Assignment Tree Attention / bench_report.json
Mechanism confirmed, baseline not beaten
1{
2 "bench_version": 1,
3 "track": "sequence",
4 "model": "transformer_tiny",
5 "metric_direction": "lower is better",
6 "n_seeds": 8,
7 "baseline": {
8 "best_cfg": {
9 "lr": 0.001,
10 "epochs": 2,
11 "K": 4,
12 "tau": 0.7,
13 "alpha": 0.5
14 },
15 "sweep": [
16 {
17 "cfg": {
18 "lr": 0.001,
19 "epochs": 2,
20 "K": 4,
21 "tau": 0.7,
22 "alpha": 0.5
23 },
24 "mean": 0.9283310621976852
25 },
26 {
27 "cfg": {
28 "lr": 0.003,
29 "epochs": 2,
30 "K": 4,
31 "tau": 0.7,
32 "alpha": 0.5
33 },
34 "mean": 1.0033762753009796
35 },
36 {
37 "cfg": {
38 "lr": 0.006,
39 "epochs": 2,
40 "K": 4,
41 "tau": 0.7,
42 "alpha": 0.5
43 },
44 "mean": 1.0917343497276306
45 }
46 ],
47 "full": {
48 "mean": 0.9580474123358727,
49 "std": 0.07569227656456869,
50 "per_seed": [
51 0.9577249884605408,
52 0.8674030900001526,
53 0.9789561033248901,
54 0.9092400670051575,
55 1.1384438276290894,
56 0.9376952052116394,
57 0.9597494006156921,
58 0.9151666164398193
59 ],
60 "n": 8
61 }
62 },
63 "idea": {
64 "mean": 0.9508912265300751,
65 "std": 0.05702490985176115,
66 "per_seed": [
67 0.9412082433700562,
68 0.8438071012496948,
69 0.9839955568313599,
70 0.9183895587921143,
71 1.0529438257217407,
72 0.9422717094421387,
73 0.9901344180107117,
74 0.9343793988227844
75 ],
76 "n": 8
77 },
78 "comparison": {
79 "delta_mean": -0.007156185805797577,
80 "idea_wins": 3,
81 "n_pairs": 8,
82 "per_seed_diffs": [
83 -0.01651674509048462,
84 -0.023595988750457764,
85 0.0050394535064697266,
86 0.009149491786956787,
87 -0.08550000190734863,
88 0.004576504230499268,
89 0.03038501739501953,
90 0.019212782382965088
91 ],
92 "p_value": 0.70955,
93 "mde": 0.030234908489597815,
94 "mde_rel_pct": 3.1558885395744953,
95 "verdict": "no measurable effect",
96 "system_worked": false
97 },
98 "mechanism_signature": {
99 "prediction": "connected bipartite tree retains 2L-1 edges independent of K",
100 "predicted_tree_edges": 63.0,
101 "observed_tree_edges": 63.0,
102 "edge_count_error": 0.0,
103 "confirmed": true,
104 "measured_on_trained_model": true
105 },
106 "idea_sweep": [
107 {
108 "cfg": {
109 "lr": 0.001,
110 "epochs": 2,
111 "K": 4,
112 "tau": 0.7,
113 "alpha": 0.5
114 },
115 "mean": 0.9508912265300751,
116 "std": 0.05702490985176115,
117 "per_seed": [
118 0.9412082433700562,
119 0.8438071012496948,
120 0.9839955568313599,
121 0.9183895587921143,
122 1.0529438257217407,
123 0.9422717094421387,
124 0.9901344180107117,
125 0.9343793988227844
126 ],
127 "n": 8
128 },
129 {
130 "cfg": {
131 "lr": 0.003,
132 "epochs": 2,
133 "K": 4,
134 "tau": 0.7,
135 "alpha": 0.5
136 },
137 "mean": 0.9828187823295593,
138 "std": 0.03844168507903302,
139 "per_seed": [
140 0.9239319562911987,
141 1.0003299713134766,
142 1.0384844541549683,
143 0.9352728724479675,
144 0.9620147347450256,
145 0.9814748167991638,
146 1.0294827222824097,
147 0.9915587306022644
148 ],
149 "n": 8
150 },
151 {
152 "cfg": {
153 "lr": 0.006,
154 "epochs": 2,
155 "K": 4,
156 "tau": 0.7,
157 "alpha": 0.5
158 },
159 "mean": 1.1042008697986603,
160 "std": 0.12163079384084556,
161 "per_seed": [
162 1.053992748260498,
163 0.8888604640960693,
164 1.2662471532821655,
165 1.2502007484436035,
166 1.0634373426437378,
167 1.0349705219268799,
168 1.2190437316894531,
169 1.056854248046875
170 ],
171 "n": 8
172 }
173 ]
174}