Excitation-Gated Latent Frame Calibration / 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": 0.01
10 },
11 "sweep": [
12 {
13 "cfg": {
14 "lr": 0.001
15 },
16 "mean": 0.0061489523795899
17 },
18 {
19 "cfg": {
20 "lr": 0.003
21 },
22 "mean": 0.0013683813158422709
23 },
24 {
25 "cfg": {
26 "lr": 0.01
27 },
28 "mean": 0.0007866571540944278
29 }
30 ],
31 "full": {
32 "mean": 0.0006869791614008136,
33 "std": 0.0001962519169577577,
34 "per_seed": [
35 0.0008465873543173075,
36 0.0005530674243345857,
37 0.0006530178943648934,
38 0.0010939559433609247,
39 0.0006159734330140054,
40 0.0003910730010829866,
41 0.0006219872157089412,
42 0.0007201710250228643
43 ],
44 "n": 8
45 }
46 },
47 "idea": {
48 "mean": 0.000751409461372532,
49 "std": 0.0002372904792985927,
50 "per_seed": [
51 0.001134159043431282,
52 0.0006296231294982135,
53 0.0008904425776563585,
54 0.0010553350439295173,
55 0.0005659241578541696,
56 0.0003978781751357019,
57 0.0006218505441211164,
58 0.0007160630193538964
59 ],
60 "n": 8,
61 "best_cfg": {
62 "lr": 0.01
63 }
64 },
65 "comparison": {
66 "delta_mean": 6.443029997171834e-05,
67 "idea_wins": 4,
68 "n_pairs": 8,
69 "per_seed_diffs": [
70 0.00028757168911397457,
71 7.655570516362786e-05,
72 0.00023742468329146504,
73 -3.862089943140745e-05,
74 -5.0049275159835815e-05,
75 6.8051740527153015e-06,
76 -1.3667158782482147e-07,
77 -4.108005668967962e-06
78 ],
79 "p_value": 0.2782,
80 "mde": 0.00010753523221065781,
81 "mde_rel_pct": 15.6533470376865,
82 "verdict": "no significant win",
83 "system_worked": false
84 },
85 "mechanism_signature": {
86 "low_q25_sigma": 0.07731843180954456,
87 "high_q75_sigma": 0.1157134473323822,
88 "low_excitation_train_mse": 0.0012986520305275917,
89 "high_excitation_train_mse": 0.0005935525405220687,
90 "observed_error_gap_low_minus_high": 0.000705099490005523,
91 "prediction": "low-excitation windows have higher residual error",
92 "confirmed": true
93 },
94 "idea_sweep": [
95 {
96 "cfg": {
97 "lr": 0.001
98 },
99 "mean": 0.00511605893552769,
100 "std": 0.0024142004953189313,
101 "per_seed": [
102 0.0015621563652530313,
103 0.005496674682945013,
104 0.007589519023895264,
105 0.006733444053679705,
106 0.0014989599585533142,
107 0.0036795316264033318,
108 0.0063422685489058495,
109 0.00802591722458601
110 ],
111 "n": 8
112 },
113 {
114 "cfg": {
115 "lr": 0.003
116 },
117 "mean": 0.001308875231188722,
118 "std": 0.00023694831535078266,
119 "per_seed": [
120 0.0010585704585537314,
121 0.0011665996862575412,
122 0.001426042872481048,
123 0.0014445277629420161,
124 0.0009118119487538934,
125 0.0012852930231019855,
126 0.0014918958768248558,
127 0.0016862602205947042
128 ],
129 "n": 8
130 },
131 {
132 "cfg": {
133 "lr": 0.01
134 },
135 "mean": 0.000751409461372532,
136 "std": 0.0002372904792985927,
137 "per_seed": [
138 0.001134159043431282,
139 0.0006296231294982135,
140 0.0008904425776563585,
141 0.0010553350439295173,
142 0.0005659241578541696,
143 0.0003978781751357019,
144 0.0006218505441211164,
145 0.0007160630193538964
146 ],
147 "n": 8
148 }
149 ],
150 "method": {
151 "description": "Per-window latent-frame consistency loss weighted by normalized smallest Jacobian singular value.",
152 "tau": 0.2,
153 "floor": 0.15,
154 "epochs": 15,
155 "batch": 128
156 },
157 "protocol_notes": "Matched dynamics track: controlled pendulum rollout and recurrent GRU. Baseline and idea share rnn_small, Adam, epochs, batch, data, and the union learning-rate grid; idea was evaluated at baseline-best and two nearby settings."
158}