maxffarrell commited on
Commit
1d942dd
·
verified ·
1 Parent(s): 2a07692

Publish completed seed-42 robustness diagnostics; full study pending (part 2)

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .gitattributes +5 -0
  2. studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/current_only/parameter_shift_2node-predictions.npz +3 -0
  3. studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/current_only/small_room_3node-context.json +0 -0
  4. studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/current_only/small_room_3node-predictions.npz +3 -0
  5. studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/current_only/support.json +1167 -0
  6. studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/full_history/auditorium_3node-context.json +0 -0
  7. studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/full_history/auditorium_3node-predictions.npz +3 -0
  8. studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/full_history/calibration.json +2082 -0
  9. studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/full_history/parameter_shift_2node-context.json +0 -0
  10. studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/full_history/parameter_shift_2node-predictions.npz +3 -0
  11. studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/full_history/small_room_3node-context.json +0 -0
  12. studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/full_history/small_room_3node-predictions.npz +3 -0
  13. studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/full_history/support.json +1267 -0
  14. studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/thermal-evaluation.json +0 -0
  15. studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/thermal_temperature_only/auditorium_3node-context.json +0 -0
  16. studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/thermal_temperature_only/auditorium_3node-predictions.npz +0 -0
  17. studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/thermal_temperature_only/calibration.json +1842 -0
  18. studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/thermal_temperature_only/parameter_shift_2node-context.json +0 -0
  19. studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/thermal_temperature_only/parameter_shift_2node-predictions.npz +0 -0
  20. studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/thermal_temperature_only/small_room_3node-context.json +0 -0
  21. studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/thermal_temperature_only/small_room_3node-predictions.npz +0 -0
  22. studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/thermal_temperature_only/support.json +1237 -0
  23. studies/robustness-seed42-20261011/configs/full-study.json +44 -0
  24. studies/robustness-seed42-20261011/configs/poc.json +9 -0
  25. studies/robustness-seed42-20261011/configs/robustness-study.json +38 -0
  26. studies/robustness-seed42-20261011/configs/smoke.json +6 -0
  27. studies/robustness-seed42-20261011/docs/context-uncertainty.md +125 -0
  28. studies/robustness-seed42-20261011/docs/contracts.md +14 -0
  29. studies/robustness-seed42-20261011/docs/robustness-study.md +29 -0
  30. studies/robustness-seed42-20261011/licenses/CityLearn-3.0.2-LICENSE +21 -0
  31. studies/robustness-seed42-20261011/pyproject.toml +28 -0
  32. studies/robustness-seed42-20261011/release.json +1313 -0
  33. studies/robustness-seed42-20261011/scripts/behavior_support.py +609 -0
  34. studies/robustness-seed42-20261011/scripts/build_robustness_release.py +218 -0
  35. studies/robustness-seed42-20261011/scripts/challenge_interventions.py +404 -0
  36. studies/robustness-seed42-20261011/scripts/citylearn_validation.py +261 -0
  37. studies/robustness-seed42-20261011/scripts/context_uncertainty.py +443 -0
  38. studies/robustness-seed42-20261011/scripts/evaluate_citylearn.py +257 -0
  39. studies/robustness-seed42-20261011/scripts/evaluate_thermal_challenges.py +298 -0
  40. studies/robustness-seed42-20261011/scripts/history_diagnostics.py +394 -0
  41. studies/robustness-seed42-20261011/scripts/thermal_challenges.py +322 -0
  42. studies/robustness-seed42-20261011/tests/test_behavior_support.py +183 -0
  43. studies/robustness-seed42-20261011/tests/test_challenge_interventions.py +120 -0
  44. studies/robustness-seed42-20261011/tests/test_citylearn_validation.py +133 -0
  45. studies/robustness-seed42-20261011/tests/test_context_uncertainty.py +185 -0
  46. studies/robustness-seed42-20261011/tests/test_evaluate_citylearn.py +55 -0
  47. studies/robustness-seed42-20261011/tests/test_history_diagnostics.py +210 -0
  48. studies/robustness-seed42-20261011/tests/test_robustness_release.py +53 -0
  49. studies/robustness-seed42-20261011/tests/test_thermal_challenges.py +122 -0
  50. studies/robustness-seed42-20261011/tests/test_thermal_evaluation.py +160 -0
.gitattributes CHANGED
@@ -11,3 +11,8 @@ studies/robustness-seed42-20261011/artifacts/robustness/history/thermal_temperat
11
  studies/robustness-seed42-20261011/artifacts/robustness/history/truncate_4-test.npz filter=lfs diff=lfs merge=lfs -text
12
  studies/robustness-seed42-20261011/artifacts/robustness/history/truncate_8-test.npz filter=lfs diff=lfs merge=lfs -text
13
  studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/current_only/auditorium_3node-predictions.npz filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
11
  studies/robustness-seed42-20261011/artifacts/robustness/history/truncate_4-test.npz filter=lfs diff=lfs merge=lfs -text
12
  studies/robustness-seed42-20261011/artifacts/robustness/history/truncate_8-test.npz filter=lfs diff=lfs merge=lfs -text
13
  studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/current_only/auditorium_3node-predictions.npz filter=lfs diff=lfs merge=lfs -text
14
+ studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/current_only/parameter_shift_2node-predictions.npz filter=lfs diff=lfs merge=lfs -text
15
+ studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/current_only/small_room_3node-predictions.npz filter=lfs diff=lfs merge=lfs -text
16
+ studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/full_history/auditorium_3node-predictions.npz filter=lfs diff=lfs merge=lfs -text
17
+ studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/full_history/parameter_shift_2node-predictions.npz filter=lfs diff=lfs merge=lfs -text
18
+ studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/full_history/small_room_3node-predictions.npz filter=lfs diff=lfs merge=lfs -text
studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/current_only/parameter_shift_2node-predictions.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5345e1042a1c9d89604671b8892b9b3b9d34b9fc2a835e097f42d2512ce65ea0
3
+ size 121579
studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/current_only/small_room_3node-context.json ADDED
The diff for this file is too large to render. See raw diff
 
studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/current_only/small_room_3node-predictions.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f119ea07d493e7efcec5a79acd21bee815fbe7d1c5a47f0211d5d73b80c24019
3
+ size 122659
studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/current_only/support.json ADDED
@@ -0,0 +1,1167 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "version": "tfm-context-uncertainty-v1",
3
+ "training_system_ids": [
4
+ "battery-00002",
5
+ "battery-00005",
6
+ "battery-00008",
7
+ "battery-00014",
8
+ "battery-00020",
9
+ "battery-00026",
10
+ "battery-00032",
11
+ "battery-00035",
12
+ "battery-00044",
13
+ "battery-00050",
14
+ "battery-00053",
15
+ "battery-00056",
16
+ "battery-00059",
17
+ "battery-00062",
18
+ "battery-00065",
19
+ "battery-00071",
20
+ "battery-00077",
21
+ "battery-00089",
22
+ "battery-00092",
23
+ "battery-00098",
24
+ "battery-00101",
25
+ "battery-00110",
26
+ "battery-00113",
27
+ "battery-00116",
28
+ "battery-00122",
29
+ "battery-00125",
30
+ "battery-00131",
31
+ "battery-00137",
32
+ "battery-00140",
33
+ "battery-00143",
34
+ "battery-00155",
35
+ "battery-00170",
36
+ "battery-00173",
37
+ "battery-00179",
38
+ "battery-00182",
39
+ "battery-00185",
40
+ "battery-00191",
41
+ "battery-00197",
42
+ "battery-00200",
43
+ "battery-00203",
44
+ "battery-00206",
45
+ "battery-00209",
46
+ "battery-00215",
47
+ "battery-00218",
48
+ "battery-00221",
49
+ "battery-00227",
50
+ "battery-00233",
51
+ "battery-00236",
52
+ "battery-00242",
53
+ "battery-00245",
54
+ "battery-00248",
55
+ "battery-00251",
56
+ "battery-00254",
57
+ "battery-00257",
58
+ "battery-00269",
59
+ "battery-00272",
60
+ "battery-00278",
61
+ "battery-00284",
62
+ "battery-00293",
63
+ "battery-00302",
64
+ "battery-00311",
65
+ "battery-00314",
66
+ "battery-00323",
67
+ "battery-00326",
68
+ "battery-00329",
69
+ "battery-00332",
70
+ "battery-00335",
71
+ "battery-00338",
72
+ "battery-00344",
73
+ "battery-00350",
74
+ "battery-00356",
75
+ "battery-00359",
76
+ "battery-00362",
77
+ "battery-00368",
78
+ "battery-00371",
79
+ "battery-00374",
80
+ "battery-00383",
81
+ "battery-00386",
82
+ "battery-00389",
83
+ "battery-00392",
84
+ "battery-00395",
85
+ "battery-00398",
86
+ "battery-00401",
87
+ "battery-00404",
88
+ "battery-00407",
89
+ "battery-00410",
90
+ "battery-00419",
91
+ "battery-00422",
92
+ "battery-00425",
93
+ "battery-00428",
94
+ "battery-00431",
95
+ "battery-00434",
96
+ "battery-00437",
97
+ "battery-00440",
98
+ "battery-00443",
99
+ "battery-00449",
100
+ "battery-00452",
101
+ "battery-00455",
102
+ "battery-00458",
103
+ "battery-00461",
104
+ "battery-00467",
105
+ "battery-00470",
106
+ "battery-00473",
107
+ "battery-00476",
108
+ "battery-00485",
109
+ "battery-00488",
110
+ "battery-00491",
111
+ "battery-00494",
112
+ "battery-00497",
113
+ "battery-00503",
114
+ "battery-00506",
115
+ "battery-00509",
116
+ "battery-00512",
117
+ "battery-00515",
118
+ "battery-00518",
119
+ "battery-00521",
120
+ "battery-00524",
121
+ "battery-00530",
122
+ "battery-00536",
123
+ "battery-00539",
124
+ "battery-00542",
125
+ "battery-00545",
126
+ "battery-00548",
127
+ "battery-00551",
128
+ "battery-00554",
129
+ "battery-00557",
130
+ "battery-00560",
131
+ "battery-00563",
132
+ "battery-00569",
133
+ "battery-00572",
134
+ "battery-00575",
135
+ "battery-00584",
136
+ "battery-00590",
137
+ "battery-00596",
138
+ "battery-00602",
139
+ "battery-00611",
140
+ "battery-00614",
141
+ "battery-00617",
142
+ "battery-00620",
143
+ "battery-00623",
144
+ "battery-00626",
145
+ "battery-00629",
146
+ "battery-00638",
147
+ "battery-00641",
148
+ "battery-00644",
149
+ "battery-00647",
150
+ "battery-00650",
151
+ "battery-00656",
152
+ "battery-00662",
153
+ "battery-00668",
154
+ "battery-00671",
155
+ "battery-00680",
156
+ "battery-00683",
157
+ "battery-00686",
158
+ "battery-00689",
159
+ "battery-00692",
160
+ "battery-00695",
161
+ "battery-00698",
162
+ "battery-00701",
163
+ "battery-00707",
164
+ "battery-00710",
165
+ "battery-00713",
166
+ "battery-00716",
167
+ "battery-00725",
168
+ "battery-00728",
169
+ "battery-00731",
170
+ "battery-00734",
171
+ "battery-00740",
172
+ "battery-00743",
173
+ "battery-00746",
174
+ "battery-00749",
175
+ "battery-00755",
176
+ "battery-00758",
177
+ "battery-00776",
178
+ "battery-00779",
179
+ "battery-00785",
180
+ "battery-00791",
181
+ "battery-00794",
182
+ "battery-00800",
183
+ "battery-00809",
184
+ "battery-00812",
185
+ "battery-00815",
186
+ "battery-00818",
187
+ "battery-00821",
188
+ "battery-00824",
189
+ "battery-00827",
190
+ "battery-00836",
191
+ "battery-00842",
192
+ "battery-00845",
193
+ "battery-00848",
194
+ "battery-00851",
195
+ "battery-00854",
196
+ "battery-00863",
197
+ "battery-00866",
198
+ "battery-00869",
199
+ "battery-00872",
200
+ "battery-00875",
201
+ "battery-00884",
202
+ "battery-00887",
203
+ "battery-00896",
204
+ "battery-00902",
205
+ "battery-00905",
206
+ "battery-00911",
207
+ "battery-00914",
208
+ "battery-00917",
209
+ "battery-00920",
210
+ "battery-00923",
211
+ "battery-00926",
212
+ "battery-00929",
213
+ "battery-00932",
214
+ "battery-00935",
215
+ "battery-00938",
216
+ "battery-00941",
217
+ "battery-00944",
218
+ "battery-00953",
219
+ "battery-00956",
220
+ "battery-00959",
221
+ "battery-00965",
222
+ "battery-00968",
223
+ "battery-00971",
224
+ "battery-00974",
225
+ "battery-00977",
226
+ "battery-00980",
227
+ "battery-00986",
228
+ "battery-00995",
229
+ "battery-00998",
230
+ "battery-01001",
231
+ "battery-01004",
232
+ "battery-01007",
233
+ "battery-01013",
234
+ "battery-01019",
235
+ "battery-01022",
236
+ "battery-01028",
237
+ "battery-01031",
238
+ "battery-01034",
239
+ "battery-01037",
240
+ "battery-01040",
241
+ "battery-01043",
242
+ "battery-01046",
243
+ "battery-01049",
244
+ "battery-01055",
245
+ "battery-01058",
246
+ "battery-01061",
247
+ "battery-01064",
248
+ "battery-01067",
249
+ "battery-01070",
250
+ "battery-01073",
251
+ "battery-01076",
252
+ "battery-01079",
253
+ "battery-01085",
254
+ "battery-01088",
255
+ "battery-01091",
256
+ "battery-01094",
257
+ "battery-01097",
258
+ "battery-01103",
259
+ "battery-01106",
260
+ "battery-01109",
261
+ "battery-01115",
262
+ "battery-01118",
263
+ "battery-01121",
264
+ "battery-01124",
265
+ "battery-01130",
266
+ "battery-01133",
267
+ "battery-01136",
268
+ "battery-01139",
269
+ "battery-01142",
270
+ "battery-01145",
271
+ "battery-01151",
272
+ "battery-01157",
273
+ "battery-01160",
274
+ "battery-01166",
275
+ "battery-01169",
276
+ "battery-01172",
277
+ "battery-01175",
278
+ "battery-01178",
279
+ "battery-01184",
280
+ "battery-01187",
281
+ "battery-01193",
282
+ "battery-01196",
283
+ "battery-01199",
284
+ "tank-00001",
285
+ "tank-00004",
286
+ "tank-00007",
287
+ "tank-00013",
288
+ "tank-00016",
289
+ "tank-00019",
290
+ "tank-00022",
291
+ "tank-00025",
292
+ "tank-00028",
293
+ "tank-00031",
294
+ "tank-00034",
295
+ "tank-00037",
296
+ "tank-00040",
297
+ "tank-00043",
298
+ "tank-00049",
299
+ "tank-00055",
300
+ "tank-00058",
301
+ "tank-00061",
302
+ "tank-00064",
303
+ "tank-00067",
304
+ "tank-00070",
305
+ "tank-00073",
306
+ "tank-00076",
307
+ "tank-00082",
308
+ "tank-00085",
309
+ "tank-00088",
310
+ "tank-00091",
311
+ "tank-00094",
312
+ "tank-00103",
313
+ "tank-00106",
314
+ "tank-00109",
315
+ "tank-00115",
316
+ "tank-00118",
317
+ "tank-00121",
318
+ "tank-00124",
319
+ "tank-00133",
320
+ "tank-00136",
321
+ "tank-00145",
322
+ "tank-00154",
323
+ "tank-00157",
324
+ "tank-00163",
325
+ "tank-00166",
326
+ "tank-00169",
327
+ "tank-00172",
328
+ "tank-00178",
329
+ "tank-00181",
330
+ "tank-00184",
331
+ "tank-00193",
332
+ "tank-00196",
333
+ "tank-00199",
334
+ "tank-00202",
335
+ "tank-00208",
336
+ "tank-00214",
337
+ "tank-00220",
338
+ "tank-00223",
339
+ "tank-00226",
340
+ "tank-00229",
341
+ "tank-00232",
342
+ "tank-00235",
343
+ "tank-00238",
344
+ "tank-00244",
345
+ "tank-00247",
346
+ "tank-00250",
347
+ "tank-00259",
348
+ "tank-00262",
349
+ "tank-00265",
350
+ "tank-00268",
351
+ "tank-00274",
352
+ "tank-00277",
353
+ "tank-00280",
354
+ "tank-00283",
355
+ "tank-00289",
356
+ "tank-00295",
357
+ "tank-00301",
358
+ "tank-00304",
359
+ "tank-00307",
360
+ "tank-00310",
361
+ "tank-00316",
362
+ "tank-00319",
363
+ "tank-00322",
364
+ "tank-00328",
365
+ "tank-00334",
366
+ "tank-00340",
367
+ "tank-00343",
368
+ "tank-00346",
369
+ "tank-00349",
370
+ "tank-00352",
371
+ "tank-00355",
372
+ "tank-00358",
373
+ "tank-00361",
374
+ "tank-00364",
375
+ "tank-00367",
376
+ "tank-00370",
377
+ "tank-00373",
378
+ "tank-00379",
379
+ "tank-00382",
380
+ "tank-00385",
381
+ "tank-00388",
382
+ "tank-00394",
383
+ "tank-00397",
384
+ "tank-00403",
385
+ "tank-00406",
386
+ "tank-00409",
387
+ "tank-00415",
388
+ "tank-00418",
389
+ "tank-00421",
390
+ "tank-00424",
391
+ "tank-00427",
392
+ "tank-00430",
393
+ "tank-00439",
394
+ "tank-00442",
395
+ "tank-00445",
396
+ "tank-00451",
397
+ "tank-00454",
398
+ "tank-00457",
399
+ "tank-00463",
400
+ "tank-00469",
401
+ "tank-00472",
402
+ "tank-00478",
403
+ "tank-00490",
404
+ "tank-00493",
405
+ "tank-00496",
406
+ "tank-00499",
407
+ "tank-00508",
408
+ "tank-00511",
409
+ "tank-00514",
410
+ "tank-00520",
411
+ "tank-00529",
412
+ "tank-00538",
413
+ "tank-00541",
414
+ "tank-00547",
415
+ "tank-00550",
416
+ "tank-00556",
417
+ "tank-00559",
418
+ "tank-00562",
419
+ "tank-00580",
420
+ "tank-00589",
421
+ "tank-00595",
422
+ "tank-00598",
423
+ "tank-00604",
424
+ "tank-00607",
425
+ "tank-00610",
426
+ "tank-00613",
427
+ "tank-00619",
428
+ "tank-00622",
429
+ "tank-00646",
430
+ "tank-00649",
431
+ "tank-00652",
432
+ "tank-00655",
433
+ "tank-00658",
434
+ "tank-00664",
435
+ "tank-00667",
436
+ "tank-00670",
437
+ "tank-00673",
438
+ "tank-00676",
439
+ "tank-00685",
440
+ "tank-00688",
441
+ "tank-00691",
442
+ "tank-00694",
443
+ "tank-00697",
444
+ "tank-00700",
445
+ "tank-00703",
446
+ "tank-00709",
447
+ "tank-00712",
448
+ "tank-00715",
