Tangential Bellman Tie Resolver / bench_report.json
Failed on benchmark
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": 0.005,
10 "wd": 0.0
11 },
12 "sweep": [
13 {
14 "cfg": {
15 "lr": 0.001,
16 "wd": 0.0
17 },
18 "mean": 0.001154391880845651
19 },
20 {
21 "cfg": {
22 "lr": 0.001,
23 "wd": 0.0001
24 },
25 "mean": 0.0011398583883419633
26 },
27 {
28 "cfg": {
29 "lr": 0.003,
30 "wd": 0.0
31 },
32 "mean": 0.0006730457389494404
33 },
34 {
35 "cfg": {
36 "lr": 0.003,
37 "wd": 0.0001
38 },
39 "mean": 0.00066975416848436
40 },
41 {
42 "cfg": {
43 "lr": 0.005,
44 "wd": 0.0
45 },
46 "mean": 0.00047177909436868504
47 },
48 {
49 "cfg": {
50 "lr": 0.005,
51 "wd": 0.0001
52 },
53 "mean": 0.0004883268411504105
54 }
55 ],
56 "full": {
57 "mean": 0.00045577605851576664,
58 "std": 9.045873661427688e-05,
59 "per_seed": [
60 0.00045988193596713245,
61 0.0005730087868869305,
62 0.0003259392105974257,
63 0.0005282864440232515,
64 0.000366326974472031,
65 0.00035140709951519966,
66 0.0005543518927879632,
67 0.0004870061238761991
68 ],
69 "n": 8
70 }
71 },
72 "idea": {
73 "mean": 0.7881884723901749,
74 "std": 0.0419254847955854,
75 "per_seed": [
76 0.7438060641288757,
77 0.8217167854309082,
78 0.8174083828926086,
79 0.857958197593689,
80 0.8094730973243713,
81 0.7287576198577881,
82 0.7656309604644775,
83 0.7607566714286804
84 ],
85 "n": 8
86 },
87 "comparison": {
88 "delta_mean": 0.7877326963316591,
89 "idea_wins": 0,
90 "n_pairs": 8,
91 "per_seed_diffs": [
92 0.7433461821929086,
93 0.8211437766440213,
94 0.8170824436820112,
95 0.8574299111496657,
96 0.8091067703498993,
97 0.7284062127582729,
98 0.7650766085716896,
99 0.7602696653048042
100 ],
101 "p_value": 0.0081,
102 "mde": 0.03746007735504655,
103 "mde_rel_pct": 8218.965576435758,
104 "verdict": "idea worse (significant)",
105 "system_worked": false
106 },
107 "mechanism_signature": {
108 "math_check": {
109 "operator_lipschitz_ratio": 0.14484610436506878,
110 "beta": 0.73,
111 "iteration_ratios": [
112 0.4969973584994038,
113 0.7098904435991482,
114 0.7246338853808736,
115 0.7286066687505649,
116 0.7293810555254636,
117 0.7299559137698793,
118 0.729981689594991,
119 0.7299953759512697
120 ],
121 "max_ratio_over_beta": 0.9999936656866709
122 },
123 "trained_behavior": {
124 "resolver_weight_span": {
125 "observed_mean": 0.0036720615407830337,
126 "per_seed": [
127 0.000908972229808569,
128 0.006066256668418646,
129 0.0012230807915329933,
130 0.019085220992565155,
131 0.0007767750066705048,
132 0.0002523639996070415,
133 0.000828575633931905,
134 0.00023524700372945517
135 ]
136 },
137 "hard_best_frequency": {
138 "observed_mean": 0.4099999964237213,
139 "per_seed": [
140 0.0,
141 1.0,
142 0.7674999833106995,
143 0.0,
144 0.512499988079071,
145 0.0,
146 1.0,
147 0.0
148 ]
149 },
150 "prediction": "resolver branch weights vary continuously with counterfactual action consequences; measured on trained models",
151 "confirmed": false
152 },
153 "idea_sweep": [
154 {
155 "cfg": {
156 "lr": 0.001,
157 "wd": 0.0
158 },
159 "mean": 0.7889741063117981
160 },
161 {
162 "cfg": {
163 "lr": 0.003,
164 "wd": 0.0
165 },
166 "mean": 0.7881884723901749
167 },
168 {
169 "cfg": {
170 "lr": 0.005,
171 "wd": 0.0
172 },
173 "mean": 0.7896546497941017
174 }
175 ]
176 },
177 "protocol_notes": {
178 "n_train": 1200,
179 "n_test": 400,
180 "epochs": 10,
181 "structural_match": "controlled pendulum dynamics; action branches are final-step control perturbations",
182 "baseline_knobs_swept": [
183 "lr",
184 "weight_decay"
185 ]
186 }
187}