PDE Sinkhorn with asymmetric geometric boundaries / bench_report.json
Mechanism confirmed, baseline not beaten
1{
2 "bench_version": 1,
3 "track": "anisotropic_oblique_bvp",
4 "model": "mlp_tiny",
5 "metric_direction": "lower is better",
6 "n_seeds": 8,
7 "baseline": {
8 "best_cfg": {
9 "lr": 0.003,
10 "weight_decay": 0.0
11 },
12 "sweep": [
13 {
14 "cfg": {
15 "lr": 0.001,
16 "weight_decay": 0.0
17 },
18 "mean": 0.00012391583004500717
19 },
20 {
21 "cfg": {
22 "lr": 0.003,
23 "weight_decay": 0.0
24 },
25 "mean": 0.00010377542184869526
26 },
27 {
28 "cfg": {
29 "lr": 0.006,
30 "weight_decay": 0.0
31 },
32 "mean": 0.00011149384499731241
33 }
34 ],
35 "full": {
36 "mean": 9.435732454221579e-05,
37 "std": 2.4013090964758765e-05,
38 "per_seed": [
39 0.00012657140905503184,
40 0.00012852151121478528,
41 5.6132030294975266e-05,
42 0.00010387673682998866,
43 8.006240386748686e-05,
44 9.653119195718318e-05,
45 9.42463957471773e-05,
46 6.891691737109795e-05
47 ],
48 "n": 8
49 }
50 },
51 "idea": {
52 "mean": 0.00011699842434609309,
53 "std": 3.095430559427717e-05,
54 "per_seed": [
55 0.00016582579701207578,
56 0.0001437541941413656,
57 6.513581320177764e-05,
58 9.932620014296845e-05,
59 0.00011056653602281585,
60 0.0001362798793707043,
61 0.00012959061132278293,
62 8.550836355425417e-05
63 ],
64 "n": 8,
65 "cfg": {
66 "lr": 0.003,
67 "weight_decay": 0.0,
68 "bc_weight": 0.1
69 },
70 "selected_cfg": {
71 "lr": 0.003,
72 "weight_decay": 0.0,
73 "bc_weight": 0.1
74 }
75 },
76 "comparison": {
77 "delta_mean": 2.2641099803877296e-05,
78 "idea_wins": 1,
79 "n_pairs": 8,
80 "per_seed_diffs": [
81 3.9254387957043946e-05,
82 1.523268292658031e-05,
83 9.003782906802371e-06,
84 -4.550536687020212e-06,
85 3.050413215532899e-05,
86 3.974868741352111e-05,
87 3.534421557560563e-05,
88 1.6591446183156222e-05
89 ],
90 "p_value": 0.0152,
91 "mde": 1.3440697763296005e-05,
92 "mde_rel_pct": 14.244466795243426,
93 "verdict": "idea worse (significant)",
94 "system_worked": false
95 },
96 "mechanism_signature": {
97 "mechanism_signature": {
98 "prediction": "boundary-aware PDE training reduces measured oblique/no-flux residuals",
99 "mean_baseline_oblique_residual": 0.02720322646200657,
100 "mean_idea_oblique_residual": 0.01003198220860213,
101 "mean_baseline_normal_residual": 0.04276570677757263,
102 "mean_idea_normal_residual": 0.008119827311020344,
103 "per_seed": [
104 {
105 "seed": 0,
106 "baseline_oblique_abs_derivative": 0.03964276611804962,
107 "idea_oblique_abs_derivative": 0.01283207256346941,
108 "baseline_normal_abs_derivative": 0.03705023229122162,
109 "idea_normal_abs_derivative": 0.009148340672254562
110 },
111 {
112 "seed": 1,
113 "baseline_oblique_abs_derivative": 0.020765259861946106,
114 "idea_oblique_abs_derivative": 0.008115937002003193,
115 "baseline_normal_abs_derivative": 0.04100044071674347,
116 "idea_normal_abs_derivative": 0.010081736370921135
117 },
118 {
119 "seed": 2,
120 "baseline_oblique_abs_derivative": 0.020734591409564018,
