Symplectic Recurrent Block / 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": 0.005
10 },
11 "sweep": [
12 {
13 "cfg": {
14 "lr": 0.001
15 },
16 "mean": 0.004298779065720737
17 },
18 {
19 "cfg": {
20 "lr": 0.003
21 },
22 "mean": 0.0034615940821822733
23 },
24 {
25 "cfg": {
26 "lr": 0.005
27 },
28 "mean": 0.0015488484350498766
29 }
30 ],
31 "full": {
32 "mean": 0.0015987288497854024,
33 "std": 0.0004723450985826372,
34 "per_seed": [
35 0.0010614258935675025,
36 0.0021564248017966747,
37 0.001086847041733563,
38 0.0018906960031017661,
39 0.0020614659879356623,
40 0.0012517774011939764,
41 0.0011406542034819722,
42 0.002140539465472102
43 ],
44 "n": 8
45 }
46 },
47 "idea": {
48 "mean": 0.006309549498837441,
49 "std": 0.0013831181007421675,
50 "per_seed": [
51 0.004766013007611036,
52 0.0075876968912780285,
53 0.00614228006452322,
54 0.005290781147778034,
55 0.005994749721139669,
56 0.0061586033552885056,
57 0.009300675243139267,
58 0.005235596559941769
59 ],
60 "n": 8
61 },
62 "comparison": {
63 "delta_mean": 0.004710820649052039,
64 "idea_wins": 0,
65 "n_pairs": 8,
66 "per_seed_diffs": [
67 0.003704587114043534,
68 0.005431272089481354,
69 0.005055433022789657,
70 0.003400085144676268,
71 0.003933283733204007,
72 0.004906825954094529,
73 0.008160021039657295,
74 0.0030950570944696665
75 ],
76 "p_value": 0.0081,
77 "mde": 0.0013600218273413377,
78 "mde_rel_pct": 85.06894884169343,
79 "verdict": "idea worse (significant)",
80 "system_worked": false
81 },
82 "idea_sweep": [
83 {
84 "cfg": {
85 "lr": 0.005,
86 "step": 0.04
87 },
88 "result": {
89 "mean": 0.006823674135375768,
90 "std": 0.0018729123134178251,
91 "per_seed": [
92 0.005168338771909475,
93 0.010784340091049671,
94 0.0066774641163647175,
95 0.005866168532520533,
96 0.0061936150304973125,
97 0.005813745316118002,
98 0.008938142098486423,
99 0.005147579126060009
100 ],
101 "n": 8
102 }
103 },
104 {
105 "cfg": {
106 "lr": 0.005,
107 "step": 0.08
108 },
109 "result": {
110 "mean": 0.006543989293277264,
111 "std": 0.00158231784982001,
112 "per_seed": [
113 0.005307216662913561,
114 0.009000664576888084,
115 0.0062515996396541595,
116 0.0055503034964203835,
117 0.005550468806177378,
118 0.006075830198824406,
119 0.00943143293261528,
120 0.005184398032724857
121 ],
122 "n": 8
123 }
124 },
125 {
126 "cfg": {
127 "lr": 0.005,
128 "step": 0.12
129 },
130 "result": {
131 "mean": 0.006309549498837441,
132 "std": 0.0013831181007421675,
133 "per_seed": [
134 0.004766013007611036,
135 0.0075876968912780285,
136 0.00614228006452322,
137 0.005290781147778034,
138 0.005994749721139669,
139 0.0061586033552885056,
140 0.009300675243139267,
141 0.005235596559941769
142 ],
143 "n": 8
144 }
145 }
146 ],
147 "audit": {
148 "n_train": 400,
149 "n_test": 200,
150 "epochs": 12,
151 "matched_architecture_role": "GRU replaced by symplectic hidden transition",
152 "baseline_lr_union": [
153 0.001,
154 0.003,
155 0.005
156 ]
157 },
158 "mechanism_signature": {
159 "prediction": "symplectic transition det should remain near 1 (log|det| near 0)",
160 "observed_idea_log_abs_det": 0.00387675897218287,
161 "observed_baseline_log_abs_det": -42.42223358154297,
162 "absolute_logdet_error_idea": 0.00387675897218287,
163 "absolute_logdet_error_baseline": 42.42223358154297,
164 "confirmed": true
165 }
166}