449
+ "tank-00718",
450
+ "tank-00721",
451
+ "tank-00724",
452
+ "tank-00727",
453
+ "tank-00730",
454
+ "tank-00733",
455
+ "tank-00736",
456
+ "tank-00739",
457
+ "tank-00742",
458
+ "tank-00745",
459
+ "tank-00748",
460
+ "tank-00751",
461
+ "tank-00760",
462
+ "tank-00763",
463
+ "tank-00766",
464
+ "tank-00778",
465
+ "tank-00781",
466
+ "tank-00787",
467
+ "tank-00790",
468
+ "tank-00793",
469
+ "tank-00802",
470
+ "tank-00805",
471
+ "tank-00811",
472
+ "tank-00814",
473
+ "tank-00817",
474
+ "tank-00820",
475
+ "tank-00823",
476
+ "tank-00829",
477
+ "tank-00832",
478
+ "tank-00835",
479
+ "tank-00838",
480
+ "tank-00844",
481
+ "tank-00847",
482
+ "tank-00850",
483
+ "tank-00853",
484
+ "tank-00856",
485
+ "tank-00859",
486
+ "tank-00865",
487
+ "tank-00868",
488
+ "tank-00871",
489
+ "tank-00877",
490
+ "tank-00889",
491
+ "tank-00892",
492
+ "tank-00895",
493
+ "tank-00898",
494
+ "tank-00901",
495
+ "tank-00904",
496
+ "tank-00907",
497
+ "tank-00910",
498
+ "tank-00916",
499
+ "tank-00919",
500
+ "tank-00922",
501
+ "tank-00925",
502
+ "tank-00931",
503
+ "tank-00937",
504
+ "tank-00946",
505
+ "tank-00949",
506
+ "tank-00955",
507
+ "tank-00958",
508
+ "tank-00961",
509
+ "tank-00964",
510
+ "tank-00970",
511
+ "tank-00973",
512
+ "tank-00979",
513
+ "tank-00982",
514
+ "tank-00985",
515
+ "tank-00988",
516
+ "tank-00991",
517
+ "tank-00997",
518
+ "tank-01000",
519
+ "tank-01003",
520
+ "tank-01006",
521
+ "tank-01012",
522
+ "tank-01018",
523
+ "tank-01021",
524
+ "tank-01024",
525
+ "tank-01027",
526
+ "tank-01030",
527
+ "tank-01033",
528
+ "tank-01036",
529
+ "tank-01039",
530
+ "tank-01042",
531
+ "tank-01045",
532
+ "tank-01048",
533
+ "tank-01054",
534
+ "tank-01057",
535
+ "tank-01066",
536
+ "tank-01069",
537
+ "tank-01075",
538
+ "tank-01084",
539
+ "tank-01087",
540
+ "tank-01090",
541
+ "tank-01096",
542
+ "tank-01105",
543
+ "tank-01108",
544
+ "tank-01111",
545
+ "tank-01117",
546
+ "tank-01120",
547
+ "tank-01126",
548
+ "tank-01135",
549
+ "tank-01138",
550
+ "tank-01141",
551
+ "tank-01144",
552
+ "tank-01147",
553
+ "tank-01150",
554
+ "tank-01153",
555
+ "tank-01159",
556
+ "tank-01162",
557
+ "tank-01165",
558
+ "tank-01168",
559
+ "tank-01174",
560
+ "tank-01177",
561
+ "tank-01180",
562
+ "tank-01186",
563
+ "tank-01195",
564
+ "thermal-00000",
565
+ "thermal-00006",
566
+ "thermal-00009",
567
+ "thermal-00012",
568
+ "thermal-00021",
569
+ "thermal-00024",
570
+ "thermal-00027",
571
+ "thermal-00033",
572
+ "thermal-00036",
573
+ "thermal-00054",
574
+ "thermal-00057",
575
+ "thermal-00060",
576
+ "thermal-00072",
577
+ "thermal-00075",
578
+ "thermal-00078",
579
+ "thermal-00081",
580
+ "thermal-00084",
581
+ "thermal-00087",
582
+ "thermal-00090",
583
+ "thermal-00096",
584
+ "thermal-00099",
585
+ "thermal-00111",
586
+ "thermal-00114",
587
+ "thermal-00120",
588
+ "thermal-00132",
589
+ "thermal-00135",
590
+ "thermal-00138",
591
+ "thermal-00141",
592
+ "thermal-00144",
593
+ "thermal-00150",
594
+ "thermal-00153",
595
+ "thermal-00156",
596
+ "thermal-00159",
597
+ "thermal-00162",
598
+ "thermal-00165",
599
+ "thermal-00168",
600
+ "thermal-00174",
601
+ "thermal-00177",
602
+ "thermal-00180",
603
+ "thermal-00183",
604
+ "thermal-00186",
605
+ "thermal-00192",
606
+ "thermal-00198",
607
+ "thermal-00201",
608
+ "thermal-00204",
609
+ "thermal-00210",
610
+ "thermal-00216",
611
+ "thermal-00219",
612
+ "thermal-00225",
613
+ "thermal-00231",
614
+ "thermal-00237",
615
+ "thermal-00240",
616
+ "thermal-00243",
617
+ "thermal-00246",
618
+ "thermal-00249",
619
+ "thermal-00252",
620
+ "thermal-00258",
621
+ "thermal-00264",
622
+ "thermal-00267",
623
+ "thermal-00273",
624
+ "thermal-00276",
625
+ "thermal-00279",
626
+ "thermal-00282",
627
+ "thermal-00285",
628
+ "thermal-00288",
629
+ "thermal-00294",
630
+ "thermal-00300",
631
+ "thermal-00303",
632
+ "thermal-00312",
633
+ "thermal-00318",
634
+ "thermal-00321",
635
+ "thermal-00324",
636
+ "thermal-00336",
637
+ "thermal-00345",
638
+ "thermal-00354",
639
+ "thermal-00357",
640
+ "thermal-00360",
641
+ "thermal-00363",
642
+ "thermal-00372",
643
+ "thermal-00378",
644
+ "thermal-00381",
645
+ "thermal-00384",
646
+ "thermal-00390",
647
+ "thermal-00393",
648
+ "thermal-00396",
649
+ "thermal-00399",
650
+ "thermal-00402",
651
+ "thermal-00405",
652
+ "thermal-00414",
653
+ "thermal-00420",
654
+ "thermal-00429",
655
+ "thermal-00432",
656
+ "thermal-00435",
657
+ "thermal-00438",
658
+ "thermal-00441",
659
+ "thermal-00444",
660
+ "thermal-00447",
661
+ "thermal-00450",
662
+ "thermal-00453",
663
+ "thermal-00456",
664
+ "thermal-00459",
665
+ "thermal-00462",
666
+ "thermal-00465",
667
+ "thermal-00468",
668
+ "thermal-00474",
669
+ "thermal-00477",
670
+ "thermal-00483",
671
+ "thermal-00486",
672
+ "thermal-00492",
673
+ "thermal-00495",
674
+ "thermal-00498",
675
+ "thermal-00504",
676
+ "thermal-00507",
677
+ "thermal-00513",
678
+ "thermal-00516",
679
+ "thermal-00519",
680
+ "thermal-00522",
681
+ "thermal-00528",
682
+ "thermal-00534",
683
+ "thermal-00537",
684
+ "thermal-00540",
685
+ "thermal-00543",
686
+ "thermal-00549",
687
+ "thermal-00552",
688
+ "thermal-00555",
689
+ "thermal-00558",
690
+ "thermal-00573",
691
+ "thermal-00576",
692
+ "thermal-00579",
693
+ "thermal-00588",
694
+ "thermal-00591",
695
+ "thermal-00594",
696
+ "thermal-00600",
697
+ "thermal-00603",
698
+ "thermal-00606",
699
+ "thermal-00609",
700
+ "thermal-00612",
701
+ "thermal-00615",
702
+ "thermal-00621",
703
+ "thermal-00627",
704
+ "thermal-00630",
705
+ "thermal-00633",
706
+ "thermal-00639",
707
+ "thermal-00642",
708
+ "thermal-00645",
709
+ "thermal-00648",
710
+ "thermal-00651",
711
+ "thermal-00657",
712
+ "thermal-00666",
713
+ "thermal-00675",
714
+ "thermal-00678",
715
+ "thermal-00681",
716
+ "thermal-00684",
717
+ "thermal-00687",
718
+ "thermal-00690",
719
+ "thermal-00693",
720
+ "thermal-00696",
721
+ "thermal-00699",
722
+ "thermal-00702",
723
+ "thermal-00705",
724
+ "thermal-00708",
725
+ "thermal-00714",
726
+ "thermal-00717",
727
+ "thermal-00723",
728
+ "thermal-00726",
729
+ "thermal-00729",
730
+ "thermal-00732",
731
+ "thermal-00735",
732
+ "thermal-00738",
733
+ "thermal-00741",
734
+ "thermal-00744",
735
+ "thermal-00747",
736
+ "thermal-00750",
737
+ "thermal-00753",
738
+ "thermal-00756",
739
+ "thermal-00759",
740
+ "thermal-00762",
741
+ "thermal-00765",
742
+ "thermal-00768",
743
+ "thermal-00771",
744
+ "thermal-00774",
745
+ "thermal-00777",
746
+ "thermal-00780",
747
+ "thermal-00783",
748
+ "thermal-00786",
749
+ "thermal-00789",
750
+ "thermal-00792",
751
+ "thermal-00795",
752
+ "thermal-00804",
753
+ "thermal-00807",
754
+ "thermal-00810",
755
+ "thermal-00813",
756
+ "thermal-00819",
757
+ "thermal-00828",
758
+ "thermal-00840",
759
+ "thermal-00843",
760
+ "thermal-00846",
761
+ "thermal-00849",
762
+ "thermal-00852",
763
+ "thermal-00855",
764
+ "thermal-00864",
765
+ "thermal-00867",
766
+ "thermal-00873",
767
+ "thermal-00876",
768
+ "thermal-00879",
769
+ "thermal-00882",
770
+ "thermal-00885",
771
+ "thermal-00894",
772
+ "thermal-00897",
773
+ "thermal-00906",
774
+ "thermal-00912",
775
+ "thermal-00915",
776
+ "thermal-00918",
777
+ "thermal-00921",
778
+ "thermal-00930",
779
+ "thermal-00933",
780
+ "thermal-00936",
781
+ "thermal-00954",
782
+ "thermal-00957",
783
+ "thermal-00963",
784
+ "thermal-00966",
785
+ "thermal-00969",
786
+ "thermal-00978",
787
+ "thermal-00981",
788
+ "thermal-00984",
789
+ "thermal-00987",
790
+ "thermal-00990",
791
+ "thermal-00993",
792
+ "thermal-00996",
793
+ "thermal-01002",
794
+ "thermal-01005",
795
+ "thermal-01008",
796
+ "thermal-01011",
797
+ "thermal-01014",
798
+ "thermal-01017",
799
+ "thermal-01020",
800
+ "thermal-01026",
801
+ "thermal-01029",
802
+ "thermal-01032",
803
+ "thermal-01035",
804
+ "thermal-01038",
805
+ "thermal-01041",
806
+ "thermal-01044",
807
+ "thermal-01047",
808
+ "thermal-01056",
809
+ "thermal-01059",
810
+ "thermal-01062",
811
+ "thermal-01065",
812
+ "thermal-01068",
813
+ "thermal-01071",
814
+ "thermal-01074",
815
+ "thermal-01080",
816
+ "thermal-01083",
817
+ "thermal-01086",
818
+ "thermal-01089",
819
+ "thermal-01095",
820
+ "thermal-01104",
821
+ "thermal-01107",
822
+ "thermal-01110",
823
+ "thermal-01113",
824
+ "thermal-01116",
825
+ "thermal-01122",
826
+ "thermal-01125",
827
+ "thermal-01128",
828
+ "thermal-01134",
829
+ "thermal-01140",
830
+ "thermal-01143",
831
+ "thermal-01146",
832
+ "thermal-01149",
833
+ "thermal-01152",
834
+ "thermal-01164",
835
+ "thermal-01167",
836
+ "thermal-01173",
837
+ "thermal-01176",
838
+ "thermal-01179",
839
+ "thermal-01185",
840
+ "thermal-01188",
841
+ "thermal-01191",
842
+ "thermal-01194",
843
+ "thermal-01197"
844
+ ],
845
+ "normalization_sha256": "33b890560ea8c2a881a2a4e125f19e2beb7a69029ec406d4b7e6a13b9880f3e4",
846
+ "transform_id": "history-diagnostics-v1:current_only",
847
+ "policy": {
848
+ "low_action_std": 0.05,
849
+ "weak_driver_std": 0.05,
850
+ "support_std_margin": 0.5,
851
+ "support_min_count": 10
852
+ },
853
+ "entries": {
854
+ "7:history_mean": {
855
+ "min": -1.8345006415060774,
856
+ "max": 1.9300461718252913,
857
+ "count": 161
858
+ },
859
+ "7:history_std": {
860
+ "min": -1e-06,
861
+ "max": 1e-06,
862
+ "count": 161
863
+ },
864
+ "7:current": {
865
+ "min": -1.8345006415060774,
866
+ "max": 1.9300461718252913,
867
+ "count": 161
868
+ },
869
+ "8:history_mean": {
870
+ "min": -2.6074347951387455,
871
+ "max": 7.818217561480909,
872
+ "count": 161
873
+ },
874
+ "8:history_std": {
875
+ "min": -1e-06,
876
+ "max": 1e-06,
877
+ "count": 161
878
+ },
879
+ "8:current": {
880
+ "min": -2.6074347951387455,
881
+ "max": 7.818217561480909,
882
+ "count": 161
883
+ },
884
+ "13:future_mean": {
885
+ "min": -2.238292363341394,
886
+ "max": 2.211721089537683,
887
+ "count": 166
888
+ },
889
+ "13:future_min": {
890
+ "min": -2.2697974004622576,
891
+ "max": 2.2432261266585467,
892
+ "count": 166
893
+ },
894
+ "13:future_max": {
895
+ "min": -2.297491891162325,
896
+ "max": 2.2709206173586143,
897
+ "count": 166
898
+ },
899
+ "19:future_mean": {
900
+ "min": -2.550761560451955,
901
+ "max": 2.681303361904592,
902
+ "count": 166
903
+ },
904
+ "19:future_min": {
905
+ "min": -2.5944152936715605,
906
+ "max": 2.55719195459075,
907
+ "count": 166
908
+ },
909
+ "19:future_max": {
910
+ "min": -2.4807157780203952,
911
+ "max": 2.7571237350974216,
912
+ "count": 166
913
+ },
914
+ "history_dt:span": {
915
+ "min": 344.999999,
916
+ "max": 345.000001,
917
+ "count": 512
918
+ },
919
+ "future_dt:span": {
920
+ "min": 104.999999,
921
+ "max": 105.000001,
922
+ "count": 512
923
+ },
924
+ "0:history_mean": {
925
+ "min": -6.9520964500721725,
926
+ "max": 5.824430453520659,
927
+ "count": 162
928
+ },
929
+ "0:history_std": {
930
+ "min": -1e-06,
931
+ "max": 1e-06,
932
+ "count": 162
933
+ },
934
+ "0:current": {
935
+ "min": -6.9520964500721725,
936
+ "max": 5.824430453520659,
937
+ "count": 162
938
+ },
939
+ "1:history_mean": {
940
+ "min": -2.297187591713113,
941
+ "max": 3.181546355407876,
942
+ "count": 116
943
+ },
944
+ "1:history_std": {
945
+ "min": -1e-06,
946
+ "max": 1e-06,
947
+ "count": 116
948
+ },
949
+ "1:current": {
950
+ "min": -2.297187591713113,
951
+ "max": 3.181546355407876,
952
+ "count": 116
953
+ },
954
+ "2:history_mean": {
955
+ "min": -1.9606586886523507,
956
+ "max": 2.693115730392768,
957
+ "count": 162
958
+ },
959
+ "2:history_std": {
960
+ "min": -1e-06,
961
+ "max": 1e-06,
962
+ "count": 162
963
+ },
964
+ "2:current": {
965
+ "min": -1.9606586886523507,
966
+ "max": 2.693115730392768,
967
+ "count": 162
968
+ },
969
+ "10:future_mean": {
970
+ "min": -2.1179169128433077,
971
+ "max": 2.063192219544968,
972
+ "count": 164
973
+ },
974
+ "10:future_min": {
975
+ "min": -2.1401726013446933,
976
+ "max": 2.0854479080463535,
977
+ "count": 164
978
+ },
979
+ "10:future_max": {
980
+ "min": -2.173832215238251,
981
+ "max": 2.1191075219399114,
982
+ "count": 164
983
+ },
984
+ "14:future_mean": {
985
+ "min": -2.7651439204986037,
986
+ "max": 2.859376241284126,
987
+ "count": 164
988
+ },
989
+ "14:future_min": {
990
+ "min": -2.858736697072939,
991
+ "max": 2.7898151310634174,
992
+ "count": 164
993
+ },
994
+ "14:future_max": {
995
+ "min": -2.602246621295409,
996
+ "max": 2.8956323653902576,
997
+ "count": 164
998
+ },
999
+ "15:future_mean": {
1000
+ "min": -1.2472839059341898,
1001
+ "max": 3.519802898358201,
1002
+ "count": 164
1003
+ },
1004
+ "15:future_min": {
1005
+ "min": -1.1967750033853077,
1006
+ "max": 3.271746560048821,
1007
+ "count": 164
1008
+ },
1009
+ "15:future_max": {
1010
+ "min": -1.2916961727275031,
1011
+ "max": 3.6574930963648935,
1012
+ "count": 164
1013
+ },
1014
+ "16:future_mean": {
1015
+ "min": -1.508917923957451,
1016
+ "max": 2.3405938110653004,
1017
+ "count": 164
1018
+ },
1019
+ "16:future_min": {
1020
+ "min": -1.530090790360272,
1021
+ "max": 2.361766677468121,
1022
+ "count": 164
1023
+ },
1024
+ "16:future_max": {
1025
+ "min": -1.5293852632568663,
1026
+ "max": 2.3854243104980775,
1027
+ "count": 164
1028
+ },
1029
+ "9:history_mean": {
1030
+ "min": -2.2023716167498044,
1031
+ "max": 2.350365538983386,
1032
+ "count": 163
1033
+ },
1034
+ "9:history_std": {
1035
+ "min": -1e-06,
1036
+ "max": 1e-06,
1037
+ "count": 163
1038
+ },
1039
+ "9:current": {
1040
+ "min": -2.2023716167498044,
1041
+ "max": 2.350365538983386,
1042
+ "count": 163
1043
+ },
1044
+ "3:history_mean": {
1045
+ "min": -2.563212187898668,
1046
+ "max": 2.2561561185245833,
1047
+ "count": 176
1048
+ },
1049
+ "3:history_std": {
1050
+ "min": -1e-06,
1051
+ "max": 1e-06,
1052
+ "count": 176
1053
+ },
1054
+ "3:current": {
1055
+ "min": -2.563212187898668,
1056
+ "max": 2.2561561185245833,
1057
+ "count": 176
1058
+ },
1059
+ "4:history_mean": {
1060
+ "min": -3.1131484877348123,
1061
+ "max": 2.602466858553904,
1062
+ "count": 182
1063
+ },
1064
+ "4:history_std": {
1065
+ "min": -1e-06,
1066
+ "max": 1e-06,
1067
+ "count": 182
1068
+ },
1069
+ "4:current": {
1070
+ "min": -3.1131484877348123,
1071
+ "max": 2.602466858553904,
1072
+ "count": 182
1073
+ },
1074
+ "5:history_mean": {
1075
+ "min": -1.480774138844447,
1076
+ "max": 3.880235885060267,
1077
+ "count": 180
1078
+ },
1079
+ "5:history_std": {
1080
+ "min": -1e-06,
1081
+ "max": 1e-06,
1082
+ "count": 180
1083
+ },
1084
+ "5:current": {
1085
+ "min": -1.480774138844447,
1086
+ "max": 3.880235885060267,
1087
+ "count": 180
1088
+ },
1089
+ "6:history_mean": {
1090
+ "min": -1.5452842631124117,
1091
+ "max": 3.9898985543035126,
1092
+ "count": 181
1093
+ },
1094
+ "6:history_std": {
1095
+ "min": -1e-06,
1096
+ "max": 1e-06,
1097
+ "count": 181
1098
+ },
1099
+ "6:current": {
1100
+ "min": -1.5452842631124117,
1101
+ "max": 3.9898985543035126,
1102
+ "count": 181
1103
+ },
1104
+ "11:future_mean": {
1105
+ "min": -2.2535837142599267,
1106
+ "max": 1.642554327739541,
1107
+ "count": 182
1108
+ },
1109
+ "11:future_min": {
1110
+ "min": -2.368155647330075,
1111
+ "max": 1.7571262608096894,
1112
+ "count": 182
1113
+ },
1114
+ "11:future_max": {
1115
+ "min": -2.225419785163575,
1116
+ "max": 1.6143903986431893,
1117
+ "count": 182
1118
+ },
1119
+ "12:future_mean": {
1120
+ "min": -1.6069942850034193,
1121
+ "max": 2.228464140311952,
1122
+ "count": 182
1123
+ },
1124
+ "12:future_min": {
1125
+ "min": -1.6163920905112403,
1126
+ "max": 2.237861945819773,
1127
+ "count": 182
1128
+ },
1129
+ "12:future_max": {
1130
+ "min": -1.6778703322386568,
1131
+ "max": 2.2993401875471893,
1132
+ "count": 182
1133
+ },
1134
+ "17:future_mean": {
1135
+ "min": -1.6140851258041038,
1136
+ "max": 2.1835690020324363,
1137
+ "count": 182
1138
+ },
1139
+ "17:future_min": {
1140
+ "min": -1.991628161844835,
1141
+ "max": 1.8608503259988878,
1142
+ "count": 182
1143
+ },
1144
+ "17:future_max": {
1145
+ "min": -1.4533015714685136,
1146
+ "max": 2.676748834709328,
1147
+ "count": 182
1148
+ },
1149
+ "18:future_mean": {
1150
+ "min": -2.522122064472071,
1151
+ "max": 2.864874044299952,
1152
+ "count": 182
1153
+ },
1154
+ "18:future_min": {
1155
+ "min": -2.7156598317395617,
1156
+ "max": 2.8018741357098986,
1157
+ "count": 182
1158
+ },
1159
+ "18:future_max": {
1160
+ "min": -2.4448202826306225,
1161
+ "max": 2.98024523036927,
1162
+ "count": 182
1163
+ }
1164
+ },
1165
+ "limitation": "Marginal feature-envelope heuristic; no guarantee of detecting hidden physical shift.",
1166
+ "sha256": "2c5c7dbc4cd207aa64c26c042e3ad824e425d6ee0f773624c35680a2852e8977"
1167
+ }
studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/full_history/auditorium_3node-context.json ADDED
The diff for this file is too large to render. See raw diff
 
studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/full_history/auditorium_3node-predictions.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bcb5acbc99ea8f73b65abfb9495c253154ae5e3be7806bf79740c964a77e6079
3
+ size 123912
studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/full_history/calibration.json ADDED
