Overshoot Budget Controller / 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": 0.03
10 },
11 "sweep": [
12 {
13 "cfg": {
14 "lr": 0.001
15 },
16 "mean": 0.0013811723329126835
17 },
18 {
19 "cfg": {
20 "lr": 0.01
21 },
22 "mean": 0.00036846242801402695
23 },
24 {
25 "cfg": {
26 "lr": 0.03
27 },
28 "mean": 0.0001452554606657941
29 }
30 ],
31 "full": {
32 "mean": 0.00016686083745298674,
33 "std": 0.0001211998237425883,
34 "per_seed": [
35 0.000324782362440601,
36 5.0498259952291846e-05,
37 6.899042637087405e-05,
38 0.00013675079389940947,
39 0.00021192267013248056,
40 0.0003892787790391594,
41 9.448564378544688e-05,
42 5.817776400363073e-05
43 ],
44 "n": 8
45 }
46 },
47 "idea": {
48 "best_cfg": {
49 "lr": 0.03
50 },
51 "sweep": [
52 {
53 "cfg": {
54 "lr": 0.001
55 },
56 "mean": 0.0015905930049484596,
57 "per_seed": [
58 0.001139691099524498,
59 0.0013835665304213762,
60 0.0010758476564660668,
61 0.001925584045238793,
62 0.0012607797980308533,
63 0.001312664826400578,
64 0.0020039849914610386,
65 0.0026226250920444727
66 ]
67 },
68 {
69 "cfg": {
70 "lr": 0.01
71 },
72 "mean": 0.00038149026113387663,
73 "per_seed": [
74 0.00029065465787425637,
75 0.0005194233381189406,
76 0.0002394697949057445,
77 0.00042430192115716636,
78 0.00026041336241178215,
79 0.00043081643525511026,
80 0.0004192772030364722,
81 0.0004675653763115406
82 ]
83 },
84 {
85 "cfg": {
86 "lr": 0.03
87 },
88 "mean": 0.00015556820017081918,
89 "per_seed": [
90 0.0003302668046671897,
91 9.448741184314713e-05,
92 0.00012044865434290841,
93 9.239319479092956e-05,
94 0.00011777623876696452,
95 0.00015112041728571057,
96 0.00016604004485998303,
97 0.00017201283480972052
98 ]
99 }
100 ],
101 "mean": 0.00015556820017081918,
102 "std": 7.185317458066955e-05,
103 "per_seed": [
104 0.0003302668046671897,
105 9.448741184314713e-05,
106 0.00012044865434290841,
107 9.239319479092956e-05,
108 0.00011777623876696452,
109 0.00015112041728571057,
110 0.00016604004485998303,
111 0.00017201283480972052
112 ],
113 "n": 8
114 },
115 "comparison": {
116 "delta_mean": -1.1292637282167561e-05,
117 "idea_wins": 3,
118 "n_pairs": 8,
119 "per_seed_diffs": [
120 5.484442226588726e-06,
121 4.398915189085528e-05,
122 5.1458227972034365e-05,
123 -4.435759910847992e-05,
124 -9.414643136551604e-05,
125 -0.00023815836175344884,
126 7.155440107453614e-05,
127 0.00011383507080608979
128 ],
129 "p_value": 0.82275,
130 "mde": 9.440018542032603e-05,
131 "mde_rel_pct": 56.57420090973917,
132 "verdict": "no significant win",
133 "system_worked": false
134 },
135 "protocol_notes": "Dynamics pendulum is structurally matched to stability/control. Adam is shared; only per-update controller scaling differs.",
136 "mechanism_signature": {
137 "predicted_cap_mean": 3.5256107662831653,
138 "observed_accepted_eta_mean": 2.748427854677808,
139 "proposal_cap_fraction": 0.3080357142857143,
140 "observed_update_norm_mean": 1.5009487083821522,
141 "gradient_ratio_to_next_loss_change_corr": -0.06326550793870624,
142 "n_trained_updates": 672,
143 "confirmed": true
144 },
145 "math_sanity": {
146 "cap_monotone_in_S": true,
147 "formula": "eta_cap=1+sqrt(1+2S)*sqrt(rho*v+eps)"
148 }
149}