121 "idea_oblique_abs_derivative": 0.009465154260396957,
122 "baseline_normal_abs_derivative": 0.06341361999511719,
123 "idea_normal_abs_derivative": 0.01047296728938818
124 },
125 {
126 "seed": 3,
127 "baseline_oblique_abs_derivative": 0.023547939956188202,
128 "idea_oblique_abs_derivative": 0.009573541581630707,
129 "baseline_normal_abs_derivative": 0.027527321130037308,
130 "idea_normal_abs_derivative": 0.00741744227707386
131 },
132 {
133 "seed": 4,
134 "baseline_oblique_abs_derivative": 0.02001587674021721,
135 "idea_oblique_abs_derivative": 0.009027393534779549,
136 "baseline_normal_abs_derivative": 0.04191877692937851,
137 "idea_normal_abs_derivative": 0.008082550019025803
138 },
139 {
140 "seed": 5,
141 "baseline_oblique_abs_derivative": 0.026504037901759148,
142 "idea_oblique_abs_derivative": 0.007130592130124569,
143 "baseline_normal_abs_derivative": 0.031208854168653488,
144 "idea_normal_abs_derivative": 0.007085439749062061
145 },
146 {
147 "seed": 6,
148 "baseline_oblique_abs_derivative": 0.0324188694357872,
149 "idea_oblique_abs_derivative": 0.016259243711829185,
150 "baseline_normal_abs_derivative": 0.03935849666595459,
151 "idea_normal_abs_derivative": 0.007526469882577658
152 },
153 {
154 "seed": 7,
155 "baseline_oblique_abs_derivative": 0.033996470272541046,
156 "idea_oblique_abs_derivative": 0.007851922884583473,
157 "baseline_normal_abs_derivative": 0.060647912323474884,
158 "idea_normal_abs_derivative": 0.005143672227859497
159 }
160 ],
161 "confirmed": true
162 },
163 "custom_track": {
164 "name": "anisotropic_oblique_bvp",
165 "file": "pde_boundary_track.py",
166 "domain": "pde"
167 },
168 "idea_sweep": [
169 {
170 "cfg": {
171 "lr": 0.003,
172 "weight_decay": 0.0,
173 "bc_weight": 0.1
174 },
175 "mean": 0.00011699842434609309,
176 "std": 3.095430559427717e-05,
177 "per_seed": [
178 0.00016582579701207578,
179 0.0001437541941413656,
180 6.513581320177764e-05,
181 9.932620014296845e-05,
182 0.00011056653602281585,
183 0.0001362798793707043,
184 0.00012959061132278293,
185 8.550836355425417e-05
186 ]
187 },
188 {
189 "cfg": {
190 "lr": 0.003,
191 "weight_decay": 0.0,
192 "bc_weight": 1.0
193 },
194 "mean": 0.00016314282402163371,
195 "std": 5.0713334733415886e-05,
196 "per_seed": [
197 0.00013824015331920236,
198 0.0002775485045276582,
199 9.841256542131305e-05,
200 0.00013486476382240653,
201 0.00015113064728211612,
202 0.00019398487347643822,
203 0.00017369228589814156,
204 0.00013726879842579365
205 ]
206 },
207 {
208 "cfg": {
209 "lr": 0.003,
210 "weight_decay": 0.0,
211 "bc_weight": 5.0
212 },
213 "mean": 0.00022120710127637722,
214 "std": 5.5719673809869554e-05,
215 "per_seed": [
216 0.00020971496996935457,
217 0.0002884801651816815,
218 0.00013389925879891962,
219 0.0002538147964514792,
220 0.00018812052439898252,
221 0.00022245440050028265,
222 0.0003070295206271112,
223 0.00016614317428320646
224 ]
225 }
226 ]
227 },
228 "protocol_note": "Custom PDE track required: no built-in track contains boundary-value/PDE structure."
229}