@@ -0,0 +1,2082 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "version": "tfm-context-uncertainty-v1",
3
+ "nominal_coverage": 0.9,
4
+ "min_count": 30,
5
+ "checkpoint_sha256": [
6
+ "41485d6dd0c72f09a9ca67ced83bc4a0725f9cde731276d1672902304675ebec"
7
+ ],
8
+ "normalization_sha256": "33b890560ea8c2a881a2a4e125f19e2beb7a69029ec406d4b7e6a13b9880f3e4",
9
+ "transform_id": "history-diagnostics-v1:full_history",
10
+ "context_policy": {
11
+ "low_action_std": 0.05,
12
+ "weak_driver_std": 0.05,
13
+ "support_std_margin": 0.5,
14
+ "support_min_count": 10
15
+ },
16
+ "calibration_system_ids": [
17
+ "battery-00068",
18
+ "battery-00074",
19
+ "battery-00083",
20
+ "battery-00107",
21
+ "battery-00188",
22
+ "battery-00281",
23
+ "battery-00287",
24
+ "battery-00299",
25
+ "battery-00317",
26
+ "battery-00353",
27
+ "battery-00365",
28
+ "battery-00377",
29
+ "battery-00413",
30
+ "battery-00464",
31
+ "battery-00479",
32
+ "battery-00527",
33
+ "battery-00566",
34
+ "battery-00653",
35
+ "battery-00674",
36
+ "battery-00677",
37
+ "battery-00761",
38
+ "battery-00767",
39
+ "battery-00833",
40
+ "battery-00839",
41
+ "battery-00881",
42
+ "battery-00950",
43
+ "battery-01010",
44
+ "battery-01016",
45
+ "battery-01100",
46
+ "battery-01190",
47
+ "tank-00010",
48
+ "tank-00052",
49
+ "tank-00112",
50
+ "tank-00139",
51
+ "tank-00292",
52
+ "tank-00313",
53
+ "tank-00331",
54
+ "tank-00400",
55
+ "tank-00412",
56
+ "tank-00433",
57
+ "tank-00448",
58
+ "tank-00475",
59
+ "tank-00484",
60
+ "tank-00544",
61
+ "tank-00553",
62
+ "tank-00586",
63
+ "tank-00616",
64
+ "tank-00679",
65
+ "tank-00682",
66
+ "tank-00775",
67
+ "tank-00796",
68
+ "tank-00883",
69
+ "tank-00934",
70
+ "tank-00940",
71
+ "tank-00967",
72
+ "tank-00994",
73
+ "tank-01009",
74
+ "tank-01051",
75
+ "tank-01063",
76
+ "tank-01198",
77
+ "thermal-00039",
78
+ "thermal-00117",
79
+ "thermal-00126",
80
+ "thermal-00189",
81
+ "thermal-00195",
82
+ "thermal-00234",
83
+ "thermal-00261",
84
+ "thermal-00270",
85
+ "thermal-00342",
86
+ "thermal-00366",
87
+ "thermal-00369",
88
+ "thermal-00375",
89
+ "thermal-00510",
90
+ "thermal-00531",
91
+ "thermal-00561",
92
+ "thermal-00567",
93
+ "thermal-00597",
94
+ "thermal-00660",
95
+ "thermal-00663",
96
+ "thermal-00831",
97
+ "thermal-00834",
98
+ "thermal-00924",
99
+ "thermal-00972",
100
+ "thermal-01023",
101
+ "thermal-01050",
102
+ "thermal-01092",
103
+ "thermal-01137",
104
+ "thermal-01158",
105
+ "thermal-01170",
106
+ "thermal-01182"
107
+ ],
108
+ "support_sha256": "81d31bd21a19a37ea8b2728debb836319a108f3083d4851e2ead3b28b6b544af",
109
+ "training_system_ids": [
110
+ "battery-00002",
111
+ "battery-00005",
112
+ "battery-00008",
113
+ "battery-00014",
114
+ "battery-00020",
115
+ "battery-00026",
116
+ "battery-00032",
117
+ "battery-00035",
118
+ "battery-00044",
119
+ "battery-00050",
120
+ "battery-00053",
121
+ "battery-00056",
122
+ "battery-00059",
123
+ "battery-00062",
124
+ "battery-00065",
125
+ "battery-00071",
126
+ "battery-00077",
127
+ "battery-00089",
128
+ "battery-00092",
129
+ "battery-00098",
130
+ "battery-00101",
131
+ "battery-00110",
132
+ "battery-00113",
133
+ "battery-00116",
134
+ "battery-00122",
135
+ "battery-00125",
136
+ "battery-00131",
137
+ "battery-00137",
138
+ "battery-00140",
139
+ "battery-00143",
140
+ "battery-00155",
141
+ "battery-00170",
142
+ "battery-00173",
143
+ "battery-00179",
144
+ "battery-00182",
145
+ "battery-00185",
146
+ "battery-00191",
147
+ "battery-00197",
148
+ "battery-00200",
149
+ "battery-00203",
150
+ "battery-00206",
151
+ "battery-00209",
152
+ "battery-00215",
153
+ "battery-00218",
154
+ "battery-00221",
155
+ "battery-00227",
156
+ "battery-00233",
157
+ "battery-00236",
158
+ "battery-00242",
159
+ "battery-00245",
160
+ "battery-00248",
161
+ "battery-00251",
162
+ "battery-00254",
163
+ "battery-00257",
164
+ "battery-00269",
165
+ "battery-00272",
166
+ "battery-00278",
167
+ "battery-00284",
168
+ "battery-00293",
169
+ "battery-00302",
170
+ "battery-00311",
171
+ "battery-00314",
172
+ "battery-00323",
173
+ "battery-00326",
174
+ "battery-00329",
175
+ "battery-00332",
176
+ "battery-00335",
177
+ "battery-00338",
178
+ "battery-00344",
179
+ "battery-00350",
180
+ "battery-00356",
181
+ "battery-00359",
182
+ "battery-00362",
183
+ "battery-00368",
184
+ "battery-00371",
185
+ "battery-00374",
186
+ "battery-00383",
187
+ "battery-00386",
188
+ "battery-00389",
189
+ "battery-00392",
190
+ "battery-00395",
191
+ "battery-00398",
192
+ "battery-00401",
193
+ "battery-00404",
194
+ "battery-00407",
195
+ "battery-00410",
196
+ "battery-00419",
197
+ "battery-00422",
198
+ "battery-00425",
199
+ "battery-00428",
200
+ "battery-00431",
201
+ "battery-00434",
202
+ "battery-00437",
203
+ "battery-00440",
204
+ "battery-00443",
205
+ "battery-00449",
206
+ "battery-00452",
207
+ "battery-00455",
208
+ "battery-00458",
209
+ "battery-00461",
210
+ "battery-00467",
211
+ "battery-00470",
212
+ "battery-00473",
213
+ "battery-00476",
214
+ "battery-00485",
215
+ "battery-00488",
216
+ "battery-00491",
217
+ "battery-00494",
218
+ "battery-00497",
219
+ "battery-00503",
220
+ "battery-00506",
221
+ "battery-00509",
222
+ "battery-00512",
223
+ "battery-00515",
224
+ "battery-00518",
225
+ "battery-00521",
226
+ "battery-00524",
227
+ "battery-00530",
228
+ "battery-00536",
229
+ "battery-00539",
230
+ "battery-00542",
231
+ "battery-00545",
232
+ "battery-00548",
233
+ "battery-00551",
234
+ "battery-00554",
235
+ "battery-00557",
236
+ "battery-00560",
237
+ "battery-00563",
238
+ "battery-00569",
239
+ "battery-00572",
240
+ "battery-00575",
241
+ "battery-00584",
242
+ "battery-00590",
243
+ "battery-00596",
244
+ "battery-00602",
245
+ "battery-00611",
246
+ "battery-00614",
247
+ "battery-00617",
248
+ "battery-00620",
249
+ "battery-00623",
250
+ "battery-00626",
251
+ "battery-00629",
252
+ "battery-00638",
253
+ "battery-00641",
254
+ "battery-00644",
255
+ "battery-00647",
256
+ "battery-00650",
257
+ "battery-00656",
258
+ "battery-00662",
259
+ "battery-00668",
260
+ "battery-00671",
261
+ "battery-00680",
262
+ "battery-00683",
263
+ "battery-00686",
264
+ "battery-00689",
265
+ "battery-00692",
266
+ "battery-00695",
267
+ "battery-00698",
268
+ "battery-00701",
269
+ "battery-00707",
270
+ "battery-00710",
271
+ "battery-00713",
272
+ "battery-00716",
273
+ "battery-00725",
274
+ "battery-00728",
275
+ "battery-00731",
276
+ "battery-00734",
277
+ "battery-00740",
278
+ "battery-00743",
279
+ "battery-00746",
280
+ "battery-00749",
281
+ "battery-00755",
282
+ "battery-00758",
283
+ "battery-00776",
284
+ "battery-00779",
285
+ "battery-00785",
286
+ "battery-00791",
287
+ "battery-00794",
288
+ "battery-00800",
289
+ "battery-00809",
290
+ "battery-00812",
291
+ "battery-00815",
292
+ "battery-00818",
293
+ "battery-00821",
294
+ "battery-00824",
295
+ "battery-00827",
296
+ "battery-00836",
297
+ "battery-00842",
298
+ "battery-00845",
299
+ "battery-00848",
300
+ "battery-00851",
301
+ "battery-00854",
302
+ "battery-00863",
303
+ "battery-00866",
304
+ "battery-00869",
305
+ "battery-00872",
306
+ "battery-00875",
307
+ "battery-00884",
308
+ "battery-00887",
309
+ "battery-00896",
310
+ "battery-00902",
311
+ "battery-00905",
312
+ "battery-00911",
313
+ "battery-00914",
314
+ "battery-00917",
315
+ "battery-00920",
316
+ "battery-00923",
317
+ "battery-00926",
318
+ "battery-00929",
319
+ "battery-00932",
320
+ "battery-00935",
321
+ "battery-00938",
322
+ "battery-00941",
323
+ "battery-00944",
324
+ "battery-00953",
325
+ "battery-00956",
326
+ "battery-00959",
327
+ "battery-00965",
328
+ "battery-00968",
329
+ "battery-00971",
330
+ "battery-00974",
331
+ "battery-00977",
332
+ "battery-00980",
333
+ "battery-00986",
334
+ "battery-00995",
335
+ "battery-00998",
336
+ "battery-01001",
337
+ "battery-01004",
338
+ "battery-01007",
339
+ "battery-01013",
340
+ "battery-01019",
341
+ "battery-01022",
342
+ "battery-01028",
343
+ "battery-01031",
344
+ "battery-01034",
345
+ "battery-01037",
346
+ "battery-01040",
347
+ "battery-01043",
348
+ "battery-01046",
349
+ "battery-01049",
350
+ "battery-01055",
351
+ "battery-01058",
352
+ "battery-01061",
353
+ "battery-01064",
354
+ "battery-01067",
355
+ "battery-01070",
356
+ "battery-01073",
357
+ "battery-01076",
358
+ "battery-01079",
359
+ "battery-01085",
360
+ "battery-01088",
361
+ "battery-01091",
362
+ "battery-01094",
363
+ "battery-01097",
364
+ "battery-01103",
365
+ "battery-01106",
366
+ "battery-01109",
367
+ "battery-01115",
368
+ "battery-01118",
369
+ "battery-01121",
370
+ "battery-01124",
371
+ "battery-01130",
372
+ "battery-01133",
373
+ "battery-01136",
374
+ "battery-01139",
375
+ "battery-01142",
376
+ "battery-01145",
377
+ "battery-01151",
378
+ "battery-01157",
379
+ "battery-01160",
380
+ "battery-01166",
381
+ "battery-01169",
382
+ "battery-01172",
383
+ "battery-01175",
384
+ "battery-01178",
385
+ "battery-01184",
386
+ "battery-01187",
387
+ "battery-01193",
388
+ "battery-01196",
389
+ "battery-01199",
390
+ "tank-00001",
391
+ "tank-00004",
392
+ "tank-00007",
393
+ "tank-00013",
394
+ "tank-00016",
395
+ "tank-00019",
396
+ "tank-00022",
397
+ "tank-00025",
398
+ "tank-00028",
399
+ "tank-00031",
400
+ "tank-00034",
401
+ "tank-00037",
402
+ "tank-00040",
403
+ "tank-00043",
404
+ "tank-00049",
405
+ "tank-00055",
406
+ "tank-00058",
407
+ "tank-00061",
408
+ "tank-00064",
409
+ "tank-00067",
410
+ "tank-00070",
411
+ "tank-00073",
412
+ "tank-00076",
413
+ "tank-00082",
414
+ "tank-00085",
415
+ "tank-00088",
416
+ "tank-00091",
417
+ "tank-00094",
418
+ "tank-00103",
419
+ "tank-00106",
420
+ "tank-00109",
421
+ "tank-00115",
422
+ "tank-00118",
423
+ "tank-00121",
424
+ "tank-00124",
425
+ "tank-00133",
426
+ "tank-00136",
427
+ "tank-00145",
428
+ "tank-00154",
429
+ "tank-00157",
430
+ "tank-00163",
431
+ "tank-00166",
432
+ "tank-00169",
433
+ "tank-00172",
434
+ "tank-00178",
435
+ "tank-00181",
436
+ "tank-00184",
437
+ "tank-00193",
438
+ "tank-00196",
439
+ "tank-00199",
440
+ "tank-00202",
441
+ "tank-00208",
442
+ "tank-00214",
443
+ "tank-00220",
444
+ "tank-00223",
445
+ "tank-00226",
446
+ "tank-00229",
447
+ "tank-00232",
448
+ "tank-00235",
449
+ "tank-00238",
450
+ "tank-00244",
451
+ "tank-00247",
452
+ "tank-00250",
453
+ "tank-00259",
454
+ "tank-00262",
455
+ "tank-00265",
456
+ "tank-00268",
457
+ "tank-00274",
458
+ "tank-00277",
459
+ "tank-00280",
460
+ "tank-00283",
461
+ "tank-00289",
462
+ "tank-00295",
463
+ "tank-00301",
464
+ "tank-00304",
465
+ "tank-00307",
466
+ "tank-00310",
467
+ "tank-00316",
468
+ "tank-00319",
469
+ "tank-00322",
470
+ "tank-00328",
471
+ "tank-00334",
472
+ "tank-00340",
473
+ "tank-00343",
474
+ "tank-00346",
475
+ "tank-00349",
476
+ "tank-00352",
477
+ "tank-00355",
478
+ "tank-00358",
479
+ "tank-00361",
480
+ "tank-00364",
481
+ "tank-00367",
482
+ "tank-00370",
483
+ "tank-00373",
484
+ "tank-00379",
485
+ "tank-00382",
486
+ "tank-00385",
487
+ "tank-00388",
488
+ "tank-00394",
489
+ "tank-00397",
490
+ "tank-00403",
491
+ "tank-00406",
492
+ "tank-00409",
493
+ "tank-00415",
494
+ "tank-00418",
495
+ "tank-00421",
496
+ "tank-00424",
497
+ "tank-00427",
498
+ "tank-00430",
499
+ "tank-00439",
500
+ "tank-00442",
501
+ "tank-00445",
502
+ "tank-00451",
503
+ "tank-00454",
504
+ "tank-00457",
505
+ "tank-00463",
506
+ "tank-00469",
507
+ "tank-00472",
508
+ "tank-00478",
509
+ "tank-00490",
510
+ "tank-00493",
511
+ "tank-00496",
512
+ "tank-00499",
513
+ "tank-00508",
514
+ "tank-00511",
515
+ "tank-00514",
516
+ "tank-00520",
517
+ "tank-00529",
518
+ "tank-00538",
519
+ "tank-00541",
520
+ "tank-00547",
521
+ "tank-00550",
522
+ "tank-00556",
523
+ "tank-00559",
524
+ "tank-00562",
525
+ "tank-00580",
526
+ "tank-00589",
527
+ "tank-00595",
528
+ "tank-00598",
529
+ "tank-00604",
530
+ "tank-00607",
531
+ "tank-00610",
532
+ "tank-00613",
533
+ "tank-00619",
534
+ "tank-00622",
535
+ "tank-00646",
536
+ "tank-00649",
537
+ "tank-00652",
538
+ "tank-00655",
539
+ "tank-00658",
540
+ "tank-00664",
541
+ "tank-00667",
542
+ "tank-00670",
543
+ "tank-00673",
544
+ "tank-00676",
545
+ "tank-00685",
546
+ "tank-00688",
547
+ "tank-00691",
548
+ "tank-00694",
549
+ "tank-00697",
550
+ "tank-00700",
551
+ "tank-00703",
552
+ "tank-00709",
553
+ "tank-00712",
554
+ "tank-00715",
555
+ "tank-00718",
556
+ "tank-00721",
557
+ "tank-00724",
558
+ "tank-00727",
559
+ "tank-00730",
560
+ "tank-00733",
561
+ "tank-00736",
562
+ "tank-00739",
563
+ "tank-00742",
564
+ "tank-00745",
565
+ "tank-00748",
566
+ "tank-00751",
567
+ "tank-00760",
568
+ "tank-00763",
569
+ "tank-00766",
570
+ "tank-00778",
571
+ "tank-00781",
572
+ "tank-00787",
573
+ "tank-00790",
574
+ "tank-00793",
575
+ "tank-00802",
576
+ "tank-00805",
577
+ "tank-00811",
578
+ "tank-00814",
579
+ "tank-00817",
580
+ "tank-00820",
581
+ "tank-00823",
582
+ "tank-00829",
583
+ "tank-00832",
584
+ "tank-00835",
585
+ "tank-00838",
586
+ "tank-00844",
587
+ "tank-00847",
588
+ "tank-00850",
589
+ "tank-00853",
590
+ "tank-00856",
591
+ "tank-00859",
592
+ "tank-00865",
593
+ "tank-00868",
594
+ "tank-00871",
595
+ "tank-00877",
596
+ "tank-00889",
597
+ "tank-00892",
598
+ "tank-00895",
599
+ "tank-00898",
600
+ "tank-00901",
601
+ "tank-00904",
602
+ "tank-00907",
603
+ "tank-00910",
604
+ "tank-00916",
605
+ "tank-00919",
606
+ "tank-00922",
607
+ "tank-00925",
608
+ "tank-00931",
609
+ "tank-00937",
610
+ "tank-00946",
611
+ "tank-00949",
612
+ "tank-00955",
613
+ "tank-00958",
614
+ "tank-00961",
615
+ "tank-00964",
616
+ "tank-00970",
617
+ "tank-00973",
618
+ "tank-00979",
619
+ "tank-00982",
620
+ "tank-00985",
621
+ "tank-00988",
622
+ "tank-00991",
623
+ "tank-00997",
624
+ "tank-01000",
625
+ "tank-01003",
626
+ "tank-01006",
627
+ "tank-01012",
628
+ "tank-01018",
629
+ "tank-01021",
630
+ "tank-01024",
631
+ "tank-01027",
632
+ "tank-01030",
633
+ "tank-01033",
634
+ "tank-01036",
635
+ "tank-01039",
636
+ "tank-01042",
637
+ "tank-01045",
638
+ "tank-01048",
639
+ "tank-01054",
640
+ "tank-01057",
641
+ "tank-01066",
642
+ "tank-01069",
643
+ "tank-01075",
644
+ "tank-01084",
645
+ "tank-01087",
646
+ "tank-01090",
647
+ "tank-01096",
648
+ "tank-01105",
649
+ "tank-01108",
650
+ "tank-01111",
651
+ "tank-01117",
652
+ "tank-01120",
653
+ "tank-01126",
654
+ "tank-01135",
655
+ "tank-01138",
656
+ "tank-01141",
657
+ "tank-01144",
658
+ "tank-01147",
659
+ "tank-01150",
660
+ "tank-01153",
661
+ "tank-01159",
662
+ "tank-01162",
663
+ "tank-01165",
664
+ "tank-01168",
665
+ "tank-01174",
666
+ "tank-01177",
667
+ "tank-01180",
668
+ "tank-01186",
669
+ "tank-01195",
670
+ "thermal-00000",
671
+ "thermal-00006",
672
+ "thermal-00009",
673
+ "thermal-00012",
674
+ "thermal-00021",
675
+ "thermal-00024",
676
+ "thermal-00027",
677
+ "thermal-00033",
678
+ "thermal-00036",
679
+ "thermal-00054",
680
+ "thermal-00057",
681
+ "thermal-00060",
682
+ "thermal-00072",
683
+ "thermal-00075",
684
+ "thermal-00078",
685
+ "thermal-00081",
686
+ "thermal-00084",
687
+ "thermal-00087",
688
+ "thermal-00090",
689
+ "thermal-00096",
690
+ "thermal-00099",
691
+ "thermal-00111",
692
+ "thermal-00114",
693
+ "thermal-00120",
694
+ "thermal-00132",
695
+ "thermal-00135",
696
+ "thermal-00138",
697
+ "thermal-00141",
698
+ "thermal-00144",
699
+ "thermal-00150",
700
+ "thermal-00153",
701
+ "thermal-00156",
702
+ "thermal-00159",
703
+ "thermal-00162",
704
+ "thermal-00165",
705
+ "thermal-00168",
706
+ "thermal-00174",
707
+ "thermal-00177",
708
+ "thermal-00180",
709
+ "thermal-00183",
710
+ "thermal-00186",
711
+ "thermal-00192",
712
+ "thermal-00198",
713
+ "thermal-00201",
714
+ "thermal-00204",
715
+ "thermal-00210",
716
+ "thermal-00216",
717
+ "thermal-00219",
718
+ "thermal-00225",
719
+ "thermal-00231",
720
+ "thermal-00237",
721
+ "thermal-00240",
722
+ "thermal-00243",
723
+ "thermal-00246",
724
+ "thermal-00249",
725
+ "thermal-00252",
726
+ "thermal-00258",
727
+ "thermal-00264",
728
+ "thermal-00267",
729
+ "thermal-00273",
730
+ "thermal-00276",
731
+ "thermal-00279",
732
+ "thermal-00282",
733
+ "thermal-00285",
734
+ "thermal-00288",
735
+ "thermal-00294",
736
+ "thermal-00300",
737
+ "thermal-00303",
738
+ "thermal-00312",
739
+ "thermal-00318",
740
+ "thermal-00321",
741
+ "thermal-00324",
742
+ "thermal-00336",
743
+ "thermal-00345",
744
+ "thermal-00354",
745
+ "thermal-00357",
746
+ "thermal-00360",
747
+ "thermal-00363",
748
+ "thermal-00372",
749
+ "thermal-00378",
750
+ "thermal-00381",
751
+ "thermal-00384",
752
+ "thermal-00390",
753
+ "thermal-00393",
754
+ "thermal-00396",
755
+ "thermal-00399",
756
+ "thermal-00402",
757
+ "thermal-00405",
758
+ "thermal-00414",
759
+ "thermal-00420",
760
+ "thermal-00429",
761
+ "thermal-00432",
762
+ "thermal-00435",
763
+ "thermal-00438",
764
+ "thermal-00441",
765
+ "thermal-00444",
766
+ "thermal-00447",
767
+ "thermal-00450",
768
+ "thermal-00453",
769
+ "thermal-00456",
770
+ "thermal-00459",
