CEGAR-certified latent-state abstraction / 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": 0.01,
10 "weight_decay": 0.0
11 },
12 "sweep": [
13 {
14 "cfg": {
15 "lr": 0.001,
16 "weight_decay": 0.0
17 },
18 "mean": 0.004545392468571663
19 },
20 {
21 "cfg": {
22 "lr": 0.003,
23 "weight_decay": 0.0
24 },
25 "mean": 0.00358188062091358
26 },
27 {
28 "cfg": {
29 "lr": 0.01,
30 "weight_decay": 0.0
31 },
32 "mean": 0.0015466218756046146
33 }
34 ],
35 "full": {
36 "mean": 0.0014403377863345668,
37 "std": 0.0003958335789015625,
38 "per_seed": [
39 0.001036701025441289,
40 0.0019527574768289924,
41 0.0010734819807112217,
42 0.0021235470194369555,
43 0.001080187619663775,
44 0.0011896038195118308,
45 0.001624842407181859,
46 0.001441580941900611
47 ],
48 "n": 8
49 }
50 },
51 "idea": {
52 "mean": 0.004763860866660252,
53 "std": 0.002933690813628907,
54 "per_seed": [
55 0.011456220410764217,
56 0.002699522068724036,
57 0.0031392518430948257,
58 0.0025344439782202244,
59 0.007165998220443726,
60 0.0028000844176858664,
61 0.003200623905286193,
62 0.005114742089062929
63 ],
64 "n": 8
65 },
66 "comparison": {
67 "delta_mean": 0.0033235230803256854,
68 "idea_wins": 0,
69 "n_pairs": 8,
70 "per_seed_diffs": [
71 0.010419519385322928,
72 0.0007467645918950438,
73 0.002065769862383604,
74 0.00041089695878326893,
75 0.006085810600779951,
76 0.0016104805981740355,
77 0.0015757814981043339,
78 0.0036731611471623182
79 ],
80 "p_value": 0.0081,
81 "mde": 0.002840810147756273,
82 "mde_rel_pct": 197.2322169638893,
83 "verdict": "idea worse (significant)",
84 "system_worked": false
85 },
86 "mechanism_signature": {
87 "idea_sweep": [
88 {
89 "cfg": {
90 "lr": 0.001,
91 "weight_decay": 0.0
92 },
93 "result": {
94 "mean": 0.004763860866660252,
95 "std": 0.002933690813628907,
96 "per_seed": [
97 0.011456220410764217,
98 0.002699522068724036,
99 0.0031392518430948257,
100 0.0025344439782202244,
101 0.007165998220443726,
102 0.0028000844176858664,
103 0.003200623905286193,
104 0.005114742089062929
105 ],
106 "n": 8
107 }
108 },
109 {
110 "cfg": {
111 "lr": 0.003,
112 "weight_decay": 0.0
113 },
114 "result": {
115 "mean": 0.0048263132775900885,
116 "std": 0.0025504190600527878,
117 "per_seed": [
118 0.0045633199624717236,
119 0.004938819445669651,
120 0.0009674743050709367,
121 0.010774293914437294,
122 0.004922726657241583,
123 0.004287341609597206,
124 0.0037127339746803045,
125 0.00444379635155201
126 ],
127 "n": 8
128 }
129 },
130 {
131 "cfg": {
132 "lr": 0.01,
133 "weight_decay": 0.0
134 },
135 "result": {
136 "mean": 0.04266263055615127,
137 "std": 0.02664842856440231,
138 "per_seed": [
139 0.07184934616088867,
140 0.04949978366494179,
141 0.023517103865742683,
142 0.029628023505210876,
143 0.008588656783103943,
144 0.09184401482343674,
145 0.048785608261823654,
146 0.017588507384061813
147 ],
148 "n": 8
149 }
150 }
151 ],
152 "selected_cfg": {
153 "lr": 0.001,
154 "weight_decay": 0.0
155 },
156 "signature": {
157 "prediction": "latent box transitions should contain observed trained-model successors",
158 "observed_transitions": 159,
159 "same_cell_fraction": 0.0,
160 "mean_observed_box_span": 0.17268966138362885,
161 "cells": 4,
162 "metric_for_signature_model": 0.011456220410764217,
163 "confirmed": true
164 }
165 }
166}