Mpemba Mode-Filtered Training / bench_report.json

Mechanism confirmed, baseline not beaten

Raw ⬇ ZIP
  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      "weight_decay": 0.0001
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.0015,
 16          "weight_decay": 0.0
 17        },
 18        "mean": 0.0016401863540522754
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.0015,
 23          "weight_decay": 0.0001
 24        },
 25        "mean": 0.0016229975590249524
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.003,
 30          "weight_decay": 0.0
 31        },
 32        "mean": 0.00105226032610517
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.003,
 37          "weight_decay": 0.0001
 38        },
 39        "mean": 0.0010288678167853504
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.006,
 44          "weight_decay": 0.0
 45        },
 46        "mean": 0.0007705322786932811
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.006,
 51          "weight_decay": 0.0001
 52        },
 53        "mean": 0.0007483311783289537
 54      }
 55    ],
 56    "full": {
 57      "mean": 0.0007712390506640077,
 58      "std": 0.00015769195775709773,
 59      "per_seed": [
 60        0.0006971143302507699,
 61        0.0008527055033482611,
 62        0.0005140206776559353,
 63        0.0009294842020608485,
 64        0.0005480053368955851,
 65        0.0008613452082499862,
 66        0.0008049127063713968,
 67        0.0009623244404792786
 68      ],
 69      "n": 8
 70    }
 71  },
 72  "idea": {
 73    "mean": 0.000772874424001202,
 74    "std": 0.00016036009165282255,
 75    "per_seed": [
 76      0.0006930093513801694,
 77      0.0008406945853494108,
 78      0.0005222903564572334,
 79      0.0009326299768872559,
 80      0.0005455962382256985,
 81      0.0008512879139743745,
 82      0.0008067093440331519,
 83      0.0009907776257023215
 84    ],
 85    "n": 8,
 86    "best_cfg": {
 87      "lr": 0.006,
 88      "weight_decay": 0.0001,
 89      "gamma": 0.01
 90    },
 91    "sweep": [
 92      {
 93        "cfg": {
 94          "lr": 0.0015,
 95          "weight_decay": 0.0001,
 96          "gamma": 0.01
 97        },
 98        "mean": 0.0017785793825169094,
 99        "per_seed": [
100          0.0009525532950647175,
101          0.0017182347364723682,
102          0.0013132664607837796,
103          0.0025045962538570166,
104          0.0009531762916594744,
105          0.001337766065262258,
106          0.002284065820276737,
107          0.0031649761367589235
108        ]
109      },
110      {
111        "cfg": {
112          "lr": 0.0015,
113          "weight_decay": 0.0001,
114          "gamma": 0.03
115        },
116        "mean": 0.001787718734703958,
117        "per_seed": [
118          0.0009378251270391047,
119          0.001730192918330431,
120          0.001301152864471078,
121          0.0025130226276814938,
122          0.0009530456154607236,
123          0.001333152875304222,
124          0.002290173200890422,
125          0.0032431846484541893
126        ]
127      },
128      {
129        "cfg": {
130          "lr": 0.0015,
131          "weight_decay": 0.0001,
132          "gamma": 0.1
133        },
134        "mean": 0.0018141339023713954,
135        "per_seed": [
136          0.0009042767342180014,
137          0.001764905173331499,
138          0.0012808581814169884,
139          0.0025383830070495605,
140          0.0009604471852071583,
141          0.0013194901403039694,
142          0.002280787331983447,
143          0.003463923465460539
144        ]
145      },
146      {
147        "cfg": {
148          "lr": 0.003,
149          "weight_decay": 0.0001,
150          "gamma": 0.01
151        },
152        "mean": 0.0010871518170461059,
153        "per_seed": [
154          0.0009137207525782287,
155          0.0008784217061474919,
156          0.0006754152127541602,
157          0.0016324227908626199,
158          0.0006958629237487912,
159          0.0008044865680858493,
160          0.0014184740139171481,
161          0.0016784105682745576
162        ]
163      },
164      {
165        "cfg": {
166          "lr": 0.003,
167          "weight_decay": 0.0001,
168          "gamma": 0.03
169        },
170        "mean": 0.001080300826288294,
171        "per_seed": [
172          0.0009089682716876268,
173          0.0008737270254641771,
174          0.0006606929819099605,
175          0.0016298916889354587,
176          0.00069375493330881,
177          0.0008071218035183847,
178          0.0014060401590541005,
179          0.001662209746427834
180        ]
181      },
182      {
183        "cfg": {
184          "lr": 0.003,
185          "weight_decay": 0.0001,
186          "gamma": 0.1
187        },