771
+ "thermal-00462",
772
+ "thermal-00465",
773
+ "thermal-00468",
774
+ "thermal-00474",
775
+ "thermal-00477",
776
+ "thermal-00483",
777
+ "thermal-00486",
778
+ "thermal-00492",
779
+ "thermal-00495",
780
+ "thermal-00498",
781
+ "thermal-00504",
782
+ "thermal-00507",
783
+ "thermal-00513",
784
+ "thermal-00516",
785
+ "thermal-00519",
786
+ "thermal-00522",
787
+ "thermal-00528",
788
+ "thermal-00534",
789
+ "thermal-00537",
790
+ "thermal-00540",
791
+ "thermal-00543",
792
+ "thermal-00549",
793
+ "thermal-00552",
794
+ "thermal-00555",
795
+ "thermal-00558",
796
+ "thermal-00573",
797
+ "thermal-00576",
798
+ "thermal-00579",
799
+ "thermal-00588",
800
+ "thermal-00591",
801
+ "thermal-00594",
802
+ "thermal-00600",
803
+ "thermal-00603",
804
+ "thermal-00606",
805
+ "thermal-00609",
806
+ "thermal-00612",
807
+ "thermal-00615",
808
+ "thermal-00621",
809
+ "thermal-00627",
810
+ "thermal-00630",
811
+ "thermal-00633",
812
+ "thermal-00639",
813
+ "thermal-00642",
814
+ "thermal-00645",
815
+ "thermal-00648",
816
+ "thermal-00651",
817
+ "thermal-00657",
818
+ "thermal-00666",
819
+ "thermal-00675",
820
+ "thermal-00678",
821
+ "thermal-00681",
822
+ "thermal-00684",
823
+ "thermal-00687",
824
+ "thermal-00690",
825
+ "thermal-00693",
826
+ "thermal-00696",
827
+ "thermal-00699",
828
+ "thermal-00702",
829
+ "thermal-00705",
830
+ "thermal-00708",
831
+ "thermal-00714",
832
+ "thermal-00717",
833
+ "thermal-00723",
834
+ "thermal-00726",
835
+ "thermal-00729",
836
+ "thermal-00732",
837
+ "thermal-00735",
838
+ "thermal-00738",
839
+ "thermal-00741",
840
+ "thermal-00744",
841
+ "thermal-00747",
842
+ "thermal-00750",
843
+ "thermal-00753",
844
+ "thermal-00756",
845
+ "thermal-00759",
846
+ "thermal-00762",
847
+ "thermal-00765",
848
+ "thermal-00768",
849
+ "thermal-00771",
850
+ "thermal-00774",
851
+ "thermal-00777",
852
+ "thermal-00780",
853
+ "thermal-00783",
854
+ "thermal-00786",
855
+ "thermal-00789",
856
+ "thermal-00792",
857
+ "thermal-00795",
858
+ "thermal-00804",
859
+ "thermal-00807",
860
+ "thermal-00810",
861
+ "thermal-00813",
862
+ "thermal-00819",
863
+ "thermal-00828",
864
+ "thermal-00840",
865
+ "thermal-00843",
866
+ "thermal-00846",
867
+ "thermal-00849",
868
+ "thermal-00852",
869
+ "thermal-00855",
870
+ "thermal-00864",
871
+ "thermal-00867",
872
+ "thermal-00873",
873
+ "thermal-00876",
874
+ "thermal-00879",
875
+ "thermal-00882",
876
+ "thermal-00885",
877
+ "thermal-00894",
878
+ "thermal-00897",
879
+ "thermal-00906",
880
+ "thermal-00912",
881
+ "thermal-00915",
882
+ "thermal-00918",
883
+ "thermal-00921",
884
+ "thermal-00930",
885
+ "thermal-00933",
886
+ "thermal-00936",
887
+ "thermal-00954",
888
+ "thermal-00957",
889
+ "thermal-00963",
890
+ "thermal-00966",
891
+ "thermal-00969",
892
+ "thermal-00978",
893
+ "thermal-00981",
894
+ "thermal-00984",
895
+ "thermal-00987",
896
+ "thermal-00990",
897
+ "thermal-00993",
898
+ "thermal-00996",
899
+ "thermal-01002",
900
+ "thermal-01005",
901
+ "thermal-01008",
902
+ "thermal-01011",
903
+ "thermal-01014",
904
+ "thermal-01017",
905
+ "thermal-01020",
906
+ "thermal-01026",
907
+ "thermal-01029",
908
+ "thermal-01032",
909
+ "thermal-01035",
910
+ "thermal-01038",
911
+ "thermal-01041",
912
+ "thermal-01044",
913
+ "thermal-01047",
914
+ "thermal-01056",
915
+ "thermal-01059",
916
+ "thermal-01062",
917
+ "thermal-01065",
918
+ "thermal-01068",
919
+ "thermal-01071",
920
+ "thermal-01074",
921
+ "thermal-01080",
922
+ "thermal-01083",
923
+ "thermal-01086",
924
+ "thermal-01089",
925
+ "thermal-01095",
926
+ "thermal-01104",
927
+ "thermal-01107",
928
+ "thermal-01110",
929
+ "thermal-01113",
930
+ "thermal-01116",
931
+ "thermal-01122",
932
+ "thermal-01125",
933
+ "thermal-01128",
934
+ "thermal-01134",
935
+ "thermal-01140",
936
+ "thermal-01143",
937
+ "thermal-01146",
938
+ "thermal-01149",
939
+ "thermal-01152",
940
+ "thermal-01164",
941
+ "thermal-01167",
942
+ "thermal-01173",
943
+ "thermal-01176",
944
+ "thermal-01179",
945
+ "thermal-01185",
946
+ "thermal-01188",
947
+ "thermal-01191",
948
+ "thermal-01194",
949
+ "thermal-01197"
950
+ ],
951
+ "validation_selection_system_ids": [
952
+ "battery-00023",
953
+ "battery-00038",
954
+ "battery-00041",
955
+ "battery-00047",
956
+ "battery-00080",
957
+ "battery-00104",
958
+ "battery-00119",
959
+ "battery-00146",
960
+ "battery-00158",
961
+ "battery-00167",
962
+ "battery-00176",
963
+ "battery-00275",
964
+ "battery-00296",
965
+ "battery-00305",
966
+ "battery-00308",
967
+ "battery-00482",
968
+ "battery-00587",
969
+ "battery-00593",
970
+ "battery-00605",
971
+ "battery-00632",
972
+ "battery-00665",
973
+ "battery-00752",
974
+ "battery-00764",
975
+ "battery-00773",
976
+ "battery-00788",
977
+ "battery-00797",
978
+ "battery-00893",
979
+ "battery-00962",
980
+ "battery-01025",
981
+ "battery-01127",
982
+ "tank-00046",
983
+ "tank-00142",
984
+ "tank-00160",
985
+ "tank-00211",
986
+ "tank-00241",
987
+ "tank-00256",
988
+ "tank-00286",
989
+ "tank-00337",
990
+ "tank-00505",
991
+ "tank-00517",
992
+ "tank-00523",
993
+ "tank-00574",
994
+ "tank-00577",
995
+ "tank-00601",
996
+ "tank-00625",
997
+ "tank-00634",
998
+ "tank-00769",
999
+ "tank-00808",
1000
+ "tank-00874",
1001
+ "tank-00913",
1002
+ "tank-00928",
1003
+ "tank-00952",
1004
+ "tank-01078",
1005
+ "tank-01081",
1006
+ "tank-01093",
1007
+ "tank-01099",
1008
+ "tank-01102",
1009
+ "tank-01114",
1010
+ "tank-01129",
1011
+ "tank-01171",
1012
+ "thermal-00003",
1013
+ "thermal-00015",
1014
+ "thermal-00030",
1015
+ "thermal-00051",
1016
+ "thermal-00063",
1017
+ "thermal-00207",
1018
+ "thermal-00213",
1019
+ "thermal-00222",
1020
+ "thermal-00255",
1021
+ "thermal-00297",
1022
+ "thermal-00309",
1023
+ "thermal-00333",
1024
+ "thermal-00348",
1025
+ "thermal-00351",
1026
+ "thermal-00426",
1027
+ "thermal-00471",
1028
+ "thermal-00501",
1029
+ "thermal-00564",
1030
+ "thermal-00618",
1031
+ "thermal-00654",
1032
+ "thermal-00669",
1033
+ "thermal-00672",
1034
+ "thermal-00798",
1035
+ "thermal-00822",
1036
+ "thermal-00903",
1037
+ "thermal-00945",
1038
+ "thermal-01077",
1039
+ "thermal-01101",
1040
+ "thermal-01131",
1041
+ "thermal-01161"
1042
+ ],
1043
+ "test_system_ids": [
1044
+ "battery-00011",
1045
+ "battery-00017",
1046
+ "battery-00029",
1047
+ "battery-00086",
1048
+ "battery-00095",
1049
+ "battery-00128",
1050
+ "battery-00134",
1051
+ "battery-00149",
1052
+ "battery-00152",
1053
+ "battery-00161",
1054
+ "battery-00164",
1055
+ "battery-00194",
1056
+ "battery-00212",
1057
+ "battery-00224",
1058
+ "battery-00230",
1059
+ "battery-00239",
1060
+ "battery-00260",
1061
+ "battery-00263",
1062
+ "battery-00266",
1063
+ "battery-00290",
1064
+ "battery-00320",
1065
+ "battery-00341",
1066
+ "battery-00347",
1067
+ "battery-00380",
1068
+ "battery-00416",
1069
+ "battery-00446",
1070
+ "battery-00500",
1071
+ "battery-00533",
1072
+ "battery-00578",
1073
+ "battery-00581",
1074
+ "battery-00599",
1075
+ "battery-00608",
1076
+ "battery-00635",
1077
+ "battery-00659",
1078
+ "battery-00704",
1079
+ "battery-00719",
1080
+ "battery-00722",
1081
+ "battery-00737",
1082
+ "battery-00770",
1083
+ "battery-00782",
1084
+ "battery-00803",
1085
+ "battery-00806",
1086
+ "battery-00830",
1087
+ "battery-00857",
1088
+ "battery-00860",
1089
+ "battery-00878",
1090
+ "battery-00890",
1091
+ "battery-00899",
1092
+ "battery-00908",
1093
+ "battery-00947",
1094
+ "battery-00983",
1095
+ "battery-00989",
1096
+ "battery-00992",
1097
+ "battery-01052",
1098
+ "battery-01082",
1099
+ "battery-01112",
1100
+ "battery-01148",
1101
+ "battery-01154",
1102
+ "battery-01163",
1103
+ "battery-01181",
1104
+ "challenge-auditorium_3node-1041230",
1105
+ "challenge-auditorium_3node-1041231",
1106
+ "challenge-auditorium_3node-1041232",
1107
+ "challenge-auditorium_3node-1041233",
1108
+ "challenge-auditorium_3node-1041234",
1109
+ "challenge-auditorium_3node-1041235",
1110
+ "challenge-auditorium_3node-1041236",
1111
+ "challenge-auditorium_3node-1041237",
1112
+ "challenge-auditorium_3node-1041238",
1113
+ "challenge-auditorium_3node-1041239",
1114
+ "challenge-auditorium_3node-1041240",
1115
+ "challenge-auditorium_3node-1041241",
1116
+ "challenge-auditorium_3node-1041242",
1117
+ "challenge-auditorium_3node-1041243",
1118
+ "challenge-auditorium_3node-1041244",
1119
+ "challenge-auditorium_3node-1041245",
1120
+ "challenge-auditorium_3node-1041246",
1121
+ "challenge-auditorium_3node-1041247",
1122
+ "challenge-auditorium_3node-1041248",
1123
+ "challenge-auditorium_3node-1041249",
1124
+ "challenge-auditorium_3node-1041250",
1125
+ "challenge-auditorium_3node-1041251",
1126
+ "challenge-auditorium_3node-1041252",
1127
+ "challenge-auditorium_3node-1041253",
1128
+ "challenge-parameter_shift_2node-841230",
1129
+ "challenge-parameter_shift_2node-841231",
1130
+ "challenge-parameter_shift_2node-841232",
1131
+ "challenge-parameter_shift_2node-841233",
1132
+ "challenge-parameter_shift_2node-841234",
1133
+ "challenge-parameter_shift_2node-841235",
1134
+ "challenge-parameter_shift_2node-841236",
1135
+ "challenge-parameter_shift_2node-841237",
1136
+ "challenge-parameter_shift_2node-841238",
1137
+ "challenge-parameter_shift_2node-841239",
1138
+ "challenge-parameter_shift_2node-841240",
1139
+ "challenge-parameter_shift_2node-841241",
1140
+ "challenge-parameter_shift_2node-841242",
1141
+ "challenge-parameter_shift_2node-841243",
1142
+ "challenge-parameter_shift_2node-841244",
1143
+ "challenge-parameter_shift_2node-841245",
1144
+ "challenge-parameter_shift_2node-841246",
1145
+ "challenge-parameter_shift_2node-841247",
1146
+ "challenge-parameter_shift_2node-841248",
1147
+ "challenge-parameter_shift_2node-841249",
1148
+ "challenge-parameter_shift_2node-841250",
1149
+ "challenge-parameter_shift_2node-841251",
1150
+ "challenge-parameter_shift_2node-841252",
1151
+ "challenge-parameter_shift_2node-841253",
1152
+ "challenge-small_room_3node-941230",
1153
+ "challenge-small_room_3node-941231",
1154
+ "challenge-small_room_3node-941232",
1155
+ "challenge-small_room_3node-941233",
1156
+ "challenge-small_room_3node-941234",
1157
+ "challenge-small_room_3node-941235",
1158
+ "challenge-small_room_3node-941236",
1159
+ "challenge-small_room_3node-941237",
1160
+ "challenge-small_room_3node-941238",
1161
+ "challenge-small_room_3node-941239",
1162
+ "challenge-small_room_3node-941240",
1163
+ "challenge-small_room_3node-941241",
1164
+ "challenge-small_room_3node-941242",
1165
+ "challenge-small_room_3node-941243",
1166
+ "challenge-small_room_3node-941244",
1167
+ "challenge-small_room_3node-941245",
1168
+ "challenge-small_room_3node-941246",
1169
+ "challenge-small_room_3node-941247",
1170
+ "challenge-small_room_3node-941248",
1171
+ "challenge-small_room_3node-941249",
1172
+ "challenge-small_room_3node-941250",
1173
+ "challenge-small_room_3node-941251",
1174
+ "challenge-small_room_3node-941252",
1175
+ "challenge-small_room_3node-941253",
1176
+ "tank-00079",
1177
+ "tank-00097",
1178
+ "tank-00100",
1179
+ "tank-00127",
1180
+ "tank-00130",
1181
+ "tank-00148",
1182
+ "tank-00151",
1183
+ "tank-00175",
1184
+ "tank-00187",
1185
+ "tank-00190",
1186
+ "tank-00205",
1187
+ "tank-00217",
1188
+ "tank-00253",
1189
+ "tank-00271",
1190
+ "tank-00298",
1191
+ "tank-00325",
1192
+ "tank-00376",
1193
+ "tank-00391",
1194
+ "tank-00436",
1195
+ "tank-00460",
1196
+ "tank-00466",
1197
+ "tank-00481",
1198
+ "tank-00487",
1199
+ "tank-00502",
1200
+ "tank-00526",
1201
+ "tank-00532",
1202
+ "tank-00535",
1203
+ "tank-00565",
1204
+ "tank-00568",
1205
+ "tank-00571",
1206
+ "tank-00583",
1207
+ "tank-00592",
1208
+ "tank-00628",
1209
+ "tank-00631",
1210
+ "tank-00637",
1211
+ "tank-00640",
1212
+ "tank-00643",
1213
+ "tank-00661",
1214
+ "tank-00706",
1215
+ "tank-00754",
1216
+ "tank-00757",
1217
+ "tank-00772",
1218
+ "tank-00784",
1219
+ "tank-00799",
1220
+ "tank-00826",
1221
+ "tank-00841",
1222
+ "tank-00862",
1223
+ "tank-00880",
1224
+ "tank-00886",
1225
+ "tank-00943",
1226
+ "tank-00976",
1227
+ "tank-01015",
1228
+ "tank-01060",
1229
+ "tank-01072",
1230
+ "tank-01123",
1231
+ "tank-01132",
1232
+ "tank-01156",
1233
+ "tank-01183",
1234
+ "tank-01189",
1235
+ "tank-01192",
1236
+ "thermal-00018",
1237
+ "thermal-00042",
1238
+ "thermal-00045",
1239
+ "thermal-00048",
1240
+ "thermal-00066",
1241
+ "thermal-00069",
1242
+ "thermal-00093",
1243
+ "thermal-00102",
1244
+ "thermal-00105",
1245
+ "thermal-00108",
1246
+ "thermal-00123",
1247
+ "thermal-00129",
1248
+ "thermal-00147",
1249
+ "thermal-00171",
1250
+ "thermal-00228",
1251
+ "thermal-00291",
1252
+ "thermal-00306",
1253
+ "thermal-00315",
1254
+ "thermal-00327",
1255
+ "thermal-00330",
1256
+ "thermal-00339",
1257
+ "thermal-00387",
1258
+ "thermal-00408",
1259
+ "thermal-00411",
1260
+ "thermal-00417",
1261
+ "thermal-00423",
1262
+ "thermal-00480",
1263
+ "thermal-00489",
1264
+ "thermal-00525",
1265
+ "thermal-00546",
1266
+ "thermal-00570",
1267
+ "thermal-00582",
1268
+ "thermal-00585",
1269
+ "thermal-00624",
1270
+ "thermal-00636",
1271
+ "thermal-00711",
1272
+ "thermal-00720",
1273
+ "thermal-00801",
1274
+ "thermal-00816",
1275
+ "thermal-00825",
1276
+ "thermal-00837",
1277
+ "thermal-00858",
1278
+ "thermal-00861",
1279
+ "thermal-00870",
1280
+ "thermal-00888",
1281
+ "thermal-00891",
1282
+ "thermal-00900",
1283
+ "thermal-00909",
1284
+ "thermal-00927",
1285
+ "thermal-00939",
1286
+ "thermal-00942",
1287
+ "thermal-00948",
1288
+ "thermal-00951",
1289
+ "thermal-00960",
1290
+ "thermal-00975",
1291
+ "thermal-00999",
1292
+ "thermal-01053",
1293
+ "thermal-01098",
1294
+ "thermal-01119",
1295
+ "thermal-01155"
1296
+ ],
1297
+ "entries": {
1298
+ "0:15|incomplete_sensors": {
1299
+ "count": 104,
1300
+ "multiplier": 1.5074912309646606
1301
+ },
1302
+ "0:15|incomplete_sensors+outside_training_support": {
1303
+ "count": 4,
1304
+ "multiplier": null
1305
+ },
1306
+ "0:15|low_observed_action_variation": {
1307
+ "count": 1,
1308
+ "multiplier": null
1309
+ },
1310
+ "0:15|observed_context": {
1311
+ "count": 41,
1312
+ "multiplier": 1.4909735918045044
1313
+ },
1314
+ "0:15|outside_training_support": {
1315
+ "count": 1,
1316
+ "multiplier": null
1317
+ },
1318
+ "0:30|incomplete_sensors": {
1319
+ "count": 105,
1320
+ "multiplier": 1.5262303352355957
1321
+ },
1322
+ "0:30|incomplete_sensors+outside_training_support": {
1323
+ "count": 4,
1324
+ "multiplier": null
1325
+ },
1326
+ "0:30|low_observed_action_variation": {
1327
+ "count": 1,
1328
+ "multiplier": null
1329
+ },
1330
+ "0:30|observed_context": {
1331
+ "count": 42,
1332
+ "multiplier": 1.6071088314056396
1333
+ },
1334
+ "0:30|outside_training_support": {
1335
+ "count": 1,
1336
+ "multiplier": null
1337
+ },
1338
+ "0:60|incomplete_sensors": {
1339
+ "count": 104,
1340
+ "multiplier": 1.7405965328216553
1341
+ },
1342
+ "0:60|incomplete_sensors+outside_training_support": {
1343
+ "count": 4,
1344
+ "multiplier": null
1345
+ },
1346
+ "0:60|low_observed_action_variation": {
1347
+ "count": 1,
1348
+ "multiplier": null
1349
+ },
1350
+ "0:60|observed_context": {
1351
+ "count": 41,
1352
+ "multiplier": 1.5666905641555786
1353
+ },
1354
+ "0:60|outside_training_support": {
1355
+ "count": 1,
1356
+ "multiplier": null
1357
+ },
1358
+ "0:120|incomplete_sensors": {
1359
+ "count": 107,
1360
+ "multiplier": 1.5060911178588867
1361
+ },
1362
+ "0:120|incomplete_sensors+outside_training_support": {
1363
+ "count": 4,
1364
+ "multiplier": null
1365
+ },
1366
+ "0:120|low_observed_action_variation": {
1367
+ "count": 1,
1368
+ "multiplier": null
1369
+ },
1370
+ "0:120|observed_context": {
1371
+ "count": 41,
1372
+ "multiplier": 1.825429916381836
1373
+ },
1374
+ "0:120|outside_training_support": {
1375
+ "count": 1,
1376
+ "multiplier": null
1377
+ },
1378
+ "1:15|incomplete_sensors": {
1379
+ "count": 73,
1380
+ "multiplier": 1.4190714359283447
1381
+ },
1382
+ "1:15|incomplete_sensors+outside_training_support": {
1383
+ "count": 4,
1384
+ "multiplier": null
1385
+ },
1386
+ "1:15|low_observed_action_variation": {
1387
+ "count": 1,
1388
+ "multiplier": null
1389
+ },
1390
+ "1:15|observed_context": {
1391
+ "count": 25,
1392
+ "multiplier": null
1393
+ },
1394
+ "1:15|outside_training_support": {
1395
+ "count": 1,
1396
+ "multiplier": null
1397
+ },
1398
+ "1:30|incomplete_sensors": {
1399
+ "count": 75,
1400
+ "multiplier": 1.470940113067627
1401
+ },
1402
+ "1:30|incomplete_sensors+outside_training_support": {
1403
+ "count": 4,
1404
+ "multiplier": null
1405
+ },
1406
+ "1:30|low_observed_action_variation": {
1407
+ "count": 1,
1408
+ "multiplier": null
1409
+ },
1410
+ "1:30|observed_context": {
1411
+ "count": 25,
1412
+ "multiplier": null
1413
+ },
1414
+ "1:30|outside_training_support": {
1415
+ "count": 1,
1416
+ "multiplier": null
1417
+ },
1418
+ "1:60|incomplete_sensors": {
1419
+ "count": 75,
1420
+ "multiplier": 1.9293900728225708
1421
+ },
1422
+ "1:60|incomplete_sensors+outside_training_support": {
