Feasibility-Ranked Group Policy Gradient / bench_report.json
Unverified
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": 0.006
10 },
11 "sweep": [
12 {
13 "cfg": {
14 "lr": 0.001
15 },
16 "mean": 0.0013222937704995275
17 },
18 {
19 "cfg": {
20 "lr": 0.003
21 },
22 "mean": 0.0008866888092597947
23 },
24 {
25 "cfg": {
26 "lr": 0.006
27 },
28 "mean": 0.0005522354185814038
29 }
30 ],
31 "full": {
32 "mean": 0.000564589565328788,
33 "std": 0.00015977200389635955,
34 "per_seed": [
35 0.00038707852945663035,
36 0.0005981649155728519,
37 0.00037658659857697785,
38 0.0008471116307191551,
39 0.0004592438053805381,
40 0.0004694555827882141,
41 0.0007397530134767294,
42 0.0006393224466592073
43 ],
44 "n": 8
45 }
46 },
47 "idea": {
48 "mean": 0.000554656424355926,
49 "std": 0.0001615859243021339,
50 "per_seed": [
51 0.000384580809623003,
52 0.0005927970050834119,
53 0.000364386651199311,
54 0.0008439605007879436,
55 0.0004653841897379607,
56 0.00047854892909526825,
57 0.0007660035626031458,
58 0.0005415897467173636
59 ],
60 "n": 8
61 },
62 "comparison": {
63 "delta_mean": -9.93314097286202e-06,
64 "idea_wins": 5,
65 "n_pairs": 8,
66 "per_seed_diffs": [
67 -2.497719833627343e-06,
68 -5.367910489439964e-06,
69 -1.2199947377666831e-05,
70 -3.1511299312114716e-06,
71 6.14038435742259e-06,
72 9.093346307054162e-06,
73 2.6250549126416445e-05,
74 -9.773269994184375e-05
75 ],
76 "p_value": 0.68235,
77 "mde": 3.123112811032005e-05,
78 "mde_rel_pct": 5.531651668435042,
79 "verdict": "no significant win",
80 "system_worked": false
81 },
82 "mechanism_signature": {
83 "observed_normalized_variance": 1.0,
84 "predicted_normalized_variance": 1.0,
85 "observed_weighted_feasible_mean": 2.821424961090088,
86 "observed_weighted_violating_mean": -0.26333290338516235,
87 "observed_margin": 3.0847578644752502,
88 "predicted_margin": 3.0847578644752502,
89 "confirmed": true,
90 "note": "Signature measured from trained rnn_small behaviour. This is a supervised transfer of trajectory ranking; the canonical dynamics track has no action-policy interface, so exact PPO ratios are not tested."
91 },
92 "idea_sweep": [
93 {
94 "cfg": {
95 "lr": 0.001,
96 "wf": 1.5,
97 "wv": 1.0
98 },
99 "result": {
100 "mean": 0.0013214157661423087,
101 "std": 0.0003696944861183539,
102 "per_seed": [
103 0.000921291415579617,
104 0.001448000897653401,
105 0.001050042686983943,
106 0.001866328064352274
107 ],
108 "n": 4
109 }
110 },
111 {
112 "cfg": {
113 "lr": 0.003,
114 "wf": 2.0,
115 "wv": 1.0
116 },
117 "result": {
118 "mean": 0.0008860538364388049,
119 "std": 0.0003121798062305589,
120 "per_seed": [
121 0.0005791240837424994,
122 0.0009322315454483032,
123 0.0006562279304489493,
124 0.0013766317861154675
125 ],
126 "n": 4
127 }
128 },
129 {
130 "cfg": {
131 "lr": 0.006,
132 "wf": 3.0,
133 "wv": 1.0
134 },
135 "result": {
136 "mean": 0.0005464312416734174,
137 "std": 0.00019365513117499712,
138 "per_seed": [
139 0.000384580809623003,
140 0.0005927970050834119,
141 0.000364386651199311,
142 0.0008439605007879436
143 ],
144 "n": 4
145 }
146 }
147 ],
148 "protocol_notes": {
149 "selection_seeds": [
150 0,
151 1,
152 2,
153 3
154 ],
155 "paired_seeds": [
156 0,
157 1,
158 2,
159 3,
160 4,
161 5,
162 6,
163 7
164 ],
165 "epochs": 10,
166 "n_train": 1200,
167 "n_test": 400,
168 "track_justification": "dynamics is the built-in controlled pendulum multi-step state/action-window task, matching control and terminal feasibility structure."
169 }
170}