PSD Spectral CNN Block / bench_report.json
Mechanism confirmed, baseline not beaten
1{
2 "bench_version": 1,
3 "track": "vision",
4 "model": "cnn_small_psd_matched",
5 "metric_direction": "lower is better",
6 "n_seeds": 8,
7 "baseline": {
8 "best_cfg": {
9 "lr": 0.006,
10 "weight_decay": 0.0
11 },
12 "sweep": [
13 {
14 "cfg": {
15 "lr": 0.001,
16 "weight_decay": 0.0
17 },
18 "mean": 0.7862499952316284
19 },
20 {
21 "cfg": {
22 "lr": 0.003,
23 "weight_decay": 0.0
24 },
25 "mean": 0.7881249636411667
26 },
27 {
28 "cfg": {
29 "lr": 0.006,
30 "weight_decay": 0.0
31 },
32 "mean": 0.7812499850988388
33 }
34 ],
35 "full": {
36 "mean": 0.785937488079071,
37 "std": 0.026188541827503842,
38 "per_seed": [
39 0.7699999809265137,
40 0.7524999976158142,
41 0.7899999618530273,
42 0.8125,
43 0.7425000071525574,
44 0.8149999976158142,
45 0.8100000023841858,
46 0.7949999570846558
47 ],
48 "n": 8
49 }
50 },
51 "idea": {
52 "mean": 0.8028124868869781,
53 "std": 0.02790714067434992,
54 "per_seed": [
55 0.7824999690055847,
56 0.75,
57 0.8125,
58 0.8100000023841858,
59 0.8399999737739563,
60 0.8100000023841858,
61 0.8349999785423279,
62 0.7824999690055847
63 ],
64 "n": 8
65 },
66 "comparison": {
67 "delta_mean": 0.016874998807907104,
68 "idea_wins": 4,
69 "n_pairs": 8,
70 "per_seed_diffs": [
71 0.012499988079071045,
72 -0.002499997615814209,
73 0.022500038146972656,
74 -0.002499997615814209,
75 0.09749996662139893,
76 -0.004999995231628418,
77 0.02499997615814209,
78 -0.012499988079071045
79 ],
80 "p_value": 0.18975,
81 "mde": 0.029493782207685552,
82 "mde_rel_pct": 3.75268805153601,
83 "verdict": "no significant win",
84 "system_worked": false
85 },
86 "idea_sweep": [
87 {
88 "cfg": {
89 "lr": 0.001,
90 "weight_decay": 0.0
91 },
92 "result": {
93 "mean": 0.8043749779462814,
94 "std": 0.028959830204555673,
95 "per_seed": [
96 0.7749999761581421,
97 0.7599999904632568,
98 0.8299999833106995,
99 0.7924999594688416,
100 0.8199999928474426,
101 0.8125,
102 0.8549999594688416,
103 0.7899999618530273
104 ],
105 "n": 8
106 }
107 },
108 {
109 "cfg": {
110 "lr": 0.003,
111 "weight_decay": 0.0
112 },
113 "result": {
114 "mean": 0.8028124868869781,
115 "std": 0.02790714067434992,
116 "per_seed": [
117 0.7824999690055847,
118 0.75,
119 0.8125,
120 0.8100000023841858,
121 0.8399999737739563,
122 0.8100000023841858,
123 0.8349999785423279,
124 0.7824999690055847
125 ],
126 "n": 8
127 }
128 },
129 {
130 "cfg": {
131 "lr": 0.006,
132 "weight_decay": 0.0
133 },
134 "result": {
135 "mean": 0.8074999824166298,
136 "std": 0.02095382139683659,
137 "per_seed": [
138 0.7949999570846558,
139 0.7749999761581421,
140 0.824999988079071,
141 0.8025000095367432,
142 0.8399999737739563,
143 0.8050000071525574,
144 0.8299999833106995,
145 0.7874999642372131
146 ],
147 "n": 8
148 }
149 }
150 ],
151 "track_rationale": "The idea changes a spatial multichannel convolution, so CIFAR-10 vision/cnn_small is structurally matched.",
152 "mechanism_signature": {
153 "predicted_min_eigenvalue": 0.0,
154 "observed_min_eigenvalue": 7.631551631807294e-06,
155 "predicted_nonnegative": true,
156 "observed_nonnegative": true,
157 "confirmed": true
158 }
159}