Histogram-Controlled Cluster Updates for Iterative GNNs / 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 "epochs": 12
11 },
12 "sweep": [
13 {
14 "cfg": {
15 "lr": 0.001,
16 "epochs": 12
17 },
18 "mean": 0.023774019442498684
19 },
20 {
21 "cfg": {
22 "lr": 0.003,
23 "epochs": 12
24 },
25 "mean": 0.01013653608970344
26 },
27 {
28 "cfg": {
29 "lr": 0.01,
30 "epochs": 12
31 },
32 "mean": 0.003741882392205298
33 }
34 ],
35 "full": {
36 "mean": 0.003835578914731741,
37 "std": 0.0008182271774829689,
38 "per_seed": [
39 0.004058382008224726,
40 0.003425257047638297,
41 0.00307759759016335,
42 0.004406292922794819,
43 0.004498522728681564,
44 0.0026040049269795418,
45 0.005265520885586739,
46 0.003349053207784891
47 ],
48 "n": 8
49 }
50 },
51 "idea": {
52 "mean": 0.006128221459221095,
53 "std": 0.0019430310597602655,
54 "per_seed": [
55 0.005886178929358721,
56 0.003636509645730257,
57 0.005781157873570919,
58 0.010280938819050789,
59 0.007521466817706823,
60 0.0044521004892885685,
61 0.006602167151868343,
62 0.004865251947194338
63 ],
64 "n": 8
65 },
66 "comparison": {
67 "delta_mean": 0.002292642544489354,
68 "idea_wins": 0,
69 "n_pairs": 8,
70 "per_seed_diffs": [
71 0.001827796921133995,
72 0.00021125259809195995,
73 0.002703560283407569,
74 0.00587464589625597,
75 0.003022944089025259,
76 0.0018480955623090267,
77 0.0013366462662816048,
78 0.0015161987394094467
79 ],
80 "p_value": 0.0081,
81 "mde": 0.0014061178480241382,
82 "mde_rel_pct": 36.65985967916089,
83 "verdict": "idea worse (significant)",
84 "system_worked": false
85 },
86 "mechanism_signature": {
87 "track_structure": "controlled pendulum multi-step dynamics; recurrent state updates",
88 "scheduler": "largest high-bin hidden-state disagreement histogram",
89 "clusters": 4,
90 "histogram_bins": 4,
91 "signature": {
92 "prediction": "trained-model synchronous-vs-sequential discrepancy scales approximately eta^2",
93 "rows": [
94 {
95 "eta": 0.2,
96 "observed_linf": 0.023520752787590027,
97 "eta2": 0.04000000000000001
98 },
99 {
100 "eta": 0.4,
101 "observed_linf": 0.09223496913909912,
102 "eta2": 0.16000000000000003
103 },
104 {
105 "eta": 0.6,
106 "observed_linf": 0.20175543427467346,
107 "eta2": 0.36
108 },
109 {
110 "eta": 0.8,
111 "observed_linf": 0.3461129665374756,
112 "eta2": 0.6400000000000001
113 }
114 ],
115 "eta2_ratio_range": [
116 0.5408015102148055,
117 0.5880188196897506
118 ],
119 "confirmed": true
120 }
121 },
122 "runtime_sec": 430.6290466785431,
123 "protocol_notes": {
124 "paired_seeds": [
125 0,
126 1,
127 2,
128 3,
129 4,
130 5,
131 6,
132 7
133 ],
134 "dataset_sizes": [
135 400,
136 100
137 ],
138 "idea_grid": [
139 {
140 "cfg": {
141 "lr": 0.001,
142 "epochs": 12
143 },
144 "mean": 0.2794509418308735
145 },
146 {
147 "cfg": {
148 "lr": 0.003,
149 "epochs": 12
150 },
151 "mean": 0.012458630255423486
152 },
153 {
154 "cfg": {
155 "lr": 0.01,
156 "epochs": 12
157 },
158 "mean": 0.006128221459221095
159 }
160 ],
161 "selected_idea_cfg": {
162 "lr": 0.01,
163 "epochs": 12
164 },
165 "baseline_grid_union": [
166 {
167 "lr": 0.001,
168 "epochs": 12
169 },
170 {
171 "lr": 0.003,
172 "epochs": 12
173 },
174 {
175 "lr": 0.01,
176 "epochs": 12
177 }
178 ]
179 }
180}