Forward-Intersection Spectral Latent Dynamics / bench_output.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 "weight_decay": 0.0
11 },
12 "sweep": [
13 {
14 "cfg": {
15 "lr": 0.001,
16 "weight_decay": 0.0
17 },
18 "mean": 0.001154391880845651
19 },
20 {
21 "cfg": {
22 "lr": 0.003,
23 "weight_decay": 0.0
24 },
25 "mean": 0.0006730457389494404
26 },
27 {
28 "cfg": {
29 "lr": 0.01,
30 "weight_decay": 0.0
31 },
32 "mean": 0.00024367208061448764
33 },
34 {
35 "cfg": {
36 "lr": 0.003,
37 "weight_decay": 0.0001
38 },
39 "mean": 0.00066975416848436
40 }
41 ],
42 "full": {
43 "mean": 0.00020419583597686142,
44 "std": 0.00010942496465474286,
45 "per_seed": [
46 0.00037123056245036423,
47 0.00033986172638833523,
48 7.115668995538726e-05,
49 0.00019243934366386384,
50 0.00017716105503495783,
51 0.00011334212467772886,
52 7.963812095113099e-05,
53 0.0002887370646931231
54 ],
55 "n": 8
56 }
57 },
58 "idea": {
59 "mean": 0.3719000369310379,
60 "std": 0.1780352000875736,
61 "per_seed": [
62 0.25049954652786255,
63 0.32161659002304077,
64 0.2485007792711258,
65 0.6178863644599915,
66 0.1744787096977234,
67 0.22901414334774017,
68 0.45399707555770874,
69 0.6792070865631104
70 ],
71 "n": 8
72 },
73 "comparison": {
74 "delta_mean": 0.37169584109506104,
75 "idea_wins": 0,
76 "n_pairs": 8,
77 "per_seed_diffs": [
78 0.2501283159654122,
79 0.32127672829665244,
80 0.2484296225811704,
81 0.6176939251163276,
82 0.17430154864268843,
83 0.22890080122306244,
84 0.4539174374367576,
85 0.6789183494984172
86 ],
87 "p_value": 0.0081,
88 "mde": 0.1591303337139686,
89 "mde_rel_pct": 77930.25403906891,
90 "verdict": "idea worse (significant)",
91 "system_worked": false
92 },
93 "mechanism_signature": {
94 "mechanism_signature": {
95 "prediction": "compatible projection reduces unsupported hidden transition energy/rank without changing task architecture",
96 "observed": [
97 {
98 "seed": 0,
99 "metric": 0.25049954652786255,
100 "projection_mse": 0.013389366678893566,
101 "rank": 2,
102 "raw_spectral_radius": 1.101452112197876
103 },
104 {
105 "seed": 1,
106 "metric": 0.32161659002304077,
107 "projection_mse": 0.013996558263897896,
108 "rank": 1,
109 "raw_spectral_radius": 1.0808608531951904
110 },
111 {
112 "seed": 2,
113 "metric": 0.2485007792711258,
114 "projection_mse": 0.011121638119220734,
115 "rank": 2,
116 "raw_spectral_radius": 1.0855185985565186
117 },
118 {
119 "seed": 3,
120 "metric": 0.6178863644599915,
121 "projection_mse": 0.018192939460277557,
122 "rank": 2,
123 "raw_spectral_radius": 1.1126620769500732
124 }
125 ],
126 "mean_projection_mse": 0.014175125630572438,
127 "mean_rank": 1.75,
128 "confirmed": true
129 },
130 "custom_track": null,
131 "idea_cfg": {
132 "lr": 0.01,
133 "weight_decay": 0.0,
134 "tau": 0.15,
135 "levels": 1
136 },
137 "idea_grid": [
138 {
139 "lr": 0.01,
140 "weight_decay": 0.0,
141 "tau": 0.1,
142 "levels": 1
143 },
144 {
145 "lr": 0.01,
146 "weight_decay": 0.0,
147 "tau": 0.15,
148 "levels": 1
149 },
150 {
151 "lr": 0.01,
152 "weight_decay": 0.0,
153 "tau": 0.25,
154 "levels": 1
155 }
156 ]
157 }
158}