Co-Prime Virtual-Aperture Attention / bench_report.json

Mechanism confirmed, baseline not beaten

Raw ⬇ ZIP
  1{
  2  "bench_version": 1,
  3  "track": "sequence",
  4  "model": "transformer_tiny",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.003
 10    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 0.1494479700922966
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.14718072488904
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.01
 27        },
 28        "mean": 0.23090055957436562
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.14619511179625988,
 33      "std": 0.023272144211599243,
 34      "per_seed": [
 35        0.19329272210597992,
 36        0.12494795769453049,
 37        0.12228920310735703,
 38        0.14819301664829254,
 39        0.1282258927822113,
 40        0.1327003538608551,
 41        0.1498585343360901,
 42        0.17005321383476257
 43      ],
 44      "n": 8
 45    },
 46    "idea_lr_selected": 0.003
 47  },
 48  "idea": {
 49    "mean": 0.1398473186418414,
 50    "std": 0.011513116013381047,
 51    "per_seed": [
 52      0.15645107626914978,
 53      0.13615846633911133,
 54      0.14752109348773956,
 55      0.13047337532043457,
 56      0.1331927627325058,
 57      0.12439116090536118,
 58      0.13342459499835968,
 59      0.1571660190820694
 60    ],
 61    "n": 8
 62  },
 63  "comparison": {
 64    "delta_mean": -0.0063477931544184685,
 65    "idea_wins": 5,
 66    "n_pairs": 8,
 67    "per_seed_diffs": [
 68      -0.03684164583683014,
 69      0.011210508644580841,
 70      0.025231890380382538,
 71      -0.01771964132785797,
 72      0.004966869950294495,
 73      -0.008309192955493927,
 74      -0.016433939337730408,
 75      -0.012887194752693176
 76    ],
 77    "p_value": 0.3787,
 78    "mde": 0.016241711340640855,
 79    "mde_rel_pct": 11.109613133492175,
 80    "verdict": "no significant win",
 81    "system_worked": false
 82  },
 83  "math_check": {
 84    "M1": 3,
 85    "M2": 4,
 86    "gcd": 1,
 87    "physical_offsets": [
 88      0,
 89      3,
 90      4,
 91      6,
 92      8,
 93      9
 94    ],
 95    "physical_count": 6,
 96    "physical_formula": 6,
 97    "virtual_offsets": [
 98      0,
 99      3,
100      4,
101      6,
102      7,
103      8,
104      9,
105      10,
106      11,
107      13,
108      14,
109      17
110    ],
111    "virtual_count": 12,
112    "graph_reachable": [
113      0,
114      3,
115      4,
116      6,
117      7,
118      8,
119      9,
120      10,
121      11,
122      13,
123      14,
124      17
125    ],
126    "noncoprime_2_4_virtual_count": 6,
127    "claim_holds": true
128  },
129  "idea_sweep": [
130    {
131      "lr": 0.001,
132      "mean": 0.1506526917219162,
133      "std": 0.011124472883780578
134    },
135    {
136      "lr": 0.003,
137      "mean": 0.1398473186418414,
138      "std": 0.011513116013381047
139    },
140    {
141      "lr": 0.01,
142      "mean": 0.3220752514898777,
143      "std": 0.1023727956412049
144    }
145  ],
146  "protocol_note": "Baseline and idea use paired sequence datasets and identical non-attention architecture.",
147  "mechanism_signature": {
148    "predicted_virtual_offsets": [
149      0,
150      3,
151      4,
152      6,
153      7,
154      8,
155      9,
156      10,
157      11,
158      13,
159      14,
160      17
161    ],
162    "observed_nonzero_input_offsets": [
163      0,
164      1,
165      2,
166      3,
167      4,
168      5,
169      6,
170      7,
171      8,
172      9,
173      10,
174      11,
175      12,
176      13,
177      14,
178      15,
179      16,
180      17,
181      18,
182      19,
183      20,
184      21,
185      22,
186      23,
187      24,
188      25,
189      26,
190      27,
191      28,
192      29,
193      30,
194      31
195    ],
196    "predicted_offsets_observed": [
197      0,
198      3,
199      4,
200      6,
201      7,
202      8,
203      9,
204      10,
205      11,
206      13,
207      14,
208      17
209    ],
210    "coverage": 1.0,
211    "confirmed": true
212  }
213}