Fourier-Mode Stability Shaping / bench_report.json
Beats tuned baseline
1{
2 "bench_version": 1,
3 "track": "dynamics",
4 "model": "cyclic_rnn",
5 "metric_direction": "lower is better",
6 "n_seeds": 8,
7 "baseline": {
8 "best_cfg": {
9 "lr": 0.01,
10 "weight_decay": 0.0
11 },
12 "sweep": [
13 {
14 "cfg": {
15 "lr": 0.001,
16 "weight_decay": 0.0
17 },
18 "mean": 0.073138820938766
19 },
20 {
21 "cfg": {
22 "lr": 0.003,
23 "weight_decay": 0.0
24 },
25 "mean": 0.01337075512856245
26 },
27 {
28 "cfg": {
29 "lr": 0.01,
30 "weight_decay": 0.0
31 },
32 "mean": 0.004462164011783898
33 }
34 ],
35 "full": {
36 "mean": 0.004087189445272088,
37 "std": 0.0010007874399045344,
38 "per_seed": [
39 0.005726089235395193,
40 0.0033913287334144115,
41 0.004283653572201729,
42 0.004447584506124258,
43 0.0049942112527787685,
44 0.0028478815220296383,
45 0.002599592786282301,
46 0.004407173953950405
47 ],
48 "n": 8
49 }
50 },
51 "idea": {
52 "mean": 0.004080623359186575,
53 "std": 0.0010004881538153208,
54 "per_seed": [
55 0.005722619593143463,
56 0.003379821078851819,
57 0.004280869383364916,
58 0.004445591010153294,
59 0.004983154591172934,
60 0.0028435608837753534,
61 0.0025958793703466654,
62 0.0043934909626841545
63 ],
64 "n": 8
65 },
66 "comparison": {
67 "delta_mean": -6.566086085513234e-06,
68 "idea_wins": 8,
69 "n_pairs": 8,
70 "per_seed_diffs": [
71 -3.469642251729965e-06,
72 -1.1507654562592506e-05,
73 -2.784188836812973e-06,
74 -1.9934959709644318e-06,
75 -1.1056661605834961e-05,
76 -4.320638254284859e-06,
77 -3.7134159356355667e-06,
78 -1.368299126625061e-05
79 ],
80 "p_value": 0.0081,
81 "mde": 3.91188445758878e-06,
82 "mde_rel_pct": 0.0957108670877467,
83 "verdict": "idea better (significant)",
84 "system_worked": true
85 },
86 "mechanism_signature": {
87 "note": "trained-model zero-input Fourier perturbation; q=0",
88 "signature": {
89 "q": 0,
90 "kappa": 0.15,
91 "rows": [
92 {
93 "system": "baseline",
94 "weights": [
95 1.1289324760437012,
96 1.1723648309707642,
97 1.1778165102005005,
98 1.2012455463409424
99 ],
100 "modes": [
101 {
102 "mode": 1,
103 "predicted": -0.09919014415880456,
104 "observed": -0.04295965945097118
105 },
106 {
107 "mode": 2,
108 "predicted": -0.3536465652791994,
109 "observed": -0.17734158165416888
110 },
111 {
112 "mode": 3,
113 "predicted": -0.6558346787734254,
114 "observed": -0.42319134910019884
115 }
116 ]
117 },
118 {
119 "system": "idea",
120 "weights": [
121 1.1295150518417358,
122 1.173497200012207,
123 1.1802256107330322,
124 1.20441734790802
125 ],
126 "modes": [
127 {
128 "mode": 1,
129 "predicted": -0.09940500360509097,
130 "observed": -0.04295041987070078
131 },
132 {
133 "mode": 2,
134 "predicted": -0.3544018135383647,
135 "observed": -0.17730516519047812
136 },
137 {
138 "mode": 3,
139 "predicted": -0.6571983150619246,
140 "observed": -0.42311798056693856
141 }
142 ]
143 }
144 ],
145 "max_relative_error": 0.5679249704438308,
146 "confirmed": false,
147 "task_metric_seed0": {
148 "baseline": 0.005726089235395193,
149 "idea": 0.005722619593143463
150 }
151 }
152 },
153 "idea_sweep": [
154 {
155 "cfg": {
156 "lr": 0.001,
157 "weight_decay": 0.0
158 },
159 "mean": 0.07314319256693125
160 },
161 {
162 "cfg": {
163 "lr": 0.003,
164 "weight_decay": 0.0
165 },
166 "mean": 0.013368312735110521
167 },
168 {
169 "cfg": {
170 "lr": 0.01,
171 "weight_decay": 0.0
172 },
173 "mean": 0.004457225266378373
174 }
175 ],
176 "protocol": {
177 "paired_seeds": [
178 0,
179 1,
180 2,
181 3,
182 4,
183 5,
184 6,
185 7
186 ],
187 "train_samples": 400,
188 "test_samples": 200,
189 "epochs": 18,
190 "architecture": "same CyclicRNN; barrier only differs"
191 }
192}