Gramian-balanced neural SSM compression / bench_results/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": 0.01
10 },
11 "sweep": [
12 {
13 "cfg": {
14 "lr": 0.001
15 },
16 "mean": 0.005056460213381797
17 },
18 {
19 "cfg": {
20 "lr": 0.003
21 },
22 "mean": 0.0029833840089850128
23 },
24 {
25 "cfg": {
26 "lr": 0.01
27 },
28 "mean": 0.0014822854354861192
29 }
30 ],
31 "full": {
32 "mean": 0.0014822854354861192,
33 "std": 0.0010530387951197904,
34 "per_seed": [
35 0.00411539152264595,
36 0.0017662415048107505,
37 0.0006022244924679399,
38 0.0013981220545247197,
39 0.0012625530362129211,
40 0.0010164444101974368,
41 0.0009094251436181366,
42 0.000787881319411099
43 ],
44 "n": 8
45 }
46 },
47 "idea": {
48 "mean": 0.0021019700725446455,
49 "std": 0.0007477001545769394,
50 "per_seed": [
51 0.0022576870396733284,
52 0.0022740764543414116,
53 0.0024945440236479044,
54 0.0014605424366891384,
55 0.0016900348709896207,
56 0.0007099835784174502,
57 0.0026052906177937984,
58 0.003323601558804512
59 ],
60 "n": 8,
61 "cfg": {
62 "lr": 0.01,
63 "rank": 40
64 },
65 "sweep": [
66 {
67 "cfg": {
68 "lr": 0.01,
69 "rank": 24
70 },
71 "mean": 0.003177800477715209,
72 "std": 0.0010355345041122954,
73 "per_seed": [
74 0.0036525779869407415,
75 0.004798321519047022,
76 0.004517960362136364,
77 0.002177618443965912,
78 0.0021414996590465307,
79 0.0027073805686086416,
80 0.003489549970254302,
81 0.0019374953117221594
82 ],
83 "n": 8
84 },
85 {
86 "cfg": {
87 "lr": 0.01,
88 "rank": 32
89 },
90 "mean": 0.002437932969769463,
91 "std": 0.0011340391219749532,
92 "per_seed": [
93 0.0013505988754332066,
94 0.0022953480947762728,
95 0.002628838177770376,
96 0.002794082509353757,
97 0.0051675527356565,
98 0.001575573580339551,
99 0.0017304965294897556,
100 0.0019609732553362846
101 ],
102 "n": 8
103 },
104 {
105 "cfg": {
106 "lr": 0.01,
107 "rank": 40
108 },
109 "mean": 0.0021019700725446455,
110 "std": 0.0007477001545769394,
111 "per_seed": [
112 0.0022576870396733284,
113 0.0022740764543414116,
114 0.0024945440236479044,
115 0.0014605424366891384,
116 0.0016900348709896207,
117 0.0007099835784174502,
118 0.0026052906177937984,
119 0.003323601558804512
120 ],
121 "n": 8
122 }
123 ]
124 },
125 "comparison": {
126 "delta_mean": 0.0006196846370585263,
127 "idea_wins": 2,
128 "n_pairs": 8,
129 "per_seed_diffs": [
130 -0.001857704482972622,
131 0.0005078349495306611,
132 0.0018923195311799645,
133 6.24203821644187e-05,
134 0.00042748183477669954,
135 -0.0003064608317799866,
136 0.0016958654741756618,
137 0.002535720239393413
138 ],
139 "p_value": 0.20345,
140 "mde": 0.0011741999248702343,
141 "mde_rel_pct": 79.21550713241356,
142 "verdict": "no significant win",
143 "system_worked": false,
144 "paired_delta": [
145 -0.001857704482972622,
146 0.0005078349495306611,
147 0.0018923195311799645,
148 6.24203821644187e-05,
149 0.00042748183477669954,
150 -0.0003064608317799866,
151 0.0016958654741756618,
152 0.002535720239393413
153 ]
154 },
155 "mechanism_signature": {
156 "predicted_state_cost_ratio": 0.625,
157 "observed_state_cost_ratio": 0.625,
158 "baseline_hidden": 64,
159 "idea_hidden": 40,
160 "baseline_mean_pred_observed_corr": 0.9994533416434941,
161 "idea_mean_pred_observed_corr": 0.9989892089422736,
162 "confirmed": true
163 }
164}