1423
+ "count": 4,
1424
+ "multiplier": null
1425
+ },
1426
+ "1:60|low_observed_action_variation": {
1427
+ "count": 1,
1428
+ "multiplier": null
1429
+ },
1430
+ "1:60|observed_context": {
1431
+ "count": 25,
1432
+ "multiplier": null
1433
+ },
1434
+ "1:60|outside_training_support": {
1435
+ "count": 1,
1436
+ "multiplier": null
1437
+ },
1438
+ "1:120|incomplete_sensors": {
1439
+ "count": 71,
1440
+ "multiplier": 2.1270620822906494
1441
+ },
1442
+ "1:120|incomplete_sensors+outside_training_support": {
1443
+ "count": 4,
1444
+ "multiplier": null
1445
+ },
1446
+ "1:120|low_observed_action_variation": {
1447
+ "count": 1,
1448
+ "multiplier": null
1449
+ },
1450
+ "1:120|observed_context": {
1451
+ "count": 24,
1452
+ "multiplier": null
1453
+ },
1454
+ "1:120|outside_training_support": {
1455
+ "count": 1,
1456
+ "multiplier": null
1457
+ },
1458
+ "2:15|incomplete_sensors": {
1459
+ "count": 111,
1460
+ "multiplier": 1.5468010902404785
1461
+ },
1462
+ "2:15|incomplete_sensors+outside_training_support": {
1463
+ "count": 2,
1464
+ "multiplier": null
1465
+ },
1466
+ "2:15|low_observed_action_variation": {
1467
+ "count": 1,
1468
+ "multiplier": null
1469
+ },
1470
+ "2:15|observed_context": {
1471
+ "count": 39,
1472
+ "multiplier": 1.6188727617263794
1473
+ },
1474
+ "2:15|outside_training_support": {
1475
+ "count": 1,
1476
+ "multiplier": null
1477
+ },
1478
+ "2:30|incomplete_sensors": {
1479
+ "count": 110,
1480
+ "multiplier": 1.5038690567016602
1481
+ },
1482
+ "2:30|incomplete_sensors+outside_training_support": {
1483
+ "count": 2,
1484
+ "multiplier": null
1485
+ },
1486
+ "2:30|low_observed_action_variation": {
1487
+ "count": 1,
1488
+ "multiplier": null
1489
+ },
1490
+ "2:30|observed_context": {
1491
+ "count": 41,
1492
+ "multiplier": 1.6709548234939575
1493
+ },
1494
+ "2:30|outside_training_support": {
1495
+ "count": 1,
1496
+ "multiplier": null
1497
+ },
1498
+ "2:60|incomplete_sensors": {
1499
+ "count": 106,
1500
+ "multiplier": 1.695938229560852
1501
+ },
1502
+ "2:60|incomplete_sensors+outside_training_support": {
1503
+ "count": 3,
1504
+ "multiplier": null
1505
+ },
1506
+ "2:60|low_observed_action_variation": {
1507
+ "count": 1,
1508
+ "multiplier": null
1509
+ },
1510
+ "2:60|observed_context": {
1511
+ "count": 42,
1512
+ "multiplier": 1.4785618782043457
1513
+ },
1514
+ "2:60|outside_training_support": {
1515
+ "count": 1,
1516
+ "multiplier": null
1517
+ },
1518
+ "2:120|incomplete_sensors": {
1519
+ "count": 111,
1520
+ "multiplier": 1.7331148386001587
1521
+ },
1522
+ "2:120|incomplete_sensors+outside_training_support": {
1523
+ "count": 3,
1524
+ "multiplier": null
1525
+ },
1526
+ "2:120|low_observed_action_variation": {
1527
+ "count": 1,
1528
+ "multiplier": null
1529
+ },
1530
+ "2:120|observed_context": {
1531
+ "count": 41,
1532
+ "multiplier": 1.4770771265029907
1533
+ },
1534
+ "2:120|outside_training_support": {
1535
+ "count": 1,
1536
+ "multiplier": null
1537
+ },
1538
+ "3:15|incomplete_sensors": {
1539
+ "count": 139,
1540
+ "multiplier": 1.602124810218811
1541
+ },
1542
+ "3:15|incomplete_sensors+outside_training_support": {
1543
+ "count": 1,
1544
+ "multiplier": null
1545
+ },
1546
+ "3:15|observed_context": {
1547
+ "count": 23,
1548
+ "multiplier": null
1549
+ },
1550
+ "3:30|incomplete_sensors": {
1551
+ "count": 138,
1552
+ "multiplier": 1.6081340312957764
1553
+ },
1554
+ "3:30|incomplete_sensors+outside_training_support": {
1555
+ "count": 1,
1556
+ "multiplier": null
1557
+ },
1558
+ "3:30|observed_context": {
1559
+ "count": 22,
1560
+ "multiplier": null
1561
+ },
1562
+ "3:60|incomplete_sensors": {
1563
+ "count": 140,
1564
+ "multiplier": 1.7659860849380493
1565
+ },
1566
+ "3:60|incomplete_sensors+outside_training_support": {
1567
+ "count": 1,
1568
+ "multiplier": null
1569
+ },
1570
+ "3:60|observed_context": {
1571
+ "count": 23,
1572
+ "multiplier": null
1573
+ },
1574
+ "3:120|incomplete_sensors": {
1575
+ "count": 136,
1576
+ "multiplier": 1.4678348302841187
1577
+ },
1578
+ "3:120|observed_context": {
1579
+ "count": 23,
1580
+ "multiplier": null
1581
+ },
1582
+ "4:15|incomplete_sensors": {
1583
+ "count": 137,
1584
+ "multiplier": 1.9247430562973022
1585
+ },
1586
+ "4:15|incomplete_sensors+outside_training_support": {
1587
+ "count": 1,
1588
+ "multiplier": null
1589
+ },
1590
+ "4:15|observed_context": {
1591
+ "count": 23,
1592
+ "multiplier": null
1593
+ },
1594
+ "4:30|incomplete_sensors": {
1595
+ "count": 136,
1596
+ "multiplier": 1.9186359643936157
1597
+ },
1598
+ "4:30|incomplete_sensors+outside_training_support": {
1599
+ "count": 1,
1600
+ "multiplier": null
1601
+ },
1602
+ "4:30|observed_context": {
1603
+ "count": 23,
1604
+ "multiplier": null
1605
+ },
1606
+ "4:60|incomplete_sensors": {
1607
+ "count": 137,
1608
+ "multiplier": 1.886003851890564
1609
+ },
1610
+ "4:60|incomplete_sensors+outside_training_support": {
1611
+ "count": 1,
1612
+ "multiplier": null
1613
+ },
1614
+ "4:60|observed_context": {
1615
+ "count": 23,
1616
+ "multiplier": null
1617
+ },
1618
+ "4:120|incomplete_sensors": {
1619
+ "count": 137,
1620
+ "multiplier": 1.8352510929107666
1621
+ },
1622
+ "4:120|incomplete_sensors+outside_training_support": {
1623
+ "count": 1,
1624
+ "multiplier": null
1625
+ },
1626
+ "4:120|observed_context": {
1627
+ "count": 22,
1628
+ "multiplier": null
1629
+ },
1630
+ "5:15|incomplete_sensors": {
1631
+ "count": 142,
1632
+ "multiplier": 1.2845878601074219
1633
+ },
1634
+ "5:15|incomplete_sensors+outside_training_support": {
1635
+ "count": 1,
1636
+ "multiplier": null
1637
+ },
1638
+ "5:15|observed_context": {
1639
+ "count": 22,
1640
+ "multiplier": null
1641
+ },
1642
+ "5:30|incomplete_sensors": {
1643
+ "count": 142,
1644
+ "multiplier": 1.1014809608459473
1645
+ },
1646
+ "5:30|incomplete_sensors+outside_training_support": {
1647
+ "count": 1,
1648
+ "multiplier": null
1649
+ },
1650
+ "5:30|observed_context": {
1651
+ "count": 21,
1652
+ "multiplier": null
1653
+ },
1654
+ "5:60|incomplete_sensors": {
1655
+ "count": 142,
1656
+ "multiplier": 1.1266205310821533
1657
+ },
1658
+ "5:60|incomplete_sensors+outside_training_support": {
1659
+ "count": 1,
1660
+ "multiplier": null
1661
+ },
1662
+ "5:60|observed_context": {
1663
+ "count": 23,
1664
+ "multiplier": null
1665
+ },
1666
+ "5:120|incomplete_sensors": {
1667
+ "count": 141,
1668
+ "multiplier": 1.3680959939956665
1669
+ },
1670
+ "5:120|incomplete_sensors+outside_training_support": {
1671
+ "count": 1,
1672
+ "multiplier": null
1673
+ },
1674
+ "5:120|observed_context": {
1675
+ "count": 22,
1676
+ "multiplier": null
1677
+ },
1678
+ "6:15|incomplete_sensors": {
1679
+ "count": 140,
1680
+ "multiplier": 1.5263735055923462
1681
+ },
1682
+ "6:15|incomplete_sensors+outside_training_support": {
1683
+ "count": 1,
1684
+ "multiplier": null
1685
+ },
1686
+ "6:15|observed_context": {
1687
+ "count": 23,
1688
+ "multiplier": null
1689
+ },
1690
+ "6:30|incomplete_sensors": {
1691
+ "count": 142,
1692
+ "multiplier": 1.5905961990356445
1693
+ },
1694
+ "6:30|incomplete_sensors+outside_training_support": {
1695
+ "count": 1,
1696
+ "multiplier": null
1697
+ },
1698
+ "6:30|observed_context": {
1699
+ "count": 22,
1700
+ "multiplier": null
1701
+ },
1702
+ "6:60|incomplete_sensors": {
1703
+ "count": 138,
1704
+ "multiplier": 1.6447550058364868
1705
+ },
1706
+ "6:60|incomplete_sensors+outside_training_support": {
1707
+ "count": 1,
1708
+ "multiplier": null
1709
+ },
1710
+ "6:60|observed_context": {
1711
+ "count": 23,
1712
+ "multiplier": null
1713
+ },
1714
+ "6:120|incomplete_sensors": {
1715
+ "count": 141,
1716
+ "multiplier": 1.6278109550476074
1717
+ },
1718
+ "6:120|incomplete_sensors+outside_training_support": {
1719
+ "count": 1,
1720
+ "multiplier": null
1721
+ },
1722
+ "6:120|observed_context": {
1723
+ "count": 22,
1724
+ "multiplier": null
1725
+ },
1726
+ "7:15|incomplete_sensors": {
1727
+ "count": 132,
1728
+ "multiplier": 1.4662518501281738
1729
+ },
1730
+ "7:15|incomplete_sensors+outside_training_support": {
1731
+ "count": 2,
1732
+ "multiplier": null
1733
+ },
1734
+ "7:15|low_observed_action_variation+incomplete_sensors": {
1735
+ "count": 1,
1736
+ "multiplier": null
1737
+ },
1738
+ "7:15|observed_context": {
1739
+ "count": 38,
1740
+ "multiplier": 1.448378086090088
1741
+ },
1742
+ "7:30|incomplete_sensors": {
1743
+ "count": 130,
1744
+ "multiplier": 1.5992043018341064
1745
+ },
1746
+ "7:30|incomplete_sensors+outside_training_support": {
1747
+ "count": 2,
1748
+ "multiplier": null
1749
+ },
1750
+ "7:30|low_observed_action_variation+incomplete_sensors": {
1751
+ "count": 1,
1752
+ "multiplier": null
1753
+ },
1754
+ "7:30|observed_context": {
1755
+ "count": 38,
1756
+ "multiplier": 1.6324751377105713
1757
+ },
1758
+ "7:60|incomplete_sensors": {
1759
+ "count": 131,
1760
+ "multiplier": 1.5418221950531006
1761
+ },
1762
+ "7:60|incomplete_sensors+outside_training_support": {
1763
+ "count": 2,
1764
+ "multiplier": null
1765
+ },
1766
+ "7:60|low_observed_action_variation+incomplete_sensors": {
1767
+ "count": 1,
1768
+ "multiplier": null
1769
+ },
1770
+ "7:60|observed_context": {
1771
+ "count": 38,
1772
+ "multiplier": 1.7413384914398193
1773
+ },
1774
+ "7:120|incomplete_sensors": {
1775
+ "count": 125,
1776
+ "multiplier": 1.5066261291503906
1777
+ },
1778
+ "7:120|incomplete_sensors+outside_training_support": {
1779
+ "count": 1,
1780
+ "multiplier": null
1781
+ },
1782
+ "7:120|low_observed_action_variation+incomplete_sensors": {
1783
+ "count": 1,
1784
+ "multiplier": null
1785
+ },
1786
+ "7:120|observed_context": {
1787
+ "count": 38,
1788
+ "multiplier": 2.2300097942352295
1789
+ },
1790
+ "8:15|incomplete_sensors": {
1791
+ "count": 136,
1792
+ "multiplier": 2.743947744369507
1793
+ },
1794
+ "8:15|incomplete_sensors+outside_training_support": {
1795
+ "count": 2,
1796
+ "multiplier": null
1797
+ },
1798
+ "8:15|low_observed_action_variation+incomplete_sensors": {
1799
+ "count": 1,
1800
+ "multiplier": null
1801
+ },
1802
+ "8:15|observed_context": {
1803
+ "count": 37,
1804
+ "multiplier": 2.638808012008667
1805
+ },
1806
+ "8:30|incomplete_sensors": {
1807
+ "count": 136,
1808
+ "multiplier": 2.544153928756714
1809
+ },
1810
+ "8:30|incomplete_sensors+outside_training_support": {
1811
+ "count": 2,
1812
+ "multiplier": null
1813
+ },
1814
+ "8:30|low_observed_action_variation+incomplete_sensors": {
1815
+ "count": 1,
1816
+ "multiplier": null
1817
+ },
1818
+ "8:30|observed_context": {
1819
+ "count": 35,
1820
+ "multiplier": 2.0292863845825195
1821
+ },
1822
+ "8:60|incomplete_sensors": {
1823
+ "count": 138,
1824
+ "multiplier": 2.4357502460479736
1825
+ },
1826
+ "8:60|incomplete_sensors+outside_training_support": {
1827
+ "count": 2,
1828
+ "multiplier": null
1829
+ },
1830
+ "8:60|low_observed_action_variation+incomplete_sensors": {
1831
+ "count": 1,
1832
+ "multiplier": null
1833
+ },
1834
+ "8:60|observed_context": {
1835
+ "count": 37,
1836
+ "multiplier": 1.743776798248291
1837
+ },
1838
+ "8:120|incomplete_sensors": {
1839
+ "count": 135,
1840
+ "multiplier": 2.4752357006073
1841
+ },
1842
+ "8:120|incomplete_sensors+outside_training_support": {
1843
+ "count": 2,
1844
+ "multiplier": null
1845
+ },
1846
+ "8:120|low_observed_action_variation+incomplete_sensors": {
1847
+ "count": 1,
1848
+ "multiplier": null
1849
+ },
1850
+ "8:120|observed_context": {
1851
+ "count": 36,
1852
+ "multiplier": 1.53176748752594
1853
+ },
1854
+ "9:15|incomplete_sensors": {
1855
+ "count": 130,
1856
+ "multiplier": 1.7122114896774292
1857
+ },
1858
+ "9:15|incomplete_sensors+outside_training_support": {
1859
+ "count": 2,
1860
+ "multiplier": null
1861
+ },
1862
+ "9:15|low_observed_action_variation+incomplete_sensors": {
1863
+ "count": 1,
1864
+ "multiplier": null
1865
+ },
1866
+ "9:15|observed_context": {
1867
+ "count": 37,
1868
+ "multiplier": 1.9634885787963867
1869
+ },
1870
+ "9:30|incomplete_sensors": {
1871
+ "count": 130,
1872
+ "multiplier": 1.6662906408309937
1873
+ },
1874
+ "9:30|incomplete_sensors+outside_training_support": {
1875
+ "count": 2,
1876
+ "multiplier": null
1877
+ },
1878
+ "9:30|low_observed_action_variation+incomplete_sensors": {
1879
+ "count": 1,
1880
+ "multiplier": null
1881
+ },
1882
+ "9:30|observed_context": {
1883
+ "count": 38,
1884
+ "multiplier": 1.8528791666030884
1885
+ },
1886
+ "9:60|incomplete_sensors": {
1887
+ "count": 130,
1888
+ "multiplier": 1.7820677757263184
1889
+ },
1890
+ "9:60|incomplete_sensors+outside_training_support": {
1891
+ "count": 2,
1892
+ "multiplier": null
1893
+ },
1894
+ "9:60|low_observed_action_variation+incomplete_sensors": {
1895
+ "count": 1,
1896
+ "multiplier": null
1897
+ },
1898
+ "9:60|observed_context": {
1899
+ "count": 36,
1900
+ "multiplier": 1.8004403114318848
1901
+ },
1902
+ "9:120|incomplete_sensors": {
1903
+ "count": 132,
1904
+ "multiplier": 1.935365915298462
1905
+ },
1906
+ "9:120|incomplete_sensors+outside_training_support": {
1907
+ "count": 2,
1908
+ "multiplier": null
1909
+ },
1910
+ "9:120|low_observed_action_variation+incomplete_sensors": {
1911
+ "count": 1,
1912
+ "multiplier": null
1913
+ },
1914
+ "9:120|observed_context": {
1915
+ "count": 37,
1916
+ "multiplier": 2.4109134674072266
1917
+ }
1918
+ },
1919
+ "pooled_entries": {
1920
+ "0:15": {
1921
+ "count": 151,
1922
+ "multiplier": 1.4238070249557495
1923
+ },
1924
+ "0:30": {
1925
+ "count": 153,
1926
+ "multiplier": 1.5262303352355957
1927
+ },
1928
+ "0:60": {
1929
+ "count": 151,
1930
+ "multiplier": 1.7154074907302856
1931
+ },
1932
+ "0:120": {
1933
+ "count": 154,
1934
+ "multiplier": 1.5502911806106567
1935
+ },
1936
+ "1:15": {
1937
+ "count": 104,
1938
+ "multiplier": 1.4871869087219238
1939
+ },
1940
+ "1:30": {
1941
+ "count": 106,
1942
+ "multiplier": 1.470940113067627
1943
+ },
1944
+ "1:60": {
1945
+ "count": 106,
1946
+ "multiplier": 1.6826138496398926
1947
+ },
1948
+ "1:120": {
1949
+ "count": 101,
1950
+ "multiplier": 1.9494343996047974
1951
+ },
1952
+ "2:15": {
1953
+ "count": 154,
1954
+ "multiplier": 1.588180422782898
1955
+ },
1956
+ "2:30": {
1957
+ "count": 155,
1958
+ "multiplier": 1.5038690567016602
1959
+ },
1960
+ "2:60": {
1961
+ "count": 153,
1962
+ "multiplier": 1.6207759380340576
1963
+ },
1964
+ "2:120": {
1965
+ "count": 157,
1966
+ "multiplier": 1.6863044500350952
1967
+ },
1968
+ "3:15": {
1969
+ "count": 163,
1970
+ "multiplier": 1.6815403699874878
1971
+ },
1972
+ "3:30": {
1973
+ "count": 161,
1974
+ "multiplier": 1.6081340312957764
1975
+ },
1976
+ "3:60": {
1977
+ "count": 164,
1978
+ "multiplier": 1.7659860849380493
1979
+ },
1980
+ "3:120": {
1981
+ "count": 159,
1982
+ "multiplier": 1.4390878677368164
1983
+ },
1984
+ "4:15": {
1985
+ "count": 161,
1986
+ "multiplier": 1.9247430562973022
1987
+ },
1988
+ "4:30": {
1989
+ "count": 160,
1990
+ "multiplier": 1.8637503385543823
1991
+ },
1992
+ "4:60": {
1993
+ "count": 161,
1994
+ "multiplier": 1.8226490020751953
1995
+ },
1996
+ "4:120": {
1997
+ "count": 160,
1998
+ "multiplier": 1.8352510929107666
1999
+ },
2000
+ "5:15": {
2001
+ "count": 165,
2002
+ "multiplier": 1.2845878601074219
2003
+ },
2004
+ "5:30": {
2005
+ "count": 164,
2006
+ "multiplier": 1.1519505977630615
2007
+ },
2008
+ "5:60": {
2009
+ "count": 166,
2010
+ "multiplier": 1.1387457847595215
2011
+ },
2012
+ "5:120": {
2013
+ "count": 164,
2014
+ "multiplier": 1.3353688716888428
2015
+ },
2016
+ "6:15": {
2017
+ "count": 164,
2018
+ "multiplier": 1.3798800706863403
2019
+ },
2020
+ "6:30": {
2021
+ "count": 165,
2022
+ "multiplier": 1.5757362842559814
2023
+ },
2024
+ "6:60": {
2025
+ "count": 162,
2026
+ "multiplier": 1.58315908908844
2027
+ },
2028
+ "6:120": {
2029
+ "count": 164,
2030
+ "multiplier": 1.5772513151168823
2031
+ },
2032
+ "7:15": {
2033
+ "count": 173,
2034
+ "multiplier": 1.4304349422454834
2035
+ },
2036
+ "7:30": {
2037
+ "count": 171,
2038
+ "multiplier": 1.5780900716781616
2039
+ },
2040
+ "7:60": {
2041
+ "count": 172,
2042
+ "multiplier": 1.5485385656356812
2043
+ },
2044
+ "7:120": {
2045
+ "count": 165,
2046
+ "multiplier": 1.5068869590759277
2047
+ },
2048
+ "8:15": {
2049
+ "count": 176,
2050
+ "multiplier": 2.663092851638794
2051
+ },
2052
+ "8:30": {
2053
+ "count": 174,
2054
+ "multiplier": 2.3664445877075195
2055
+ },
2056
+ "8:60": {
2057
+ "count": 178,
2058
+ "multiplier": 2.1575307846069336
2059
+ },
2060
+ "8:120": {
2061
+ "count": 174,
2062
+ "multiplier": 2.195991277694702
2063
+ },
2064
+ "9:15": {
2065
+ "count": 170,
2066
+ "multiplier": 1.6777907609939575
2067
+ },
2068
+ "9:30": {
2069
+ "count": 171,
2070
+ "multiplier": 1.6662906408309937
2071
+ },
2072
+ "9:60": {
2073
+ "count": 169,
2074
+ "multiplier": 1.7003358602523804
2075
+ },
2076
+ "9:120": {
2077
+ "count": 172,
2078
+ "multiplier": 1.8664488792419434
2079
+ }
2080
+ },
2081
+ "limitation": "Sampled windows are correlated. Finite-sample coverage requires exchangeability; no system-level or shift guarantee. Pooled fallback has no stratum-conditional coverage claim."
