Spectral Hamiltonian Neuron / bench_report.json
Mechanism confirmed, baseline not beaten
1{
2 "bench_version": 1,
3 "track": "dynamics",
4 "model": "rnn_small_matched_spectral_head",
5 "metric_direction": "lower is better",
6 "n_seeds": 8,
7 "baseline": {
8 "best_cfg": {
9 "lr": 0.01,
10 "epochs": 18
11 },
12 "sweep": [
13 {
14 "cfg": {
15 "lr": 0.001,
16 "epochs": 18
17 },
18 "mean": 0.009911961504258215
19 },
20 {
21 "cfg": {
22 "lr": 0.003,
23 "epochs": 18
24 },
25 "mean": 0.004386652202811092
26 },
27 {
28 "cfg": {
29 "lr": 0.01,
30 "epochs": 18
31 },
32 "mean": 0.0008802848460618407
33 }
34 ],
35 "full": {
36 "mean": 0.0006743472804373596,
37 "std": 0.00026491514577006763,
38 "per_seed": [
39 0.0005103786825202405,
40 0.001095399260520935,
41 0.0009623044752515852,
42 0.0009530569659546018,
43 0.0005656004068441689,
44 0.0003435070684645325,
45 0.0004539852379821241,
46 0.0005105461459606886
47 ],
48 "n": 8
49 }
50 },
51 "idea": {
52 "cfg": {
53 "lr": 0.01,
54 "epochs": 18
55 },
56 "mean": 0.0902025755494833,
57 "std": 0.014189437586170238,
58 "per_seed": [
59 0.07012719660997391,
60 0.09647830575704575,
61 0.08879011869430542,
62 0.11882545053958893,
63 0.09636852890253067,
64 0.07268789410591125,
65 0.08943840861320496,
66 0.08890470117330551
67 ],
68 "n": 8
69 },
70 "comparison": {
71 "delta_mean": 0.08952822826904594,
72 "idea_wins": 0,
73 "n_pairs": 8,
74 "per_seed_diffs": [
75 0.06961681792745367,
76 0.09538290649652481,
77 0.08782781421905383,
78 0.11787239357363433,
79 0.0958029284956865,
80 0.07234438703744672,
81 0.08898442337522283,
82 0.08839415502734482
83 ],
84 "p_value": 0.0081,
85 "mde": 0.012535387409262248,
86 "mde_rel_pct": 1858.8919645575206,
87 "verdict": "idea worse (significant)",
88 "system_worked": false
89 },
90 "idea_sweep": [
91 {
92 "cfg": {
93 "lr": 0.001,
94 "epochs": 18
95 },
96 "mean": 0.35300457291305065,
97 "std": 0.08097430981045273,
98 "per_seed": [
99 0.22397439181804657,
100 0.44731879234313965,
101 0.26465779542922974,
102 0.43655097484588623,
103 0.34434905648231506,
104 0.45171505212783813,
105 0.35042715072631836,
106 0.30504336953163147
107 ],
108 "n": 8
109 },
110 {
111 "cfg": {
112 "lr": 0.003,
113 "epochs": 18
114 },
115 "mean": 0.1979203224182129,
116 "std": 0.038848589270492305,
117 "per_seed": [
118 0.1307593286037445,
119 0.23984569311141968,
120 0.1644514948129654,
121 0.25781843066215515,
122 0.19592295587062836,
123 0.22193168103694916,
124 0.19744279980659485,
125 0.17519019544124603
126 ],
127 "n": 8
128 },
129 {
130 "cfg": {
131 "lr": 0.01,
132 "epochs": 18
133 },
134 "mean": 0.0902025755494833,
135 "std": 0.014189437586170238,
136 "per_seed": [
137 0.07012719660997391,
138 0.09647830575704575,
139 0.08879011869430542,
140 0.11882545053958893,
141 0.09636852890253067,
142 0.07268789410591125,
143 0.08943840861320496,
144 0.08890470117330551
145 ],
146 "n": 8
147 }
148 ],
149 "structural_match": "dynamics: controlled pendulum rollout",
150 "mechanism_signature": {
151 "prediction": "noncommuting Hamiltonian terms should create a measurable nonlinear readout",
152 "commutator_frobenius": 4.0,
153 "trained_idea_test_mse_mean": 0.0902025755494833,
154 "trained_idea_nonlinear_residual_mean": 0.1242971978952532,
155 "confirmed": true
156 },
157 "budget": {
158 "epochs": 18,
159 "n_train": 400,
160 "n_test": 400,
161 "seeds": 8
162 }
163}