Finite-horizon Lyapunov regularization for neural updates / bench_report.json
Failed on benchmark
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": 0.01
10 },
11 "sweep": [
12 {
13 "cfg": {
14 "lr": 0.001
15 },
16 "mean": 0.015694387606345117
17 },
18 {
19 "cfg": {
20 "lr": 0.003
21 },
22 "mean": 0.0023102034465409815
23 },
24 {
25 "cfg": {
26 "lr": 0.01
27 },
28 "mean": 0.0013842678163200617
29 }
30 ],
31 "full": {
32 "mean": 0.0015652802903787233,
33 "std": 0.0007422459790989128,
34 "per_seed": [
35 0.0013847877271473408,
36 0.0011940097901970148,
37 0.0015944147016853094,
38 0.0013638590462505817,
39 0.0008487981976941228,
40 0.000910585920792073,
41 0.001885351026430726,
42 0.0033404359128326178
43 ],
44 "n": 8
45 }
46 },
47 "idea": {
48 "mean": 0.0015269540235749446,
49 "std": 0.000787764827387549,
50 "per_seed": [
51 0.0010921545326709747,
52 0.0012567200465127826,
53 0.0015678050694987178,
54 0.0012287304271012545,
55 0.0008476584334857762,
56 0.0009015121031552553,
57 0.0018893856322392821,
58 0.0034316659439355135
59 ],
60 "n": 8
61 },
62 "comparison": {
63 "delta_mean": -3.832626680377871e-05,
64 "idea_wins": 5,
65 "n_pairs": 8,
66 "per_seed_diffs": [
67 -0.00029263319447636604,
68 6.271025631576777e-05,
69 -2.6609632186591625e-05,
70 -0.00013512861914932728,
71 -1.1397642083466053e-06,
72 -9.073817636817694e-06,
73 4.03460580855608e-06,
74 9.123003110289574e-05
75 ],
76 "p_value": 0.46705,
77 "mde": 0.00010250800277371827,
78 "mde_rel_pct": 6.548859230120136,
79 "verdict": "no significant win",
80 "system_worked": false
81 },
82 "mechanism_signature": {
83 "prediction": "finite-horizon gradient energy should contract on average and unstable training should have positive residual violations",
84 "trained_model_measurements": {
85 "idea_cfg": {
86 "lr": 0.01
87 },
88 "idea_signature_per_seed": [
89 {
90 "violation_rate": 0.02631578947368421,
91 "mean_v_ratio": 2.168702542242643,
92 "tail_grad_std": 0.46059960229461117,
93 "tail_loss_std": 0.11873548939891664
94 },
95 {
96 "violation_rate": 0.05263157894736842,
97 "mean_v_ratio": 3.6939305728301406,
98 "tail_grad_std": 0.4575504708189517,
99 "tail_loss_std": 0.1556151981825947
100 },
101 {
102 "violation_rate": 0.02631578947368421,
103 "mean_v_ratio": 2.2387936332900273,
104 "tail_grad_std": 0.3685155160785214,
105 "tail_loss_std": 0.14080336151872447
106 },
107 {
108 "violation_rate": 0.02631578947368421,
109 "mean_v_ratio": 1.2345037685805245,
110 "tail_grad_std": 0.38392368145194206,
111 "tail_loss_std": 0.14072047870995807
112 },
113 {
114 "violation_rate": 0.05263157894736842,
115 "mean_v_ratio": 0.8230216467851087,
116 "tail_grad_std": 0.36693779977631946,
117 "tail_loss_std": 0.10801473858501799
118 },
119 {
120 "violation_rate": 0.02631578947368421,
121 "mean_v_ratio": 1.8151202782204277,
122 "tail_grad_std": 0.44894377114615497,
123 "tail_loss_std": 0.16700399421982076
124 },
125 {
126 "violation_rate": 0.05263157894736842,
127 "mean_v_ratio": 1.0862176747698533,
128 "tail_grad_std": 0.3875248868001556,
129 "tail_loss_std": 0.14996397260121175
130 },
131 {
132 "violation_rate": 0.05263157894736842,
133 "mean_v_ratio": 1.5587420168479806,
134 "tail_grad_std": 0.40622251364286266,
135 "tail_loss_std": 0.1526901696628486
136 }
137 ]
138 },
139 "confirmed": false,
140 "idea_sweep": [
141 {
142 "cfg": {
143 "lr": 0.001
144 },
145 "mean": 0.01564739688183181
146 },
147 {
148 "cfg": {
149 "lr": 0.003
150 },
151 "mean": 0.002292555815074593
152 },
153 {
154 "cfg": {
155 "lr": 0.01
156 },
157 "mean": 0.0012863525189459324
158 }
159 ],
160 "certificate": {
161 "M": 4,
162 "alpha": 0.1,
163 "lambda": 0.002,
164 "beta": 0.9
165 },
166 "observed_mean_v_ratio": 1.8273790166958381,
167 "observed_violation_rate": 0.039473684210526314
168 }
169}