2082
+ }
studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/full_history/parameter_shift_2node-context.json ADDED
The diff for this file is too large to render. See raw diff
 
studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/full_history/parameter_shift_2node-predictions.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:01aee73627cde532e0e973b80a1d97b837684e6cae6878756b7d55a1e642b0c0
3
+ size 122761
studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/full_history/small_room_3node-context.json ADDED
The diff for this file is too large to render. See raw diff
 
studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/full_history/small_room_3node-predictions.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9052b77ab6661143e5688260ea8cdd84b5e52bc12b4d1d7d3b751d1e85f34b9e
3
+ size 124068
studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/full_history/support.json ADDED
@@ -0,0 +1,1267 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "version": "tfm-context-uncertainty-v1",
3
+ "training_system_ids": [
4
+ "battery-00002",
5
+ "battery-00005",
6
+ "battery-00008",
7
+ "battery-00014",
8
+ "battery-00020",
9
+ "battery-00026",
10
+ "battery-00032",
11
+ "battery-00035",
12
+ "battery-00044",
13
+ "battery-00050",
14
+ "battery-00053",
15
+ "battery-00056",
16
+ "battery-00059",
17
+ "battery-00062",
18
+ "battery-00065",
19
+ "battery-00071",
20
+ "battery-00077",
21
+ "battery-00089",
22
+ "battery-00092",
23
+ "battery-00098",
24
+ "battery-00101",
25
+ "battery-00110",
26
+ "battery-00113",
27
+ "battery-00116",
28
+ "battery-00122",
29
+ "battery-00125",
30
+ "battery-00131",
31
+ "battery-00137",
32
+ "battery-00140",
33
+ "battery-00143",
34
+ "battery-00155",
35
+ "battery-00170",
36
+ "battery-00173",
37
+ "battery-00179",
38
+ "battery-00182",
39
+ "battery-00185",
40
+ "battery-00191",
41
+ "battery-00197",
42
+ "battery-00200",
43
+ "battery-00203",
44
+ "battery-00206",
45
+ "battery-00209",
46
+ "battery-00215",
47
+ "battery-00218",
48
+ "battery-00221",
49
+ "battery-00227",
50
+ "battery-00233",
51
+ "battery-00236",
52
+ "battery-00242",
53
+ "battery-00245",
54
+ "battery-00248",
55
+ "battery-00251",
56
+ "battery-00254",
57
+ "battery-00257",
58
+ "battery-00269",
59
+ "battery-00272",
60
+ "battery-00278",
61
+ "battery-00284",
62
+ "battery-00293",
63
+ "battery-00302",
64
+ "battery-00311",
65
+ "battery-00314",
66
+ "battery-00323",
67
+ "battery-00326",
68
+ "battery-00329",
69
+ "battery-00332",
70
+ "battery-00335",
71
+ "battery-00338",
72
+ "battery-00344",
73
+ "battery-00350",
74
+ "battery-00356",
75
+ "battery-00359",
76
+ "battery-00362",
77
+ "battery-00368",
78
+ "battery-00371",
79
+ "battery-00374",
80
+ "battery-00383",
81
+ "battery-00386",
82
+ "battery-00389",
83
+ "battery-00392",
84
+ "battery-00395",
85
+ "battery-00398",
86
+ "battery-00401",
87
+ "battery-00404",
88
+ "battery-00407",
89
+ "battery-00410",
90
+ "battery-00419",
91
+ "battery-00422",
92
+ "battery-00425",
93
+ "battery-00428",
94
+ "battery-00431",
95
+ "battery-00434",
96
+ "battery-00437",
97
+ "battery-00440",
98
+ "battery-00443",
99
+ "battery-00449",
100
+ "battery-00452",
101
+ "battery-00455",
102
+ "battery-00458",
103
+ "battery-00461",
104
+ "battery-00467",
105
+ "battery-00470",
106
+ "battery-00473",
107
+ "battery-00476",
108
+ "battery-00485",
109
+ "battery-00488",
110
+ "battery-00491",
111
+ "battery-00494",
112
+ "battery-00497",
113
+ "battery-00503",
114
+ "battery-00506",
115
+ "battery-00509",
116
+ "battery-00512",
117
+ "battery-00515",
118
+ "battery-00518",
119
+ "battery-00521",
120
+ "battery-00524",
121
+ "battery-00530",
122
+ "battery-00536",
123
+ "battery-00539",
124
+ "battery-00542",
125
+ "battery-00545",
126
+ "battery-00548",
127
+ "battery-00551",
128
+ "battery-00554",
129
+ "battery-00557",
130
+ "battery-00560",
131
+ "battery-00563",
132
+ "battery-00569",
133
+ "battery-00572",
134
+ "battery-00575",
135
+ "battery-00584",
136
+ "battery-00590",
137
+ "battery-00596",
138
+ "battery-00602",
139
+ "battery-00611",
140
+ "battery-00614",
141
+ "battery-00617",
142
+ "battery-00620",
143
+ "battery-00623",
144
+ "battery-00626",
145
+ "battery-00629",
146
+ "battery-00638",
147
+ "battery-00641",
148
+ "battery-00644",
149
+ "battery-00647",
150
+ "battery-00650",
151
+ "battery-00656",
152
+ "battery-00662",
153
+ "battery-00668",
154
+ "battery-00671",
155
+ "battery-00680",
156
+ "battery-00683",
157
+ "battery-00686",
158
+ "battery-00689",
159
+ "battery-00692",
160
+ "battery-00695",
161
+ "battery-00698",
162
+ "battery-00701",
163
+ "battery-00707",
164
+ "battery-00710",
165
+ "battery-00713",
166
+ "battery-00716",
167
+ "battery-00725",
168
+ "battery-00728",
169
+ "battery-00731",
170
+ "battery-00734",
171
+ "battery-00740",
172
+ "battery-00743",
173
+ "battery-00746",
174
+ "battery-00749",
175
+ "battery-00755",
176
+ "battery-00758",
177
+ "battery-00776",
178
+ "battery-00779",
179
+ "battery-00785",
180
+ "battery-00791",
181
+ "battery-00794",
182
+ "battery-00800",
183
+ "battery-00809",
184
+ "battery-00812",
185
+ "battery-00815",
186
+ "battery-00818",
187
+ "battery-00821",
188
+ "battery-00824",
189
+ "battery-00827",
190
+ "battery-00836",
191
+ "battery-00842",
192
+ "battery-00845",
193
+ "battery-00848",
194
+ "battery-00851",
195
+ "battery-00854",
196
+ "battery-00863",
197
+ "battery-00866",
198
+ "battery-00869",
199
+ "battery-00872",
200
+ "battery-00875",
201
+ "battery-00884",
202
+ "battery-00887",
203
+ "battery-00896",
204
+ "battery-00902",
205
+ "battery-00905",
206
+ "battery-00911",
207
+ "battery-00914",
208
+ "battery-00917",
209
+ "battery-00920",
210
+ "battery-00923",
211
+ "battery-00926",
212
+ "battery-00929",
213
+ "battery-00932",
214
+ "battery-00935",
215
+ "battery-00938",
216
+ "battery-00941",
217
+ "battery-00944",
218
+ "battery-00953",
219
+ "battery-00956",
220
+ "battery-00959",
221
+ "battery-00965",
222
+ "battery-00968",
223
+ "battery-00971",
224
+ "battery-00974",
225
+ "battery-00977",
226
+ "battery-00980",
227
+ "battery-00986",
228
+ "battery-00995",
229
+ "battery-00998",
230
+ "battery-01001",
231
+ "battery-01004",
232
+ "battery-01007",
233
+ "battery-01013",
234
+ "battery-01019",
235
+ "battery-01022",
236
+ "battery-01028",
237
+ "battery-01031",
238
+ "battery-01034",
239
+ "battery-01037",
240
+ "battery-01040",
241
+ "battery-01043",
242
+ "battery-01046",
243
+ "battery-01049",
244
+ "battery-01055",
245
+ "battery-01058",
246
+ "battery-01061",
247
+ "battery-01064",
248
+ "battery-01067",
249
+ "battery-01070",
250
+ "battery-01073",
251
+ "battery-01076",
252
+ "battery-01079",
253
+ "battery-01085",
254
+ "battery-01088",
255
+ "battery-01091",
256
+ "battery-01094",
257
+ "battery-01097",
258
+ "battery-01103",
259
+ "battery-01106",
260
+ "battery-01109",
261
+ "battery-01115",
262
+ "battery-01118",
263
+ "battery-01121",
264
+ "battery-01124",
265
+ "battery-01130",
266
+ "battery-01133",
267
+ "battery-01136",
268
+ "battery-01139",
269
+ "battery-01142",
270
+ "battery-01145",
271
+ "battery-01151",
272
+ "battery-01157",
273
+ "battery-01160",
274
+ "battery-01166",
275
+ "battery-01169",
276
+ "battery-01172",
277
+ "battery-01175",
278
+ "battery-01178",
279
+ "battery-01184",
280
+ "battery-01187",
281
+ "battery-01193",
282
+ "battery-01196",
283
+ "battery-01199",
284
+ "tank-00001",
285
+ "tank-00004",
286
+ "tank-00007",
287
+ "tank-00013",
288
+ "tank-00016",
289
+ "tank-00019",
290
+ "tank-00022",
291
+ "tank-00025",
292
+ "tank-00028",
293
+ "tank-00031",
294
+ "tank-00034",
295
+ "tank-00037",
296
+ "tank-00040",
297
+ "tank-00043",
298
+ "tank-00049",
299
+ "tank-00055",
300
+ "tank-00058",
301
+ "tank-00061",
302
+ "tank-00064",
303
+ "tank-00067",
304
+ "tank-00070",
305
+ "tank-00073",
306
+ "tank-00076",
307
+ "tank-00082",
308
+ "tank-00085",
309
+ "tank-00088",
310
+ "tank-00091",
311
+ "tank-00094",
312
+ "tank-00103",
313
+ "tank-00106",
314
+ "tank-00109",
315
+ "tank-00115",
316
+ "tank-00118",
317
+ "tank-00121",
318
+ "tank-00124",
319
+ "tank-00133",
320
+ "tank-00136",
321
+ "tank-00145",
322
+ "tank-00154",
323
+ "tank-00157",
324
+ "tank-00163",
325
+ "tank-00166",
326
+ "tank-00169",
327
+ "tank-00172",
328
+ "tank-00178",
329
+ "tank-00181",
330
+ "tank-00184",
331
+ "tank-00193",
332
+ "tank-00196",
333
+ "tank-00199",
334
+ "tank-00202",
335
+ "tank-00208",
336
+ "tank-00214",
337
+ "tank-00220",
338
+ "tank-00223",
339
+ "tank-00226",
340
+ "tank-00229",
341
+ "tank-00232",
342
+ "tank-00235",
343
+ "tank-00238",
344
+ "tank-00244",
345
+ "tank-00247",
346
+ "tank-00250",
347
+ "tank-00259",
348
+ "tank-00262",
349
+ "tank-00265",
350
+ "tank-00268",
351
+ "tank-00274",
352
+ "tank-00277",
353
+ "tank-00280",
354
+ "tank-00283",
355
+ "tank-00289",
356
+ "tank-00295",
357
+ "tank-00301",
358
+ "tank-00304",
359
+ "tank-00307",
360
+ "tank-00310",
361
+ "tank-00316",
362
+ "tank-00319",
363
+ "tank-00322",
364
+ "tank-00328",
365
+ "tank-00334",
366
+ "tank-00340",
367
+ "tank-00343",
368
+ "tank-00346",
369
+ "tank-00349",
370
+ "tank-00352",
371
+ "tank-00355",
372
+ "tank-00358",
373
+ "tank-00361",
374
+ "tank-00364",
375
+ "tank-00367",
376
+ "tank-00370",
377
+ "tank-00373",
378
+ "tank-00379",
379
+ "tank-00382",
380
+ "tank-00385",
381
+ "tank-00388",
382
+ "tank-00394",
383
+ "tank-00397",
384
+ "tank-00403",
385
+ "tank-00406",
386
+ "tank-00409",
387
+ "tank-00415",
388
+ "tank-00418",
389
+ "tank-00421",
390
+ "tank-00424",
391
+ "tank-00427",
392
+ "tank-00430",
393
+ "tank-00439",
394
+ "tank-00442",
395
+ "tank-00445",
396
+ "tank-00451",
397
+ "tank-00454",
398
+ "tank-00457",
399
+ "tank-00463",
400
+ "tank-00469",
401
+ "tank-00472",
402
+ "tank-00478",
403
+ "tank-00490",
404
+ "tank-00493",
405
+ "tank-00496",
406
+ "tank-00499",
407
+ "tank-00508",
408
+ "tank-00511",
409
+ "tank-00514",
410
+ "tank-00520",
411
+ "tank-00529",
412
+ "tank-00538",
413
+ "tank-00541",
414
+ "tank-00547",
415
+ "tank-00550",
416
+ "tank-00556",
417
+ "tank-00559",
418
+ "tank-00562",
419
+ "tank-00580",
420
+ "tank-00589",
421
+ "tank-00595",
422
+ "tank-00598",
423
+ "tank-00604",
424
+ "tank-00607",
425
+ "tank-00610",
426
+ "tank-00613",
427
+ "tank-00619",
428
+ "tank-00622",
429
+ "tank-00646",
430
+ "tank-00649",
431
+ "tank-00652",
432
+ "tank-00655",
433
+ "tank-00658",
434
+ "tank-00664",
435
+ "tank-00667",
436
+ "tank-00670",
437
+ "tank-00673",
438
+ "tank-00676",
439
+ "tank-00685",
440
+ "tank-00688",
441
+ "tank-00691",
442
+ "tank-00694",
443
+ "tank-00697",
444
+ "tank-00700",
445
+ "tank-00703",
446
+ "tank-00709",
447
+ "tank-00712",
448
+ "tank-00715",
449
+ "tank-00718",
450
+ "tank-00721",
451
+ "tank-00724",
452
+ "tank-00727",
453
+ "tank-00730",
454
+ "tank-00733",
455
+ "tank-00736",
456
+ "tank-00739",
457
+ "tank-00742",
458
+ "tank-00745",
459
+ "tank-00748",
460
+ "tank-00751",
461
+ "tank-00760",
462
+ "tank-00763",
463
+ "tank-00766",
464
+ "tank-00778",
465
+ "tank-00781",
466
+ "tank-00787",
467
+ "tank-00790",
468
+ "tank-00793",
469
+ "tank-00802",
470
+ "tank-00805",
471
+ "tank-00811",
472
+ "tank-00814",
473
+ "tank-00817",
474
+ "tank-00820",
475
+ "tank-00823",
476
+ "tank-00829",
477
+ "tank-00832",
478
+ "tank-00835",
479
+ "tank-00838",
480
+ "tank-00844",
481
+ "tank-00847",
482
+ "tank-00850",
483
+ "tank-00853",
484
+ "tank-00856",
485
+ "tank-00859",
486
+ "tank-00865",
487
+ "tank-00868",
488
+ "tank-00871",
489
+ "tank-00877",
490
+ "tank-00889",
491
+ "tank-00892",
492
+ "tank-00895",
493
+ "tank-00898",
494
+ "tank-00901",
495
+ "tank-00904",
496
+ "tank-00907",
497
+ "tank-00910",
498
+ "tank-00916",
499
+ "tank-00919",
500
+ "tank-00922",
501
+ "tank-00925",
502
+ "tank-00931",
503
+ "tank-00937",
504
+ "tank-00946",
505
+ "tank-00949",
506
+ "tank-00955",
507
+ "tank-00958",
508
+ "tank-00961",
509
+ "tank-00964",
510
+ "tank-00970",
511
+ "tank-00973",
512
+ "tank-00979",
513
+ "tank-00982",
514
+ "tank-00985",
515
+ "tank-00988",
516
+ "tank-00991",
517
+ "tank-00997",
518
+ "tank-01000",
519
+ "tank-01003",
520
+ "tank-01006",
521
+ "tank-01012",
522
+ "tank-01018",
523
+ "tank-01021",
524
+ "tank-01024",
525
+ "tank-01027",
526
+ "tank-01030",
527
+ "tank-01033",
528
+ "tank-01036",
529
+ "tank-01039",
530
+ "tank-01042",
531
+ "tank-01045",
532
+ "tank-01048",
533
+ "tank-01054",
534
+ "tank-01057",
535
+ "tank-01066",
536
+ "tank-01069",
537
+ "tank-01075",
538
+ "tank-01084",
539
+ "tank-01087",
540
+ "tank-01090",
541
+ "tank-01096",
542
+ "tank-01105",
543
+ "tank-01108",
544
+ "tank-01111",
545
+ "tank-01117",
546
+ "tank-01120",
547
+ "tank-01126",
548
+ "tank-01135",
549
+ "tank-01138",
550
+ "tank-01141",
551
+ "tank-01144",
552
+ "tank-01147",
553
+ "tank-01150",
554
+ "tank-01153",
555
+ "tank-01159",
556
+ "tank-01162",
557
+ "tank-01165",
558
+ "tank-01168",
559
+ "tank-01174",
560
+ "tank-01177",
561
+ "tank-01180",
562
+ "tank-01186",
563
+ "tank-01195",
564
+ "thermal-00000",
565
+ "thermal-00006",
566
+ "thermal-00009",
567
+ "thermal-00012",
568
+ "thermal-00021",
569
+ "thermal-00024",
570
+ "thermal-00027",
571
+ "thermal-00033",
572
+ "thermal-00036",
573
+ "thermal-00054",
574
+ "thermal-00057",
575
+ "thermal-00060",
576
+ "thermal-00072",
577
+ "thermal-00075",
578
+ "thermal-00078",
579
+ "thermal-00081",
580
+ "thermal-00084",
581
+ "thermal-00087",
582
+ "thermal-00090",
583
+ "thermal-00096",
584
+ "thermal-00099",
585
+ "thermal-00111",
586
+ "thermal-00114",
587
+ "thermal-00120",
588
+ "thermal-00132",
589
+ "thermal-00135",
590
+ "thermal-00138",
591
+ "thermal-00141",
592
+ "thermal-00144",
593
+ "thermal-00150",
594
+ "thermal-00153",
595
+ "thermal-00156",
596
+ "thermal-00159",
597
+ "thermal-00162",
598
+ "thermal-00165",
599
+ "thermal-00168",
600
+ "thermal-00174",
601
+ "thermal-00177",
602
+ "thermal-00180",
603
+ "thermal-00183",
604
+ "thermal-00186",
605
+ "thermal-00192",
606
+ "thermal-00198",
607
+ "thermal-00201",
608
+ "thermal-00204",
609
+ "thermal-00210",
610
+ "thermal-00216",
611
+ "thermal-00219",
612
+ "thermal-00225",
613
+ "thermal-00231",
614
+ "thermal-00237",
615
+ "thermal-00240",
616
+ "thermal-00243",
617
+ "thermal-00246",
618
+ "thermal-00249",
619
+ "thermal-00252",
620
+ "thermal-00258",
621
+ "thermal-00264",
622
+ "thermal-00267",
623
+ "thermal-00273",
624
+ "thermal-00276",
625
+ "thermal-00279",
626
+ "thermal-00282",
627
+ "thermal-00285",
628
+ "thermal-00288",
629
+ "thermal-00294",
630
+ "thermal-00300",
631
+ "thermal-00303",
632
+ "thermal-00312",
633
+ "thermal-00318",
634
+ "thermal-00321",
635
+ "thermal-00324",
636
+ "thermal-00336",
637
+ "thermal-00345",
638
+ "thermal-00354",
639
+ "thermal-00357",
640
+ "thermal-00360",
641
+ "thermal-00363",
642
+ "thermal-00372",
643
+ "thermal-00378",
644
+ "thermal-00381",
645
+ "thermal-00384",
646
+ "thermal-00390",
647
+ "thermal-00393",
648
+ "thermal-00396",
649
+ "thermal-00399",
650
+ "thermal-00402",
651
+ "thermal-00405",
652
+ "thermal-00414",
653
+ "thermal-00420",
654
+ "thermal-00429",
655
+ "thermal-00432",
656
+ "thermal-00435",
657
+ "thermal-00438",
658
+ "thermal-00441",
659
+ "thermal-00444",
660
+ "thermal-00447",
661
+ "thermal-00450",
662
+ "thermal-00453",
663
+ "thermal-00456",
664
+ "thermal-00459",
665
+ "thermal-00462",
666
+ "thermal-00465",
667
+ "thermal-00468",
668
+ "thermal-00474",
669
+ "thermal-00477",
670
+ "thermal-00483",
671
+ "thermal-00486",
672
+ "thermal-00492",
673
+ "thermal-00495",
674
+ "thermal-00498",
675
+ "thermal-00504",
676
+ "thermal-00507",
677
+ "thermal-00513",
678
+ "thermal-00516",
679
+ "thermal-00519",
680
+ "thermal-00522",
681
+ "thermal-00528",
682
+ "thermal-00534",
683
+ "thermal-00537",
684
+ "thermal-00540",
685
+ "thermal-00543",
686
+ "thermal-00549",
687
+ "thermal-00552",
688
+ "thermal-00555",
689
+ "thermal-00558",
690
+ "thermal-00573",
691
+ "thermal-00576",
692
+ "thermal-00579",
693
+ "thermal-00588",
694
+ "thermal-00591",
695
+ "thermal-00594",
696
+ "thermal-00600",
697
+ "thermal-00603",
698
+ "thermal-00606",
699
+ "thermal-00609",
700
+ "thermal-00612",
701