188        "mean": 0.0010688186739571393,
189        "per_seed": [
190          0.0008881261455826461,
191          0.0008660596213303506,
192          0.0006356937228702009,
193          0.001619623857550323,
194          0.0006789837498217821,
195          0.0008229009690694511,
196          0.0013761218870058656,
197          0.0016630394384264946
198        ]
199      },
200      {
201        "cfg": {
202          "lr": 0.006,
203          "weight_decay": 0.0001,
204          "gamma": 0.01
205        },
206        "mean": 0.000772874424001202,
207        "per_seed": [
208          0.0006930093513801694,
209          0.0008406945853494108,
210          0.0005222903564572334,
211          0.0009326299768872559,
212          0.0005455962382256985,
213          0.0008512879139743745,
214          0.0008067093440331519,
215          0.0009907776257023215
216        ]
217      },
218      {
219        "cfg": {
220          "lr": 0.006,
221          "weight_decay": 0.0001,
222          "gamma": 0.03
223        },
224        "mean": 0.0007740218861727044,
225        "per_seed": [
226          0.0006858397973701358,
227          0.0008134624222293496,
228          0.0005332473665475845,
229          0.0009384911973029375,
230          0.000545005255844444,
231          0.0008184867911040783,
232          0.0008121238206513226,
233          0.0010455184383317828
234        ]
235      },
236      {
237        "cfg": {
238          "lr": 0.006,
239          "weight_decay": 0.0001,
240          "gamma": 0.1
241        },
242        "mean": 0.0007736754487268627,
243        "per_seed": [
244          0.0006801989511586726,
245          0.000721059856005013,
246          0.0005168036441318691,
247          0.0009618915501050651,
248          0.0006032949313521385,
249          0.0007013163412921131,
250          0.000832107849419117,
251          0.001172730466350913
252        ]
253      }
254    ]
255  },
256  "comparison": {
257    "delta_mean": 1.6353733371943235e-06,
258    "idea_wins": 4,
259    "n_pairs": 8,
260    "per_seed_diffs": [
261      -4.104978870600462e-06,
262      -1.2010917998850346e-05,
263      8.269678801298141e-06,
264      3.1457748264074326e-06,
265      -2.409098669886589e-06,
266      -1.0057294275611639e-05,
267      1.796637661755085e-06,
268      2.8453185223042965e-05
269    ],
270    "p_value": 0.79715,
271    "mde": 1.066015482771443e-05,
272    "mde_rel_pct": 1.382211497010744,
273    "verdict": "no measurable effect",
274    "system_worked": false
275  },
276  "mechanism_signature": {
277    "track_match": "dynamics: controlled pendulum rollout; odd sign-reversal channel",
278    "prediction": "post-probe control suppresses symmetry-breaking mode, so late A2 proxy falls below probe A2 proxy",
279    "predicted_vs_observed": {
280      "predicted_late_to_probe_ratio": "< 1.0",
281      "observed_mean_ratio": 0.16168932957134385,
282      "per_seed": [
283        {
284          "seed": 0,
285          "baseline_A2_proxy": 0.0066139898262917995,
286          "idea_probe_A2_proxy": 0.028515702579170465,
287          "idea_late_A2_proxy": 0.007312557194381952
288        },
289        {
290          "seed": 1,
291          "baseline_A2_proxy": 0.00954736303538084,
292          "idea_probe_A2_proxy": 0.06434168418248494,
293          "idea_late_A2_proxy": 0.013520444432894388
294        },
295        {
296          "seed": 2,
297          "baseline_A2_proxy": 0.009730289690196514,
298          "idea_probe_A2_proxy": 0.12297303229570389,
299          "idea_late_A2_proxy": 0.011106709018349648
300        },
301        {
302          "seed": 3,
303          "baseline_A2_proxy": 0.005885358899831772,
304          "idea_probe_A2_proxy": 0.02378938440233469,
305          "idea_late_A2_proxy": 0.006580424960702658
306        },
307        {
308          "seed": 4,
309          "baseline_A2_proxy": 0.004799993708729744,
310          "idea_probe_A2_proxy": 0.03924484054247538,
311          "idea_late_A2_proxy": 0.0038459034791837135
312        },
313        {
314          "seed": 5,
315          "baseline_A2_proxy": 0.014292827807366848,
316          "idea_probe_A2_proxy": 0.0702164676040411,
317          "idea_late_A2_proxy": 0.00915219308808446
318        },
319        {
320          "seed": 6,
321          "baseline_A2_proxy": 0.0031815164256840944,
322          "idea_probe_A2_proxy": 0.11305895199378331,
323          "idea_late_A2_proxy": 0.006548060104250908
324        },
325        {
326          "seed": 7,
327          "baseline_A2_proxy": 0.013455699197947979,
328          "idea_probe_A2_proxy": 0.07279480372865994,
329          "idea_late_A2_proxy": 0.012648270465433598
330        }
331      ]
332    },
333    "confirmed": true
334  }
335}