Complete Log-Barrier Natural Gradient / bench_report.json
Mechanism confirmed, baseline not beaten
1{
2 "bench_version": 1,
3 "track": "tabular",
4 "model": "mlp_tiny",
5 "metric_direction": "lower is better",
6 "n_seeds": 8,
7 "baseline": {
8 "sweep": {
9 "best_cfg": {
10 "lr": 0.009
11 },
12 "sweep": [
13 {
14 "cfg": {
15 "lr": 0.001
16 },
17 "mean": 15.798457145690918
18 },
19 {
20 "cfg": {
21 "lr": 0.003
22 },
23 "mean": 9.297731161117554
24 },
25 {
26 "cfg": {
27 "lr": 0.009
28 },
29 "mean": 6.843065023422241
30 }
31 ],
32 "full": {
33 "mean": 6.911551356315613,
34 "std": 0.25102731009725415,
35 "per_seed": [
36 6.589434623718262,
37 6.794177055358887,
38 7.121844291687012,
39 6.866804122924805,
40 6.845246315002441,
41 7.470037460327148,
42 6.813384056091309,
43 6.791482925415039
44 ],
45 "n": 8
46 }
47 },
48 "best_config": {
49 "lr": 0.009,
50 "epochs": 15
51 },
52 "full": {
53 "mean": 6.911551356315613,
54 "std": 0.25102731009725415,
55 "per_seed": [
56 6.589434623718262,
57 6.794177055358887,
58 7.121844291687012,
59 6.866804122924805,
60 6.845246315002441,
61 7.470037460327148,
62 6.813384056091309,
63 6.791482925415039
64 ],
65 "n": 8
66 }
67 },
68 "idea": {
69 "config": {
70 "lr": 0.009,
71 "epochs": 15,
72 "R": 2.5,
73 "block": "first_layer_bias",
74 "metric": "mse"
75 },
76 "per_seed": [
77 6.712672233581543,
78 6.749656677246094,
79 6.953096389770508,
80 6.8154120445251465,
81 6.669680118560791,
82 7.3650336265563965,
83 6.754833221435547,
84 6.995800971984863
85 ],
86 "tuning_sweep": [
87 {
88 "lr": 0.001,
89 "mean_mse": 15.960896253585815,
90 "per_seed": [
91 15.142996788024902,
92 15.989051818847656,
93 16.95168685913086,
94 15.759849548339844
95 ]
96 },
97 {
98 "lr": 0.003,
99 "mean_mse": 8.806230664253235,
100 "per_seed": [
101 7.58334493637085,
102 9.324054718017578,
103 9.102656364440918,
104 9.214866638183594
105 ]
106 },
107 {
108 "lr": 0.009,
109 "mean_mse": 6.807709336280823,
110 "per_seed": [
111 6.712672233581543,
112 6.749656677246094,
113 6.953096389770508,
114 6.8154120445251465
115 ]
116 }
117 ],
118 "mean": 6.877023160457611,
119 "std": 0.2130142060373052,
120 "n": 8
121 },
122 "comparison": {
123 "delta_mean": -0.03452819585800171,
124 "idea_wins": 6,
125 "n_pairs": 8,
126 "per_seed_diffs": [
127 0.12323760986328125,
128 -0.04452037811279297,
129 -0.1687479019165039,
130 -0.0513920783996582,
131 -0.1755661964416504,
132 -0.10500383377075195,
133 -0.05855083465576172,
134 0.20431804656982422
135 ],
136 "p_value": 0.46775,
137 "mde": 0.11208041708387258,
138 "mde_rel_pct": 1.6216390692294593,
139 "verdict": "no measurable effect",
140 "system_worked": false
141 },
142 "math_check": {
143 "max_intrinsic_grad_g_squared": 0.9999989999995,
144 "max_bound_violation": 0.0,
145 "max_radial_formula_error": 3.3306690738754696e-16,
146 "step_inverse_metric_ratio_r0.9999": 0.0006689383838711185
147 },
148 "selection_note": "baseline and idea share lr union; baseline tuned on seeds 0-3; full comparison uses paired seeds 0-7",
149 "runtime_sec": 30.761850833892822,
150 "mechanism_signature": {
151 "prediction": "inverse barrier metric decreases toward boundary; bounded intrinsic barrier gradient",
152 "predicted_at_center": 3.125,
153 "observed_mean_inverse_metric_at_trained_points": 1.705785185098648,
154 "trained_model_points": [
155 {
156 "seed": 0,
157 "observed_mean_inverse_metric": 1.6924855709075928,
158 "final_q": 0.5548654556274414
159 },
160 {
161 "seed": 1,
162 "observed_mean_inverse_metric": 1.7231124639511108,
163 "final_q": 0.5643916320800781
164 },
165 {
166 "seed": 2,
167 "observed_mean_inverse_metric": 1.6797361373901367,
168 "final_q": 0.550905647277832
169 },
170 {
171 "seed": 3,
172 "observed_mean_inverse_metric": 1.7380695343017578,
173 "final_q": 0.5688753509521485
174 },
175 {
176 "seed": 4,
177 "observed_mean_inverse_metric": 1.864469289779663,
178 "final_q": 0.6083690643310546
179 },
180 {
181 "seed": 5,
182 "observed_mean_inverse_metric": 1.7799270153045654,
183 "final_q": 0.5821066665649415
184 },
185 {
186 "seed": 6,
187 "observed_mean_inverse_metric": 1.812181830406189,
188 "final_q": 0.592091178894043
189 },
190 {
191 "seed": 7,
192 "observed_mean_inverse_metric": 1.356299638748169,
193 "final_q": 0.44984680175781255
194 }
195 ],
196 "confirmed": true
197 }
198}