Implicit Higher-Order TPR Memory / bench_report.json
Mechanism confirmed, baseline not beaten
1{
2 "bench_version": 1,
3 "track": "sequence",
4 "model": "shared_transformer_token_attention",
5 "metric_direction": "lower is better",
6 "n_seeds": 8,
7 "baseline": {
8 "best_cfg": {
9 "lr": 0.003,
10 "tau": 0.25
11 },
12 "sweep": [
13 {
14 "cfg": {
15 "lr": 0.001,
16 "tau": 0.25
17 },
18 "mean": 1.0073408037424088
19 },
20 {
21 "cfg": {
22 "lr": 0.001,
23 "tau": 0.5
24 },
25 "mean": 1.0072430819272995
26 },
27 {
28 "cfg": {
29 "lr": 0.003,
30 "tau": 0.25
31 },
32 "mean": 0.9897160083055496
33 },
34 {
35 "cfg": {
36 "lr": 0.003,
37 "tau": 0.5
38 },
39 "mean": 0.9971326589584351
40 },
41 {
42 "cfg": {
43 "lr": 0.01,
44 "tau": 0.25
45 },
46 "mean": 1.041932761669159
47 },
48 {
49 "cfg": {
50 "lr": 0.01,
51 "tau": 0.5
52 },
53 "mean": 1.0047774612903595
54 }
55 ],
56 "full": {
57 "mean": 1.0231029689311981,
58 "std": 0.11072729156070826,
59 "per_seed": [
60 0.9939484000205994,
61 0.8335484266281128,
62 1.1223986148834229,
63 1.0089685916900635,
64 0.9049072861671448,
65 1.0517245531082153,
66 1.2106995582580566,
67 1.0586283206939697
68 ],
69 "n": 8
70 }
71 },
72 "idea": {
73 "mean": 1.0160572454333305,
74 "std": 0.10683002426665564,
75 "per_seed": [
76 0.9757922291755676,
77 0.856389582157135,
78 1.1297253370285034,
79 1.0210782289505005,
80 0.8922733664512634,
81 1.0258088111877441,
82 1.2052284479141235,
83 1.0221619606018066
84 ],
85 "n": 8
86 },
87 "comparison": {
88 "delta_mean": -0.007045723497867584,
89 "idea_wins": 5,
90 "n_pairs": 8,
91 "per_seed_diffs": [
92 -0.01815617084503174,
93 0.022841155529022217,
94 0.007326722145080566,
95 0.012109637260437012,
96 -0.012633919715881348,
97 -0.02591574192047119,
98 -0.0054711103439331055,
99 -0.036466360092163086
100 ],
101 "p_value": 0.34205,
102 "mde": 0.016858466245238783,
103 "mde_rel_pct": 1.6477780592162945,
104 "verdict": "no measurable effect",
105 "system_worked": false
106 },
107 "idea_best_cfg": {
108 "lr": 0.003,
109 "tau": 0.25
110 },
111 "idea_configs": [
112 {
113 "cfg": {
114 "lr": 0.001,
115 "tau": 0.25
116 },
117 "mean": 1.0335221663117409
118 },
119 {
120 "cfg": {
121 "lr": 0.003,
122 "tau": 0.25
123 },
124 "mean": 1.0160572454333305
125 },
126 {
127 "cfg": {
128 "lr": 0.01,
129 "tau": 0.25
130 },
131 "mean": 1.0669345706701279
132 }
133 ],
134 "runtime_sec": 50.11480712890625,
135 "mechanism_signature": {
136 "prediction": "two-factor contraction equals product of trained token-query similarities",
137 "predicted_max_abs_error": 0.0,
138 "observed_max_abs_error": 0.0,
139 "predicted_correlation": 1.0,
140 "observed_correlation": 0.9999999403953552,
141 "confirmed": true
142 },
143 "track_justification": "Sequence forecast has multi-token correlations; conjunction attention operates over token roles/factors.",
144 "custom_track": null
145}