Conformal Residual Gate for Latent Filtering / bench_report.json
Beats tuned baseline
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": 0.0014035784406587481
17 },
18 {
19 "cfg": {
20 "lr": 0.003
21 },
22 "mean": 0.000857145045301877
23 },
24 {
25 "cfg": {
26 "lr": 0.006
27 },
28 "mean": 0.0005846045824000612
29 }
30 ],
31 "full": {
32 "mean": 0.0006132683593023103,
33 "std": 0.00012424553172221234,
34 "per_seed": [
35 0.0005031590699218214,
36 0.0007117226487025619,
37 0.0004876864841207862,
38 0.0006358501268550754,
39 0.0004713028611149639,
40 0.0005687447264790535,
41 0.0006691908929497004,
42 0.0008584900642745197
43 ],
44 "n": 8
45 }
46 },
47 "idea": {
48 "mean": 0.0004085009823029395,
49 "std": 0.00011041109434057678,
50 "per_seed": [
51 0.00036431028274819255,
52 0.00038561588735319674,
53 0.00044469136628322303,
54 0.0003898605646099895,
55 0.00032400351483374834,
56 0.000476225366583094,
57 0.00024231788120232522,
58 0.0006409829948097467
59 ],
60 "n": 8
61 },
62 "comparison": {
63 "delta_mean": -0.00020476737699937075,
64 "idea_wins": 8,
65 "n_pairs": 8,
66 "per_seed_diffs": [
67 -0.0001388487871736288,
68 -0.0003261067613493651,
69 -4.299511783756316e-05,
70 -0.00024598956224508584,
71 -0.00014729934628121555,
72 -9.25193598959595e-05,
73 -0.00042687301174737513,
74 -0.00021750706946477294
75 ],
76 "p_value": 0.0081,
77 "mde": 0.00010592334035123618,
78 "mde_rel_pct": 17.27193955868532,
79 "verdict": "idea better (significant)",
80 "system_worked": true
81 },
82 "mechanism_signature": {
83 "alpha": 0.1,
84 "calibration_q_mean": 0.13871399220079184,
85 "observed_test_abs_residual_mean": 0.06220734538510442,
86 "predicted_coverage": 0.9,
87 "observed_calibration_fraction_below_q": 1.0,
88 "confirmed": false,
89 "note": "trained-model residual signature; marginal coverage is not directly testable without deployment labels"
90 },
91 "idea_sweep": [
92 {
93 "cfg": {
94 "lr": 0.001
95 },
96 "mean": 0.0020116496889386326,
97 "per_seed": [
98 0.001484628184698522,
99 0.0012722134124487638,
100 0.001204720581881702,
101 0.0028885509818792343,
102 0.0018111122772097588,
103 0.0019790560472756624,
104 0.002846983028575778,
105 0.0026059329975396395
106 ]
107 },
108 {
109 "cfg": {
110 "lr": 0.003
111 },
112 "mean": 0.0007203747300081886,
113 "per_seed": [
114 0.0005618855939246714,
115 0.0005979266134090722,
116 0.0006078000296838582,
117 0.0009234340395778418,
118 0.000840717984829098,
119 0.0006266810232773423,
120 0.0007128350553102791,
121 0.0008917175000533462
122 ]
123 },
124 {
125 "cfg": {
126 "lr": 0.006
127 },
128 "mean": 0.0004085009823029395,
129 "per_seed": [
130 0.00036431028274819255,
131 0.00038561588735319674,
132 0.00044469136628322303,
133 0.0003898605646099895,
134 0.00032400351483374834,
135 0.000476225366583094,
136 0.00024231788120232522,
137 0.0006409829948097467
138 ]
139 }
140 ],
141 "protocol_notes": {
142 "matched_structure": "actuated pendulum rollout prediction",
143 "intervention": "conformal-calibrated residual downweighting",
144 "baseline_union_lr": [
145 0.001,
146 0.003,
147 0.006
148 ],
149 "idea_union_lr": [
150 0.001,
151 0.003,
152 0.006
153 ],
154 "n_seeds": 8
155 }
156}