Dyadic Hankel Boundary 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 "temperature": 1.0
11 },
12 "sweep": [
13 {
14 "cfg": {
15 "lr": 0.001,
16 "temperature": 1.0
17 },
18 "mean": 0.34783465415239334
19 },
20 {
21 "cfg": {
22 "lr": 0.003,
23 "temperature": 1.0
24 },
25 "mean": 0.3491433709859848
26 },
27 {
28 "cfg": {
29 "lr": 0.005,
30 "temperature": 1.0
31 },
32 "mean": 0.40824612975120544
33 }
34 ],
35 "full": {
36 "mean": 0.36294760182499886,
37 "std": 0.054345281417173805,
38 "per_seed": [
39 0.39175504446029663,
40 0.3038984537124634,
41 0.30141177773475647,
42 0.3942733407020569,
43 0.35095956921577454,
44 0.3010490834712982,
45 0.4579026699066162,
46 0.4023308753967285
47 ],
48 "n": 8
49 }
50 },
51 "idea": {
52 "mean": 0.3599541373550892,
53 "std": 0.05518226062931197,
54 "per_seed": [
55 0.3880409896373749,
56 0.2888491451740265,
57 0.32456186413764954,
58 0.3811890184879303,
59 0.3370722830295563,
60 0.29420310258865356,
61 0.4586992561817169,
62 0.40701743960380554
63 ],
64 "n": 8
65 },
66 "comparison": {
67 "delta_mean": -0.002993464469909668,
68 "idea_wins": 5,
69 "n_pairs": 8,
70 "per_seed_diffs": [
71 -0.003714054822921753,
72 -0.01504930853843689,
73 0.023150086402893066,
74 -0.013084322214126587,
75 -0.013887286186218262,
76 -0.006845980882644653,
77 0.000796586275100708,
78 0.004686564207077026
79 ],
80 "p_value": 0.53235,
81 "mde": 0.010661430606466927,
82 "mde_rel_pct": 2.9374572397939445,
83 "verdict": "no measurable effect",
84 "system_worked": false
85 },
86 "mechanism_signature": {
87 "idea_config": {
88 "lr": 0.001,
89 "rank": 4
90 },
91 "candidate_results": [
92 {
93 "cfg": {
94 "lr": 0.001,
95 "rank": 2
96 },
97 "result": {
98 "mean": 0.3600270599126816,
99 "std": 0.05461125040141159,
100 "per_seed": [
101 0.3894391655921936,
102 0.2890293002128601,
103 0.3245948255062103,
104 0.3806515336036682,
105 0.33679887652397156,
106 0.2958018183708191,
107 0.4570717513561249,
108 0.40682920813560486
109 ],
110 "n": 8
111 }
112 },
113 {
114 "cfg": {
115 "lr": 0.001,
116 "rank": 4
117 },
118 "result": {
119 "mean": 0.3599541373550892,
120 "std": 0.05518226062931197,
121 "per_seed": [
122 0.3880409896373749,
123 0.2888491451740265,
124 0.32456186413764954,
125 0.3811890184879303,
126 0.3370722830295563,
127 0.29420310258865356,
128 0.4586992561817169,
129 0.40701743960380554
130 ],
131 "n": 8
132 }
133 },
134 {
135 "cfg": {
136 "lr": 0.001,
137 "rank": 6
138 },
139 "result": {
140 "mean": 0.3600568361580372,
141 "std": 0.05501619060413237,
142 "per_seed": [
143 0.38814911246299744,
144 0.2887286841869354,
145 0.32432663440704346,
146 0.38098469376564026,
147 0.3378050625324249,
148 0.29527634382247925,
149 0.45912855863571167,
150 0.40605559945106506
151 ],
152 "n": 8
153 }
154 }
155 ],
156 "signature": {
157 "epsilon": 0.001,
158 "predicted_rank_bound": 5,
159 "observed_rank": 4,
160 "trained_model_cross_output_relative_error": 3.261449819547124e-05,
161 "predicted_error_bound": 0.001,
162 "confirmed": true
163 }
164 },
165 "protocol_notes": {
166 "paired_seeds": [
167 0,
168 1,
169 2,
170 3,
171 4,
172 5,
173 6,
174 7
175 ],
176 "epochs": 20,
177 "architecture_difference_only": "dense self-attention vs within-side dense plus cached dyadic Hankel cross branch"
178 }
179}