Phase-Delay Spectral Margin for Attractor RNNs / bench_report.json
Mechanism confirmed, baseline not beaten
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": 0.0
11 },
12 "sweep": [
13 {
14 "cfg": {
15 "lr": 0.0015,
16 "weight_decay": 0.0
17 },
18 "mean": 0.0009279211371904239
19 },
20 {
21 "cfg": {
22 "lr": 0.003,
23 "weight_decay": 0.0
24 },
25 "mean": 0.0005964032097836025
26 },
27 {
28 "cfg": {
29 "lr": 0.006,
30 "weight_decay": 0.0
31 },
32 "mean": 0.0003292812507424969
33 }
34 ],
35 "full": {
36 "mean": 0.000254957890319929,
37 "std": 0.00010644342536032148,
38 "per_seed": [
39 0.00032660356373526156,
40 0.00046600779751315713,
41 0.00021259563800413162,
42 0.0003119180037174374,
43 0.00017991565982811153,
44 0.0002721373166423291,
45 0.0001093827813747339,
46 0.00016110236174426973
47 ],
48 "n": 8
49 }
50 },
51 "idea": {
52 "mean": 0.00025879375971271656,
53 "std": 6.582941024778054e-05,
54 "per_seed": [
55 0.00030997759313322604,
56 0.0002871895849239081,
57 0.00023464152764063329,
58 0.0003278508083894849,
59 0.00023518447414971888,
60 0.0002398034994257614,
61 0.00011388212442398071,
62 0.0003218204656150192
63 ],
64 "n": 8
65 },
66 "comparison": {
67 "delta_mean": 3.835869392787572e-06,
68 "idea_wins": 3,
69 "n_pairs": 8,
70 "per_seed_diffs": [
71 -1.6625970602035522e-05,
72 -0.00017881821258924901,
73 2.204588963650167e-05,
74 1.5932804672047496e-05,
75 5.526881432160735e-05,
76 -3.2333817216567695e-05,
77 4.499343049246818e-06,
78 0.00016071810387074947
79 ],
80 "p_value": 0.84895,
81 "mde": 7.921439496717511e-05,
82 "mde_rel_pct": 31.069599323941087,
83 "verdict": "no significant win",
84 "system_worked": false
85 },
86 "idea_sweep": [
87 {
88 "cfg": {
89 "lr": 0.0015,
90 "weight_decay": 0.0
91 },
92 "mean": 0.0009607886240701191,
93 "std": 0.00023662483911888167,
94 "per_seed": [
95 0.0006990795955061913,
96 0.001059096772223711,
97 0.000746679084841162,
98 0.0012032907688990235,
99 0.0006439350545406342,
100 0.0009082331089302897,
101 0.0010811877436935902,
102 0.001344806863926351
103 ],
104 "n": 8
105 },
106 {
107 "cfg": {
108 "lr": 0.003,
109 "weight_decay": 0.0
110 },
111 "mean": 0.0005813516982016154,
112 "std": 0.00011166105013956393,
113 "per_seed": [
114 0.0004897246835753322,
115 0.0006705834530293941,
116 0.000494384381454438,
117 0.0007663700962439179,
118 0.0004182068514637649,
119 0.0005766836693510413,
120 0.0005388451390899718,
121 0.0006960153114050627
122 ],
123 "n": 8
124 },
125 {
126 "cfg": {
127 "lr": 0.006,
128 "weight_decay": 0.0
129 },
130 "mean": 0.00025879375971271656,
131 "std": 6.582941024778054e-05,
132 "per_seed": [
133 0.00030997759313322604,
134 0.0002871895849239081,
135 0.00023464152764063329,
136 0.0003278508083894849,
137 0.00023518447414971888,
138 0.0002398034994257614,
139 0.00011388212442398071,
140 0.0003218204656150192
141 ],
142 "n": 8
143 }
144 ],
145 "permutation_pvalue": 0.84895,
146 "mechanism_signature": {
147 "predicted_boundary_amplification": {
148 "0.8": 0.9999999999999996,
149 "1.0": 0.9999999999999999,
150 "1.2": 1.3999999999999986
151 },
152 "observed_trained_model_mean_min_real_eigenvalue": 3.3919921815395355,
153 "observed_trained_model_mean_max_euler_amplification": 0.6608031913638115,
154 "confirmed": true
155 }
156}