+ "thermal-00615",
702
+ "thermal-00621",
703
+ "thermal-00627",
704
+ "thermal-00630",
705
+ "thermal-00633",
706
+ "thermal-00639",
707
+ "thermal-00642",
708
+ "thermal-00645",
709
+ "thermal-00648",
710
+ "thermal-00651",
711
+ "thermal-00657",
712
+ "thermal-00666",
713
+ "thermal-00675",
714
+ "thermal-00678",
715
+ "thermal-00681",
716
+ "thermal-00684",
717
+ "thermal-00687",
718
+ "thermal-00690",
719
+ "thermal-00693",
720
+ "thermal-00696",
721
+ "thermal-00699",
722
+ "thermal-00702",
723
+ "thermal-00705",
724
+ "thermal-00708",
725
+ "thermal-00714",
726
+ "thermal-00717",
727
+ "thermal-00723",
728
+ "thermal-00726",
729
+ "thermal-00729",
730
+ "thermal-00732",
731
+ "thermal-00735",
732
+ "thermal-00738",
733
+ "thermal-00741",
734
+ "thermal-00744",
735
+ "thermal-00747",
736
+ "thermal-00750",
737
+ "thermal-00753",
738
+ "thermal-00756",
739
+ "thermal-00759",
740
+ "thermal-00762",
741
+ "thermal-00765",
742
+ "thermal-00768",
743
+ "thermal-00771",
744
+ "thermal-00774",
745
+ "thermal-00777",
746
+ "thermal-00780",
747
+ "thermal-00783",
748
+ "thermal-00786",
749
+ "thermal-00789",
750
+ "thermal-00792",
751
+ "thermal-00795",
752
+ "thermal-00804",
753
+ "thermal-00807",
754
+ "thermal-00810",
755
+ "thermal-00813",
756
+ "thermal-00819",
757
+ "thermal-00828",
758
+ "thermal-00840",
759
+ "thermal-00843",
760
+ "thermal-00846",
761
+ "thermal-00849",
762
+ "thermal-00852",
763
+ "thermal-00855",
764
+ "thermal-00864",
765
+ "thermal-00867",
766
+ "thermal-00873",
767
+ "thermal-00876",
768
+ "thermal-00879",
769
+ "thermal-00882",
770
+ "thermal-00885",
771
+ "thermal-00894",
772
+ "thermal-00897",
773
+ "thermal-00906",
774
+ "thermal-00912",
775
+ "thermal-00915",
776
+ "thermal-00918",
777
+ "thermal-00921",
778
+ "thermal-00930",
779
+ "thermal-00933",
780
+ "thermal-00936",
781
+ "thermal-00954",
782
+ "thermal-00957",
783
+ "thermal-00963",
784
+ "thermal-00966",
785
+ "thermal-00969",
786
+ "thermal-00978",
787
+ "thermal-00981",
788
+ "thermal-00984",
789
+ "thermal-00987",
790
+ "thermal-00990",
791
+ "thermal-00993",
792
+ "thermal-00996",
793
+ "thermal-01002",
794
+ "thermal-01005",
795
+ "thermal-01008",
796
+ "thermal-01011",
797
+ "thermal-01014",
798
+ "thermal-01017",
799
+ "thermal-01020",
800
+ "thermal-01026",
801
+ "thermal-01029",
802
+ "thermal-01032",
803
+ "thermal-01035",
804
+ "thermal-01038",
805
+ "thermal-01041",
806
+ "thermal-01044",
807
+ "thermal-01047",
808
+ "thermal-01056",
809
+ "thermal-01059",
810
+ "thermal-01062",
811
+ "thermal-01065",
812
+ "thermal-01068",
813
+ "thermal-01071",
814
+ "thermal-01074",
815
+ "thermal-01080",
816
+ "thermal-01083",
817
+ "thermal-01086",
818
+ "thermal-01089",
819
+ "thermal-01095",
820
+ "thermal-01104",
821
+ "thermal-01107",
822
+ "thermal-01110",
823
+ "thermal-01113",
824
+ "thermal-01116",
825
+ "thermal-01122",
826
+ "thermal-01125",
827
+ "thermal-01128",
828
+ "thermal-01134",
829
+ "thermal-01140",
830
+ "thermal-01143",
831
+ "thermal-01146",
832
+ "thermal-01149",
833
+ "thermal-01152",
834
+ "thermal-01164",
835
+ "thermal-01167",
836
+ "thermal-01173",
837
+ "thermal-01176",
838
+ "thermal-01179",
839
+ "thermal-01185",
840
+ "thermal-01188",
841
+ "thermal-01191",
842
+ "thermal-01194",
843
+ "thermal-01197"
844
+ ],
845
+ "normalization_sha256": "33b890560ea8c2a881a2a4e125f19e2beb7a69029ec406d4b7e6a13b9880f3e4",
846
+ "transform_id": "history-diagnostics-v1:full_history",
847
+ "policy": {
848
+ "low_action_std": 0.05,
849
+ "weak_driver_std": 0.05,
850
+ "support_std_margin": 0.5,
851
+ "support_min_count": 10
852
+ },
853
+ "entries": {
854
+ "7:history_mean": {
855
+ "min": -1.7290941794445034,
856
+ "max": 1.8246395905544277,
857
+ "count": 166
858
+ },
859
+ "7:history_std": {
860
+ "min": -0.14753900378313156,
861
+ "max": 1.3250237545308123,
862
+ "count": 166
863
+ },
864
+ "7:current": {
865
+ "min": -1.8345006415060774,
866
+ "max": 1.9300461718252913,
867
+ "count": 161
868
+ },
869
+ "8:history_mean": {
870
+ "min": -2.7116945162741843,
871
+ "max": 5.300033368202895,
872
+ "count": 166
873
+ },
874
+ "8:history_std": {
875
+ "min": -0.05179354612176362,
876
+ "max": 1.5166016823465103,
877
+ "count": 166
878
+ },
879
+ "8:current": {
880
+ "min": -2.6074347951387455,
881
+ "max": 7.818217561480909,
882
+ "count": 161
883
+ },
884
+ "9:history_mean": {
885
+ "min": -2.056825511125415,
886
+ "max": 2.2097568437136106,
887
+ "count": 166
888
+ },
889
+ "9:history_std": {
890
+ "min": -0.17823313635116123,
891
+ "max": 1.9088055984330703,
892
+ "count": 166
893
+ },
894
+ "13:history_mean": {
895
+ "min": -2.0936771567892185,
896
+ "max": 1.7503260310720552,
897
+ "count": 166
898
+ },
899
+ "13:history_std": {
900
+ "min": -0.1447562429817589,
901
+ "max": 1.683897765702215,
902
+ "count": 166
903
+ },
904
+ "13:future_mean": {
905
+ "min": -2.238292363341394,
906
+ "max": 2.211721089537683,
907
+ "count": 166
908
+ },
909
+ "13:future_min": {
910
+ "min": -2.2697974004622576,
911
+ "max": 2.2432261266585467,
912
+ "count": 166
913
+ },
914
+ "13:future_max": {
915
+ "min": -2.297491891162325,
916
+ "max": 2.2709206173586143,
917
+ "count": 166
918
+ },
919
+ "19:history_mean": {
920
+ "min": -2.512724211988596,
921
+ "max": 2.6421154053784885,
922
+ "count": 166
923
+ },
924
+ "19:history_std": {
925
+ "min": -0.01629896148844199,
926
+ "max": 0.6177024400214716,
927
+ "count": 166
928
+ },
929
+ "19:future_mean": {
930
+ "min": -2.550761560451955,
931
+ "max": 2.681303361904592,
932
+ "count": 166
933
+ },
934
+ "19:future_min": {
935
+ "min": -2.5944152936715605,
936
+ "max": 2.55719195459075,
937
+ "count": 166
938
+ },
939
+ "19:future_max": {
940
+ "min": -2.4807157780203952,
941
+ "max": 2.7571237350974216,
942
+ "count": 166
943
+ },
944
+ "history_dt:span": {
945
+ "min": 344.999999,
946
+ "max": 345.000001,
947
+ "count": 512
948
+ },
949
+ "future_dt:span": {
950
+ "min": 104.999999,
951
+ "max": 105.000001,
952
+ "count": 512
953
+ },
954
+ "0:history_mean": {
955
+ "min": -6.85931327257949,
956
+ "max": 3.8700592747767555,
957
+ "count": 164
958
+ },
959
+ "0:history_std": {
960
+ "min": -0.14812077248330266,
961
+ "max": 3.648339465650274,
962
+ "count": 164
963
+ },
964
+ "0:current": {
965
+ "min": -6.9520964500721725,
966
+ "max": 5.824430453520659,
967
+ "count": 162
968
+ },
969
+ "1:history_mean": {
970
+ "min": -2.35983811478255,
971
+ "max": 3.2966424093687317,
972
+ "count": 118
973
+ },
974
+ "1:history_std": {
975
+ "min": -0.03135561728577398,
976
+ "max": 0.4052086208919742,
977
+ "count": 118
978
+ },
979
+ "1:current": {
980
+ "min": -2.297187591713113,
981
+ "max": 3.181546355407876,
982
+ "count": 116
983
+ },
984
+ "2:history_mean": {
985
+ "min": -2.005037231719855,
986
+ "max": 2.6499844029310977,
987
+ "count": 164
988
+ },
989
+ "2:history_std": {
990
+ "min": -0.0028785374531488366,
991
+ "max": 0.09490449452343988,
992
+ "count": 164
993
+ },
994
+ "2:current": {
995
+ "min": -1.9606586886523507,
996
+ "max": 2.693115730392768,
997
+ "count": 162
998
+ },
999
+ "10:history_mean": {
1000
+ "min": -1.740817883887489,
1001
+ "max": 1.853309133925636,
1002
+ "count": 164
1003
+ },
1004
+ "10:history_std": {
1005
+ "min": -0.0890423621527541,
1006
+ "max": 1.6749548133007872,
1007
+ "count": 164
1008
+ },
1009
+ "10:future_mean": {
1010
+ "min": -2.1179169128433077,
1011
+ "max": 2.063192219544968,
1012
+ "count": 164
1013
+ },
1014
+ "10:future_min": {
1015
+ "min": -2.1401726013446933,
1016
+ "max": 2.0854479080463535,
1017
+ "count": 164
1018
+ },
1019
+ "10:future_max": {
1020
+ "min": -2.173832215238251,
1021
+ "max": 2.1191075219399114,
1022
+ "count": 164
1023
+ },
1024
+ "14:history_mean": {
1025
+ "min": -2.689308242364716,
1026
+ "max": 3.0338695807893985,
1027
+ "count": 164
1028
+ },
1029
+ "14:history_std": {
1030
+ "min": -0.018989615475842492,
1031
+ "max": 0.7289486307260484,
1032
+ "count": 164
1033
+ },
1034
+ "14:future_mean": {
1035
+ "min": -2.7651439204986037,
1036
+ "max": 2.859376241284126,
1037
+ "count": 164
1038
+ },
1039
+ "14:future_min": {
1040
+ "min": -2.858736697072939,
1041
+ "max": 2.7898151310634174,
1042
+ "count": 164
1043
+ },
1044
+ "14:future_max": {
1045
+ "min": -2.602246621295409,
1046
+ "max": 2.8956323653902576,
1047
+ "count": 164
1048
+ },
1049
+ "15:history_mean": {
1050
+ "min": -1.1876729007038567,
1051
+ "max": 3.167044984844634,
1052
+ "count": 164
1053
+ },
1054
+ "15:history_std": {
1055
+ "min": -0.1774676965720896,
1056
+ "max": 1.407122425938392,
1057
+ "count": 164
1058
+ },
1059
+ "15:future_mean": {
1060
+ "min": -1.2472839059341898,
1061
+ "max": 3.519802898358201,
1062
+ "count": 164
1063
+ },
1064
+ "15:future_min": {
1065
+ "min": -1.1967750033853077,
1066
+ "max": 3.271746560048821,
1067
+ "count": 164
1068
+ },
1069
+ "15:future_max": {
1070
+ "min": -1.2916961727275031,
1071
+ "max": 3.6574930963648935,
1072
+ "count": 164
1073
+ },
1074
+ "16:history_mean": {
1075
+ "min": -1.4585570580650746,
1076
+ "max": 2.278922486989635,
1077
+ "count": 164
1078
+ },
1079
+ "16:history_std": {
1080
+ "min": -0.20281806972541155,
1081
+ "max": 1.6377656796268398,
1082
+ "count": 164
1083
+ },
1084
+ "16:future_mean": {
1085
+ "min": -1.508917923957451,
1086
+ "max": 2.3405938110653004,
1087
+ "count": 164
1088
+ },
1089
+ "16:future_min": {
1090
+ "min": -1.530090790360272,
1091
+ "max": 2.361766677468121,
1092
+ "count": 164
1093
+ },
1094
+ "16:future_max": {
1095
+ "min": -1.5293852632568663,
1096
+ "max": 2.3854243104980775,
1097
+ "count": 164
1098
+ },
1099
+ "9:current": {
1100
+ "min": -2.2023716167498044,
1101
+ "max": 2.350365538983386,
1102
+ "count": 163
1103
+ },
1104
+ "3:history_mean": {
1105
+ "min": -2.2399926701024677,
1106
+ "max": 2.2418771305516865,
1107
+ "count": 182
1108
+ },
1109
+ "3:history_std": {
1110
+ "min": -0.05676209614502155,
1111
+ "max": 0.5641803710244772,
1112
+ "count": 182
1113
+ },
1114
+ "3:current": {
1115
+ "min": -2.563212187898668,
1116
+ "max": 2.2561561185245833,
1117
+ "count": 176
1118
+ },
1119
+ "4:history_mean": {
1120
+ "min": -3.2996764706891586,
1121
+ "max": 3.761055760316234,
1122
+ "count": 182
1123
+ },
1124
+ "4:history_std": {
1125
+ "min": -0.04884014944697704,
1126
+ "max": 0.9120465918924936,
1127
+ "count": 182
1128
+ },
1129
+ "4:current": {
1130
+ "min": -3.1131484877348123,
1131
+ "max": 2.602466858553904,
1132
+ "count": 182
1133
+ },
1134
+ "5:history_mean": {
1135
+ "min": -1.1944803404030346,
1136
+ "max": 2.729539375227406,
1137
+ "count": 182
1138
+ },
1139
+ "5:history_std": {
1140
+ "min": -0.08287015431103012,
1141
+ "max": 1.9152921351092747,
1142
+ "count": 182
1143
+ },
1144
+ "5:current": {
1145
+ "min": -1.480774138844447,
1146
+ "max": 3.880235885060267,
1147
+ "count": 180
1148
+ },
1149
+ "6:history_mean": {
1150
+ "min": -1.1502409290800177,
1151
+ "max": 2.670196266127595,
1152
+ "count": 182
1153
+ },
1154
+ "6:history_std": {
1155
+ "min": -0.19670381183905647,
1156
+ "max": 2.0513433881214085,
1157
+ "count": 182
1158
+ },
1159
+ "6:current": {
1160
+ "min": -1.5452842631124117,
1161
+ "max": 3.9898985543035126,
1162
+ "count": 181
1163
+ },
1164
+ "11:history_mean": {
1165
+ "min": -1.6990408255271952,
1166
+ "max": 1.4302197052650492,
1167
+ "count": 182
1168
+ },
1169
+ "11:history_std": {
1170
+ "min": -0.16567403674837314,
1171
+ "max": 1.5934025704931851,
1172
+ "count": 182
1173
+ },
1174
+ "11:future_mean": {
1175
+ "min": -2.2535837142599267,
1176
+ "max": 1.642554327739541,
1177
+ "count": 182
1178
+ },
1179
+ "11:future_min": {
1180
+ "min": -2.368155647330075,
1181
+ "max": 1.7571262608096894,
1182
+ "count": 182
1183
+ },
1184
+ "11:future_max": {
1185
+ "min": -2.225419785163575,
1186
+ "max": 1.6143903986431893,
1187
+ "count": 182
1188
+ },
1189
+ "12:history_mean": {
1190
+ "min": -1.4981196618844557,
1191
+ "max": 2.06382686861053,
1192
+ "count": 182
1193
+ },
1194
+ "12:history_std": {
1195
+ "min": -0.15577708269347942,
1196
+ "max": 1.5672527411102275,
1197
+ "count": 182
1198
+ },
1199
+ "12:future_mean": {
1200
+ "min": -1.6069942850034193,
1201
+ "max": 2.228464140311952,
1202
+ "count": 182
1203
+ },
1204
+ "12:future_min": {
1205
+ "min": -1.6163920905112403,
1206
+ "max": 2.237861945819773,
1207
+ "count": 182
1208
+ },
1209
+ "12:future_max": {
1210
+ "min": -1.6778703322386568,
1211
+ "max": 2.2993401875471893,
1212
+ "count": 182
1213
+ },
1214
+ "17:history_mean": {
1215
+ "min": -1.326923140336929,
1216
+ "max": 1.4096587980288846,
1217
+ "count": 182
1218
+ },
1219
+ "17:history_std": {
1220
+ "min": 0.06630291095130086,
1221
+ "max": 1.2772367740596091,
1222
+ "count": 182
1223
+ },
1224
+ "17:future_mean": {
1225
+ "min": -1.6140851258041038,
1226
+ "max": 2.1835690020324363,
1227
+ "count": 182
1228
+ },
1229
+ "17:future_min": {
1230
+ "min": -1.991628161844835,
1231
+ "max": 1.8608503259988878,
1232
+ "count": 182
1233
+ },
1234
+ "17:future_max": {
1235
+ "min": -1.4533015714685136,
1236
+ "max": 2.676748834709328,
1237
+ "count": 182
1238
+ },
1239
+ "18:history_mean": {
1240
+ "min": -2.8351903908626053,
1241
+ "max": 2.4837180130855057,
1242
+ "count": 182
1243
+ },
1244
+ "18:history_std": {
1245
+ "min": -0.02398234947422148,
1246
+ "max": 0.6540905643890202,
1247
+ "count": 182
1248
+ },
1249
+ "18:future_mean": {
1250
+ "min": -2.522122064472071,
1251
+ "max": 2.864874044299952,
1252
+ "count": 182
1253
+ },
1254
+ "18:future_min": {
1255
+ "min": -2.7156598317395617,
1256
+ "max": 2.8018741357098986,
1257
+ "count": 182
1258
+ },
1259
+ "18:future_max": {
1260
+ "min": -2.4448202826306225,
1261
+ "max": 2.98024523036927,
1262
+ "count": 182
1263
+ }
1264
+ },
1265
+ "limitation": "Marginal feature-envelope heuristic; no guarantee of detecting hidden physical shift.",
1266
+ "sha256": "81d31bd21a19a37ea8b2728debb836319a108f3083d4851e2ead3b28b6b544af"
1267
+ }
studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/thermal-evaluation.json ADDED
The diff for this file is too large to render. See raw diff
 
studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/thermal_temperature_only/auditorium_3node-context.json ADDED
The diff for this file is too large to render. See raw diff
 
studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/thermal_temperature_only/auditorium_3node-predictions.npz ADDED
Binary file (89.5 kB). View file
 
studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/thermal_temperature_only/calibration.json ADDED
@@ -0,0 +1,1842 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "version": "tfm-context-uncertainty-v1",
3
+ "nominal_coverage": 0.9,
4
+ "min_count": 30,
5
+ "checkpoint_sha256": [
6
+ "41485d6dd0c72f09a9ca67ced83bc4a0725f9cde731276d1672902304675ebec"
7
+ ],
8
+ "normalization_sha256": "33b890560ea8c2a881a2a4e125f19e2beb7a69029ec406d4b7e6a13b9880f3e4",
9
+ "transform_id": "history-diagnostics-v1:thermal_temperature_only",
10
+ "context_policy": {
11
+ "low_action_std": 0.05,
12
+ "weak_driver_std": 0.05,
13
+ "support_std_margin": 0.5,
14
+ "support_min_count": 10
15
+ },
16
+ "calibration_system_ids": [
17
+ "battery-00068",
18
+ "battery-00074",
19
+ "battery-00083",
20
+ "battery-00107",
21
+ "battery-00188",
22
+ "battery-00281",
23
+ "battery-00287",
24
+ "battery-00299",
25
+ "battery-00317",
26
+ "battery-00353",
27
+ "battery-00365",
28
+ "battery-00377",
29
+ "battery-00413",
30
+ "battery-00464",
31
+ "battery-00479",
32
+ "battery-00527",
33
+ "battery-00566",
34
+ "battery-00653",
35
+ "battery-00674",
36
+ "battery-00677",
37
+ "battery-00761",
38
+ "battery-00767",
39
+ "battery-00833",
40
+ "battery-00839",
41
+ "battery-00881",
42
+ "battery-00950",
43
+ "battery-01010",
44
+ "battery-01016",
45
+ "battery-01100",
46
+ "battery-01190",
47
+ "tank-00010",
48
+ "tank-00052",
49
+ "tank-00112",
50
+ "tank-00139",
51
+ "tank-00292",
52
+ "tank-00313",
53
+ "tank-00331",
54
+ "tank-00400",
55
+ "tank-00412",
56
+ "tank-00433",
57
+ "tank-00448",
58
+ "tank-00475",
59
+ "tank-00484",
60
+ "tank-00544",
61
+ "tank-00553",
62
+ "tank-00586",
63
+ "tank-00616",
64
+ "tank-00679",
65
+ "tank-00682",
66
+ "tank-00775",
67
+ "tank-00796",
