Response-from-Hessian Regularizer / bench_report.json
Failed on benchmark
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": 0.004298779065720737
17 },
18 {
19 "cfg": {
20 "lr": 0.003
21 },
22 "mean": 0.0034615940821822733
23 },
24 {
25 "cfg": {
26 "lr": 0.01
27 },
28 "mean": 0.0014865802077110857
29 }
30 ],
31 "full": {
32 "mean": 0.0014198302378645167,
33 "std": 0.00039839114901047844,
34 "per_seed": [
35 0.0009327387670055032,
36 0.0018641706556081772,
37 0.0010785290505737066,
38 0.0020708823576569557,
39 0.0009260809747502208,
40 0.0014055331703275442,
41 0.0016443432541564107,
42 0.001436363672837615
43 ],
44 "n": 8
45 }
46 },
47 "idea": {
48 "mean": 0.0012014269523206167,
49 "std": 0.0003833906235964023,
50 "per_seed": [
51 0.0005308947875164449,
52 0.00110043294262141,
53 0.0015973668778315187,
54 0.0017763767391443253,
55 0.0008540595881640911,
56 0.001374695566482842,
57 0.000991623499430716,
58 0.0013859656173735857
59 ],
60 "n": 8
61 },
62 "comparison": {
63 "delta_mean": -0.00021840328554389998,
64 "idea_wins": 7,
65 "n_pairs": 8,
66 "per_seed_diffs": [
67 -0.00040184397948905826,
68 -0.0007637377129867673,
69 0.000518837827257812,
70 -0.00029450561851263046,
71 -7.202138658612967e-05,
72 -3.0837603844702244e-05,
73 -0.0006527197547256947,
74 -5.039805546402931e-05
75 ],
76 "p_value": 0.17895,
77 "mde": 0.00033947390618403674,
78 "mde_rel_pct": 23.90947150798954,
79 "verdict": "no significant win",
80 "system_worked": false
81 },
82 "mechanism_signature": {
83 "idea_config": {
84 "lr": 0.01
85 },
86 "idea_sweep": [
87 {
88 "cfg": {
89 "lr": 0.001
90 },
91 "result": {
92 "mean": 0.004815780295757577,
93 "std": 0.0014414284088542476,
94 "per_seed": [
95 0.007190660573542118,
96 0.0025714822113513947,
97 0.0030915497336536646,
98 0.005778642371296883,
99 0.005150299984961748,
100 0.0042128548957407475,
101 0.004502150695770979,
102 0.00602860189974308
103 ],
104 "n": 8
105 }
106 },
107 {
108 "cfg": {
109 "lr": 0.003
110 },
111 "result": {
112 "mean": 0.0031467917578993365,
113 "std": 0.0019972773089779808,
114 "per_seed": [
115 0.003412304213270545,
116 0.002708759158849716,
117 0.001344470540061593,
118 0.008165908977389336,
119 0.0023987311869859695,
120 0.002598112216219306,
121 0.0016315100947394967,
122 0.00291453767567873
123 ],
124 "n": 8
125 }
126 },
127 {
128 "cfg": {
129 "lr": 0.01
130 },
131 "result": {
132 "mean": 0.0012014269523206167,
133 "std": 0.0003833906235964023,
134 "per_seed": [
135 0.0005308947875164449,
136 0.00110043294262141,
137 0.0015973668778315187,
138 0.0017763767391443253,
139 0.0008540595881640911,
140 0.001374695566482842,
141 0.000991623499430716,
142 0.0013859656173735857
143 ],
144 "n": 8
145 }
146 }
147 ],
148 "prediction": "Hessian response magnitude is finite and tracks curvature scale",
149 "rayleigh_q": 5.0501101213740185e-05,
150 "observed_hvp_norm": 0.17713873088359833,
151 "predicted_curvature_scale": 5.0501101213740185e-05,
152 "gradient_norm": 1.5206302404403687,
153 "observed_to_predicted_ratio": 3506.9267275331645,
154 "confirmed": false
155 }
156}