Forcing-Consistency Training Constraint / bench_report.json
Mechanism confirmed, baseline not beaten
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": 0.03,
10 "weight_decay": 0.0
11 },
12 "sweep": [
13 {
14 "cfg": {
15 "lr": 0.01,
16 "weight_decay": 0.0
17 },
18 "mean": 0.0014865802077110857
19 },
20 {
21 "cfg": {
22 "lr": 0.01,
23 "weight_decay": 0.0001
24 },
25 "mean": 0.0014641905727330595
26 },
27 {
28 "cfg": {
29 "lr": 0.03,
30 "weight_decay": 0.0
31 },
32 "mean": 0.0012827248574467376
33 },
34 {
35 "cfg": {
36 "lr": 0.03,
37 "weight_decay": 0.0001
38 },
39 "mean": 0.0013233663485152647
40 },
41 {
42 "cfg": {
43 "lr": 0.1,
44 "weight_decay": 0.0
45 },
46 "mean": 0.006263854738790542
47 },
48 {
49 "cfg": {
50 "lr": 0.1,
51 "weight_decay": 0.0001
52 },
53 "mean": 0.014768014312721789
54 }
55 ],
56 "full": {
57 "mean": 0.0010541816263867076,
58 "std": 0.0005985854110248448,
59 "per_seed": [
60 0.0023917919024825096,
61 0.0013993943575769663,
62 0.0006620299536734819,
63 0.0006776832160539925,
64 0.0012405221350491047,
65 0.001023401622660458,
66 0.000684684026055038,
67 0.0003539457975421101
68 ],
69 "n": 8
70 }
71 },
72 "idea": {
73 "mean": 0.002501493159797974,
74 "std": 0.000829447102954745,
75 "per_seed": [
76 0.0022331809159368277,
77 0.001679917797446251,
78 0.0021533623803406954,
79 0.003017958952113986,
80 0.0014837837079539895,
81 0.003271000226959586,
82 0.0020722420886158943,
83 0.0041004992090165615
84 ],
85 "n": 8
86 },
87 "comparison": {
88 "delta_mean": 0.0014473115334112663,
89 "idea_wins": 1,
90 "n_pairs": 8,
91 "per_seed_diffs": [
92 -0.00015861098654568195,
93 0.00028052343986928463,
94 0.0014913324266672134,
95 0.0023402757360599935,
96 0.00024326157290488482,
97 0.002247598604299128,
98 0.0013875580625608563,
99 0.0037465534114744514
100 ],
101 "p_value": 0.01485,
102 "mde": 0.0011002418315854009,
103 "mde_rel_pct": 104.36928552402951,
104 "verdict": "idea worse (significant)",
105 "system_worked": false
106 },
107 "mechanism_signature": {
108 "math_check": {
109 "js_equal": 0.0,
110 "js_unequal": 0.27539611524877045,
111 "js_symmetric_error": 0.0,
112 "intersection_feasible": true,
113 "intersection_empty": true,
114 "passed": true
115 },
116 "idea_sweep": [
117 {
118 "cfg": {
119 "lr": 0.01,
120 "weight_decay": 0.0,
121 "fc_weight": 0.08,
122 "safety_weight": 0.08
123 },
124 "mean": 0.00227110501145944
125 },
126 {
127 "cfg": {
128 "lr": 0.03,
129 "weight_decay": 0.0,
130 "fc_weight": 0.08,
131 "safety_weight": 0.08
132 },
133 "mean": 0.004607509879861027
134 },
135 {
136 "cfg": {
137 "lr": 0.1,
138 "weight_decay": 0.0,
139 "fc_weight": 0.08,
140 "safety_weight": 0.08
141 },
142 "mean": 0.01652669277973473
143 }
144 ],
145 "best_idea_cfg": {
146 "lr": 0.01,
147 "weight_decay": 0.0,
148 "fc_weight": 0.08,
149 "safety_weight": 0.08
150 },
151 "signature": {
152 "prediction": "FC should lower observational kappa on trained recurrent models",
153 "baseline_kappa_mean": 0.005694788647815585,
154 "idea_kappa_mean": 0.005446167720947415,
155 "baseline_safety_hinge_mean": 0.2168501764535904,
156 "idea_safety_hinge_mean": 0.1869419850409031,
157 "confirmed": true
158 }
159 }
160}