Jacobian-Ranked Simplex Features / bench_report_registered.json
Beats tuned baseline
1{
2 "bench_version": 1,
3 "track": "jacobian_simplex_triangle",
4 "model": "mlp_tiny",
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": 0.02803073776885867
17 },
18 {
19 "cfg": {
20 "lr": 0.003
21 },
22 "mean": 0.014160370687022805
23 },
24 {
25 "cfg": {
26 "lr": 0.01
27 },
28 "mean": 0.01129886822309345
29 }
30 ],
31 "full": {
32 "mean": 0.010346994968131185,
33 "std": 0.002795628226111498,
34 "per_seed": [
35 0.011259164661169052,
36 0.013376608490943909,
37 0.012956458143889904,
38 0.0076032415963709354,
39 0.011506251059472561,
40 0.006055573001503944,
41 0.013024956919252872,
42 0.006993705872446299
43 ],
44 "n": 8
45 }
46 },
47 "idea": {
48 "mean": 0.006576284009497613,
49 "std": 0.0011506745983477893,
50 "per_seed": [
51 0.005950353108346462,
52 0.007979393936693668,
53 0.007719683926552534,
54 0.005885654129087925,
55 0.005792700685560703,
56 0.00467478483915329,
57 0.008065154775977135,
58 0.006542546674609184
59 ],
60 "n": 8
61 },
62 "comparison": {
63 "delta_mean": -0.003770710958633572,
64 "idea_wins": 8,
65 "n_pairs": 8,
66 "per_seed_diffs": [
67 -0.00530881155282259,
68 -0.00539721455425024,
69 -0.00523677421733737,
70 -0.0017175874672830105,
71 -0.005713550373911858,
72 -0.0013807881623506546,
73 -0.004959802143275738,
74 -0.00045115919783711433
75 ],
76 "p_value": 0.0081,
77 "mde": 0.0018236052438909767,
78 "mde_rel_pct": 17.624491453873258,
79 "verdict": "idea better (significant)",
80 "system_worked": true
81 },
82 "custom_track": {
83 "name": "jacobian_simplex_triangle",
84 "file": "simplex_track.py",
85 "domain": "geometric_graph"
86 },
87 "math_check": {
88 "max_abs_autodiff_finite_difference": 3.032483153475596e-10,
89 "jacobian_norm": 1.0437423273887312,
90 "positive_sigma": true
91 },
92 "idea_sweep": [
93 {
94 "cfg": {
95 "lr": 0.001
96 },
97 "result": {
98 "mean": 0.027510648826137185,
99 "std": 0.010645224655473157,
100 "per_seed": [
101 0.01384858600795269,
102 0.04354790970683098,
103 0.04501708969473839,
104 0.027162743732333183,
105 0.025257226079702377,
106 0.021136121824383736,
107 0.02710472047328949,
108 0.017010793089866638
109 ],
110 "n": 8
111 }
112 },
113 {
114 "cfg": {
115 "lr": 0.003
116 },
117 "result": {
118 "mean": 0.010628473362885416,
119 "std": 0.0050250987460242836,
120 "per_seed": [
121 0.006476277951151133,
122 0.016148243099451065,
123 0.020230555906891823,
124 0.007130751386284828,
125 0.00772904371842742,
126 0.007939266972243786,
127 0.013715370558202267,
128 0.005658277310431004
129 ],
130 "n": 8
131 }
132 },
133 {
134 "cfg": {
135 "lr": 0.01
136 },
137 "result": {
138 "mean": 0.006576284009497613,
139 "std": 0.0011506745983477893,
140 "per_seed": [
141 0.005950353108346462,
142 0.007979393936693668,
143 0.007719683926552534,
144 0.005885654129087925,
145 0.005792700685560703,
146 0.00467478483915329,
147 0.008065154775977135,
148 0.006542546674609184
149 ],
150 "n": 8
151 }
152 }
153 ],
154 "epochs": 25,
155 "n_train": 400,
156 "n_test": 400,
157 "parameter_parity": true,
158 "mechanism_signature": {
159 "model_test_metric_seed0": 0.006476277951151133,
160 "mean_predicted_abs_output_delta": 0.00010509323328733444,
161 "mean_observed_abs_output_delta": 0.00010509914136491716,
162 "mean_observed_to_predicted_ratio": 1.002894401550293,
163 "finite_difference_agreement": 7.362902465501975e-08,
164 "confirmed": true
165 }
166}