
173          "batch": 128
174        },
175        "result": {
176          "mean": 0.017499999434221536,
177          "std": 0.005590169558486883,
178          "per_seed": [
179            0.017500000074505806,
180            0.019999999552965164,
181            0.014999999664723873,
182            0.02499999850988388,
183            0.017500000074505806,
184            0.02499999850988388,
185            0.007499999832361937,
186            0.01249999925494194
187          ],
188          "n": 8
189        }
190      }
191    ]
192  },
193  "math_check": {
194    "predicted_exponents": {
195      "I_S_vs_s": 2.0,
196      "I_D_vs_s": 6.0
197    },
198    "observed_exponents": {
199      "I_S_vs_s": 2.000000000000001,
200      "I_D_vs_s": 6.000000000000005
201    },
202    "gate_at_bench": {
203      "I_S": 1.3061224489795922,
204      "I_D": 0.025600000000000008,
205      "prediction": "parent_block_with_child_ambiguity"
206    }
207  }
208}