68
+ "tank-00883",
69
+ "tank-00934",
70
+ "tank-00940",
71
+ "tank-00967",
72
+ "tank-00994",
73
+ "tank-01009",
74
+ "tank-01051",
75
+ "tank-01063",
76
+ "tank-01198",
77
+ "thermal-00039",
78
+ "thermal-00117",
79
+ "thermal-00126",
80
+ "thermal-00189",
81
+ "thermal-00195",
82
+ "thermal-00234",
83
+ "thermal-00261",
84
+ "thermal-00270",
85
+ "thermal-00342",
86
+ "thermal-00366",
87
+ "thermal-00369",
88
+ "thermal-00375",
89
+ "thermal-00510",
90
+ "thermal-00531",
91
+ "thermal-00561",
92
+ "thermal-00567",
93
+ "thermal-00597",
94
+ "thermal-00660",
95
+ "thermal-00663",
96
+ "thermal-00831",
97
+ "thermal-00834",
98
+ "thermal-00924",
99
+ "thermal-00972",
100
+ "thermal-01023",
101
+ "thermal-01050",
102
+ "thermal-01092",
103
+ "thermal-01137",
104
+ "thermal-01158",
105
+ "thermal-01170",
106
+ "thermal-01182"
107
+ ],
108
+ "support_sha256": "fea344dd6ad5066c3c0328c18bb31d57e34bc6704bc2ea7bed51df6d90f7b992",
109
+ "training_system_ids": [
110
+ "battery-00002",
111
+ "battery-00005",
112
+ "battery-00008",
113
+ "battery-00014",
114
+ "battery-00020",
115
+ "battery-00026",
116
+ "battery-00032",
117
+ "battery-00035",
118
+ "battery-00044",
119
+ "battery-00050",
120
+ "battery-00053",
121
+ "battery-00056",
122
+ "battery-00059",
123
+ "battery-00062",
124
+ "battery-00065",
125
+ "battery-00071",
126
+ "battery-00077",
127
+ "battery-00089",
128
+ "battery-00092",
129
+ "battery-00098",
130
+ "battery-00101",
131
+ "battery-00110",
132
+ "battery-00113",
133
+ "battery-00116",
134
+ "battery-00122",
135
+ "battery-00125",
136
+ "battery-00131",
137
+ "battery-00137",
138
+ "battery-00140",
139
+ "battery-00143",
140
+ "battery-00155",
141
+ "battery-00170",
142
+ "battery-00173",
143
+ "battery-00179",
144
+ "battery-00182",
145
+ "battery-00185",
146
+ "battery-00191",
147
+ "battery-00197",
148
+ "battery-00200",
149
+ "battery-00203",
150
+ "battery-00206",
151
+ "battery-00209",
152
+ "battery-00215",
153
+ "battery-00218",
154
+ "battery-00221",
155
+ "battery-00227",
156
+ "battery-00233",
157
+ "battery-00236",
158
+ "battery-00242",
159
+ "battery-00245",
160
+ "battery-00248",
161
+ "battery-00251",
162
+ "battery-00254",
163
+ "battery-00257",
164
+ "battery-00269",
165
+ "battery-00272",
166
+ "battery-00278",
167
+ "battery-00284",
168
+ "battery-00293",
169
+ "battery-00302",
170
+ "battery-00311",
171
+ "battery-00314",
172
+ "battery-00323",
173
+ "battery-00326",
174
+ "battery-00329",
175
+ "battery-00332",
176
+ "battery-00335",
177
+ "battery-00338",
178
+ "battery-00344",
179
+ "battery-00350",
180
+ "battery-00356",
181
+ "battery-00359",
182
+ "battery-00362",
183
+ "battery-00368",
184
+ "battery-00371",
185
+ "battery-00374",
186
+ "battery-00383",
187
+ "battery-00386",
188
+ "battery-00389",
189
+ "battery-00392",
190
+ "battery-00395",
191
+ "battery-00398",
192
+ "battery-00401",
193
+ "battery-00404",
194
+ "battery-00407",
195
+ "battery-00410",
196
+ "battery-00419",
197
+ "battery-00422",
198
+ "battery-00425",
199
+ "battery-00428",
200
+ "battery-00431",
201
+ "battery-00434",
202
+ "battery-00437",
203
+ "battery-00440",
204
+ "battery-00443",
205
+ "battery-00449",
206
+ "battery-00452",
207
+ "battery-00455",
208
+ "battery-00458",
209
+ "battery-00461",
210
+ "battery-00467",
211
+ "battery-00470",
212
+ "battery-00473",
213
+ "battery-00476",
214
+ "battery-00485",
215
+ "battery-00488",
216
+ "battery-00491",
217
+ "battery-00494",
218
+ "battery-00497",
219
+ "battery-00503",
220
+ "battery-00506",
221
+ "battery-00509",
222
+ "battery-00512",
223
+ "battery-00515",
224
+ "battery-00518",
225
+ "battery-00521",
226
+ "battery-00524",
227
+ "battery-00530",
228
+ "battery-00536",
229
+ "battery-00539",
230
+ "battery-00542",
231
+ "battery-00545",
232
+ "battery-00548",
233
+ "battery-00551",
234
+ "battery-00554",
235
+ "battery-00557",
236
+ "battery-00560",
237
+ "battery-00563",
238
+ "battery-00569",
239
+ "battery-00572",
240
+ "battery-00575",
241
+ "battery-00584",
242
+ "battery-00590",
243
+ "battery-00596",
244
+ "battery-00602",
245
+ "battery-00611",
246
+ "battery-00614",
247
+ "battery-00617",
248
+ "battery-00620",
249
+ "battery-00623",
250
+ "battery-00626",
251
+ "battery-00629",
252
+ "battery-00638",
253
+ "battery-00641",
254
+ "battery-00644",
255
+ "battery-00647",
256
+ "battery-00650",
257
+ "battery-00656",
258
+ "battery-00662",
259
+ "battery-00668",
260
+ "battery-00671",
261
+ "battery-00680",
262
+ "battery-00683",
263
+ "battery-00686",
264
+ "battery-00689",
265
+ "battery-00692",
266
+ "battery-00695",
267
+ "battery-00698",
268
+ "battery-00701",
269
+ "battery-00707",
270
+ "battery-00710",
271
+ "battery-00713",
272
+ "battery-00716",
273
+ "battery-00725",
274
+ "battery-00728",
275
+ "battery-00731",
276
+ "battery-00734",
277
+ "battery-00740",
278
+ "battery-00743",
279
+ "battery-00746",
280
+ "battery-00749",
281
+ "battery-00755",
282
+ "battery-00758",
283
+ "battery-00776",
284
+ "battery-00779",
285
+ "battery-00785",
286
+ "battery-00791",
287
+ "battery-00794",
288
+ "battery-00800",
289
+ "battery-00809",
290
+ "battery-00812",
291
+ "battery-00815",
292
+ "battery-00818",
293
+ "battery-00821",
294
+ "battery-00824",
295
+ "battery-00827",
296
+ "battery-00836",
297
+ "battery-00842",
298
+ "battery-00845",
299
+ "battery-00848",
300
+ "battery-00851",
301
+ "battery-00854",
302
+ "battery-00863",
303
+ "battery-00866",
304
+ "battery-00869",
305
+ "battery-00872",
306
+ "battery-00875",
307
+ "battery-00884",
308
+ "battery-00887",
309
+ "battery-00896",
310
+ "battery-00902",
311
+ "battery-00905",
312
+ "battery-00911",
313
+ "battery-00914",
314
+ "battery-00917",
315
+ "battery-00920",
316
+ "battery-00923",
317
+ "battery-00926",
318
+ "battery-00929",
319
+ "battery-00932",
320
+ "battery-00935",
321
+ "battery-00938",
322
+ "battery-00941",
323
+ "battery-00944",
324
+ "battery-00953",
325
+ "battery-00956",
326
+ "battery-00959",
327
+ "battery-00965",
328
+ "battery-00968",
329
+ "battery-00971",
330
+ "battery-00974",
331
+ "battery-00977",
332
+ "battery-00980",
333
+ "battery-00986",
334
+ "battery-00995",
335
+ "battery-00998",
336
+ "battery-01001",
337
+ "battery-01004",
338
+ "battery-01007",
339
+ "battery-01013",
340
+ "battery-01019",
341
+ "battery-01022",
342
+ "battery-01028",
343
+ "battery-01031",
344
+ "battery-01034",
345
+ "battery-01037",
346
+ "battery-01040",
347
+ "battery-01043",
348
+ "battery-01046",
349
+ "battery-01049",
350
+ "battery-01055",
351
+ "battery-01058",
352
+ "battery-01061",
353
+ "battery-01064",
354
+ "battery-01067",
355
+ "battery-01070",
356
+ "battery-01073",
357
+ "battery-01076",
358
+ "battery-01079",
359
+ "battery-01085",
360
+ "battery-01088",
361
+ "battery-01091",
362
+ "battery-01094",
363
+ "battery-01097",
364
+ "battery-01103",
365
+ "battery-01106",
366
+ "battery-01109",
367
+ "battery-01115",
368
+ "battery-01118",
369
+ "battery-01121",
370
+ "battery-01124",
371
+ "battery-01130",
372
+ "battery-01133",
373
+ "battery-01136",
374
+ "battery-01139",
375
+ "battery-01142",
376
+ "battery-01145",
377
+ "battery-01151",
378
+ "battery-01157",
379
+ "battery-01160",
380
+ "battery-01166",
381
+ "battery-01169",
382
+ "battery-01172",
383
+ "battery-01175",
384
+ "battery-01178",
385
+ "battery-01184",
386
+ "battery-01187",
387
+ "battery-01193",
388
+ "battery-01196",
389
+ "battery-01199",
390
+ "tank-00001",
391
+ "tank-00004",
392
+ "tank-00007",
393
+ "tank-00013",
394
+ "tank-00016",
395
+ "tank-00019",
396
+ "tank-00022",
397
+ "tank-00025",
398
+ "tank-00028",
399
+ "tank-00031",
400
+ "tank-00034",
401
+ "tank-00037",
402
+ "tank-00040",
403
+ "tank-00043",
404
+ "tank-00049",
405
+ "tank-00055",
406
+ "tank-00058",
407
+ "tank-00061",
408
+ "tank-00064",
409
+ "tank-00067",
410
+ "tank-00070",
411
+ "tank-00073",
412
+ "tank-00076",
413
+ "tank-00082",
414
+ "tank-00085",
415
+ "tank-00088",
416
+ "tank-00091",
417
+ "tank-00094",
418
+ "tank-00103",
419
+ "tank-00106",
420
+ "tank-00109",
421
+ "tank-00115",
422
+ "tank-00118",
423
+ "tank-00121",
424
+ "tank-00124",
425
+ "tank-00133",
426
+ "tank-00136",
427
+ "tank-00145",
428
+ "tank-00154",
429
+ "tank-00157",
430
+ "tank-00163",
431
+ "tank-00166",
432
+ "tank-00169",
433
+ "tank-00172",
434
+ "tank-00178",
435
+ "tank-00181",
436
+ "tank-00184",
437
+ "tank-00193",
438
+ "tank-00196",
439
+ "tank-00199",
440
+ "tank-00202",
441
+ "tank-00208",
442
+ "tank-00214",
443
+ "tank-00220",
444
+ "tank-00223",
445
+ "tank-00226",
446
+ "tank-00229",
447
+ "tank-00232",
448
+ "tank-00235",
449
+ "tank-00238",
450
+ "tank-00244",
451
+ "tank-00247",
452
+ "tank-00250",
453
+ "tank-00259",
454
+ "tank-00262",
455
+ "tank-00265",
456
+ "tank-00268",
457
+ "tank-00274",
458
+ "tank-00277",
459
+ "tank-00280",
460
+ "tank-00283",
461
+ "tank-00289",
462
+ "tank-00295",
463
+ "tank-00301",
464
+ "tank-00304",
465
+ "tank-00307",
466
+ "tank-00310",
467
+ "tank-00316",
468
+ "tank-00319",
469
+ "tank-00322",
470
+ "tank-00328",
471
+ "tank-00334",
472
+ "tank-00340",
473
+ "tank-00343",
474
+ "tank-00346",
475
+ "tank-00349",
476
+ "tank-00352",
477
+ "tank-00355",
478
+ "tank-00358",
479
+ "tank-00361",
480
+ "tank-00364",
481
+ "tank-00367",
482
+ "tank-00370",
483
+ "tank-00373",
484
+ "tank-00379",
485
+ "tank-00382",
486
+ "tank-00385",
487
+ "tank-00388",
488
+ "tank-00394",
489
+ "tank-00397",
490
+ "tank-00403",
491
+ "tank-00406",
492
+ "tank-00409",
493
+ "tank-00415",
494
+ "tank-00418",
495
+ "tank-00421",
496
+ "tank-00424",
497
+ "tank-00427",
498
+ "tank-00430",
499
+ "tank-00439",
500
+ "tank-00442",
501
+ "tank-00445",
502
+ "tank-00451",
503
+ "tank-00454",
504
+ "tank-00457",
505
+ "tank-00463",
506
+ "tank-00469",
507
+ "tank-00472",
508
+ "tank-00478",
509
+ "tank-00490",
510
+ "tank-00493",
511
+ "tank-00496",
512
+ "tank-00499",
513
+ "tank-00508",
514
+ "tank-00511",
515
+ "tank-00514",
516
+ "tank-00520",
517
+ "tank-00529",
518
+ "tank-00538",
519
+ "tank-00541",
520
+ "tank-00547",
521
+ "tank-00550",
522
+ "tank-00556",
523
+ "tank-00559",
524
+ "tank-00562",
525
+ "tank-00580",
526
+ "tank-00589",
527
+ "tank-00595",
528
+ "tank-00598",
529
+ "tank-00604",
530
+ "tank-00607",
531
+ "tank-00610",
532
+ "tank-00613",
533
+ "tank-00619",
534
+ "tank-00622",
535
+ "tank-00646",
536
+ "tank-00649",
537
+ "tank-00652",
538
+ "tank-00655",
539
+ "tank-00658",
540
+ "tank-00664",
541
+ "tank-00667",
542
+ "tank-00670",
543
+ "tank-00673",
544
+ "tank-00676",
545
+ "tank-00685",
546
+ "tank-00688",
547
+ "tank-00691",
548
+ "tank-00694",
549
+ "tank-00697",
550
+ "tank-00700",
551
+ "tank-00703",
552
+ "tank-00709",
553
+ "tank-00712",
554
+ "tank-00715",
555
+ "tank-00718",
556
+ "tank-00721",
557
+ "tank-00724",
558
+ "tank-00727",
559
+ "tank-00730",
560
+ "tank-00733",
561
+ "tank-00736",
562
+ "tank-00739",
563
+ "tank-00742",
564
+ "tank-00745",
565
+ "tank-00748",
566
+ "tank-00751",
567
+ "tank-00760",
568
+ "tank-00763",
569
+ "tank-00766",
570
+ "tank-00778",
571
+ "tank-00781",
572
+ "tank-00787",
573
+ "tank-00790",
574
+ "tank-00793",
575
+ "tank-00802",
576
+ "tank-00805",
577
+ "tank-00811",
578
+ "tank-00814",
579
+ "tank-00817",
580
+ "tank-00820",
581
+ "tank-00823",
582
+ "tank-00829",
583
+ "tank-00832",
584
+ "tank-00835",
585
+ "tank-00838",
586
+ "tank-00844",
587
+ "tank-00847",
588
+ "tank-00850",
589
+ "tank-00853",
590
+ "tank-00856",
591
+ "tank-00859",
592
+ "tank-00865",
593
+ "tank-00868",
594
+ "tank-00871",
595
+ "tank-00877",
596
+ "tank-00889",
597
+ "tank-00892",
598
+ "tank-00895",
599
+ "tank-00898",
600
+ "tank-00901",
601
+ "tank-00904",
602
+ "tank-00907",
603
+ "tank-00910",
604
+ "tank-00916",
605
+ "tank-00919",
606
+ "tank-00922",
607
+ "tank-00925",
608
+ "tank-00931",
609
+ "tank-00937",
610
+ "tank-00946",
611
+ "tank-00949",
612
+ "tank-00955",
613
+ "tank-00958",
614
+ "tank-00961",
615
+ "tank-00964",
616
+ "tank-00970",
617
+ "tank-00973",
618
+ "tank-00979",
619
+ "tank-00982",
620
+ "tank-00985",
621
+ "tank-00988",
622
+ "tank-00991",
623
+ "tank-00997",
624
+ "tank-01000",
625
+ "tank-01003",
626
+ "tank-01006",
627
+ "tank-01012",
628
+ "tank-01018",
629
+ "tank-01021",
630
+ "tank-01024",
631
+ "tank-01027",
632
+ "tank-01030",
633
+ "tank-01033",
634
+ "tank-01036",
635
+ "tank-01039",
636
+ "tank-01042",
637
+ "tank-01045",
638
+ "tank-01048",
639
+ "tank-01054",
640
+ "tank-01057",
641
+ "tank-01066",
642
+ "tank-01069",
643
+ "tank-01075",
644
+ "tank-01084",
645
+ "tank-01087",
646
+ "tank-01090",
647
+ "tank-01096",
648
+ "tank-01105",
649
+ "tank-01108",
650
+ "tank-01111",
651
+ "tank-01117",
652
+ "tank-01120",
653
+ "tank-01126",
654
+ "tank-01135",
655
+ "tank-01138",
656
+ "tank-01141",
657
+ "tank-01144",
658
+ "tank-01147",
659
+ "tank-01150",
660
+ "tank-01153",
661
+ "tank-01159",
662
+ "tank-01162",
663
+ "tank-01165",
664
+ "tank-01168",
665
+ "tank-01174",
666
+ "tank-01177",
667
+ "tank-01180",
668
+ "tank-01186",
669
+ "tank-01195",
670
+ "thermal-00000",
671
+ "thermal-00006",
672
+ "thermal-00009",
673
+ "thermal-00012",
674
+ "thermal-00021",
675
+ "thermal-00024",
676
+ "thermal-00027",
677
+ "thermal-00033",
678
+ "thermal-00036",
679
+ "thermal-00054",
680
+ "thermal-00057",
681
+ "thermal-00060",
682
+ "thermal-00072",
683
+ "thermal-00075",
684
+ "thermal-00078",
685
+ "thermal-00081",
686
+ "thermal-00084",
687
+ "thermal-00087",
688
+ "thermal-00090",
689
+ "thermal-00096",
690
+ "thermal-00099",
691
+ "thermal-00111",
692
+ "thermal-00114",
693
+ "thermal-00120",
694
+ "thermal-00132",
695
+ "thermal-00135",
696
+ "thermal-00138",
697
+ "thermal-00141",
698
+ "thermal-00144",
699
+ "thermal-00150",
700
+ "thermal-00153",
701
+ "thermal-00156",
702
+ "thermal-00159",
703
+ "thermal-00162",
704
+ "thermal-00165",
705
+ "thermal-00168",
706
+ "thermal-00174",
707
+ "thermal-00177",
708
+ "thermal-00180",
709
+ "thermal-00183",
710
+ "thermal-00186",
711
+ "thermal-00192",
712
+ "thermal-00198",
713
+ "thermal-00201",
714
+ "thermal-00204",
715
+ "thermal-00210",
716
+ "thermal-00216",
717
+ "thermal-00219",
718
+ "thermal-00225",
719
+ "thermal-00231",
720
+ "thermal-00237",
721
+ "thermal-00240",
722
+ "thermal-00243",
723
+ "thermal-00246",
724
+ "thermal-00249",
725
+ "thermal-00252",
726
+ "thermal-00258",
727
+ "thermal-00264",
728
+ "thermal-00267",
729
+ "thermal-00273",
730
+ "thermal-00276",
731
+ "thermal-00279",
732
+ "thermal-00282",
733
+ "thermal-00285",
734
+ "thermal-00288",
735
+ "thermal-00294",
736
+ "thermal-00300",
737
+ "thermal-00303",
738
+ "thermal-00312",
739
+ "thermal-00318",
740
+ "thermal-00321",
741
+ "thermal-00324",
742
+ "thermal-00336",
743
+ "thermal-00345",
744
+ "thermal-00354",
745
+ "thermal-00357",
746
+ "thermal-00360",
747
+ "thermal-00363",
748
+ "thermal-00372",
749
+ "thermal-00378",
750
+ "thermal-00381",
751
+ "thermal-00384",
752
+ "thermal-00390",
753
+ "thermal-00393",
754
+ "thermal-00396",
755
+ "thermal-00399",
756
+ "thermal-00402",
757
+ "thermal-00405",
758
+ "thermal-00414",
759
+ "thermal-00420",
760
+ "thermal-00429",
761
+ "thermal-00432",
762
+ "thermal-00435",
763
+ "thermal-00438",
764
+ "thermal-00441",
765
+ "thermal-00444",
766
+ "thermal-00447",
767
+ "thermal-00450",
768
+ "thermal-00453",
769
+ "thermal-00456",
770
+ "thermal-00459",
771
+ "thermal-00462",
772
+ "thermal-00465",
773
+ "thermal-00468",
774
+ "thermal-00474",
775
+ "thermal-00477",
776
+ "thermal-00483",
777
+ "thermal-00486",
778
+ "thermal-00492",
779
+ "thermal-00495",
780
+ "thermal-00498",
781
+ "thermal-00504",
782
+ "thermal-00507",
783
+ "thermal-00513",
784
+ "thermal-00516",
785
+ "thermal-00519",
786
+ "thermal-00522",
787
+ "thermal-00528",
788
+ "thermal-00534",
789
+ "thermal-00537",
790
+ "thermal-00540",
791
+ "thermal-00543",
792
+ "thermal-00549",
793
+ "thermal-00552",
794
+ "thermal-00555",
795
+ "thermal-00558",
796
+ "thermal-00573",
797
+ "thermal-00576",
798
+ "thermal-00579",
799
+ "thermal-00588",
800
+ "thermal-00591",
801
+ "thermal-00594",
802
+ "thermal-00600",
803
+ "thermal-00603",
804
+ "thermal-00606",
805
+ "thermal-00609",
806
+ "thermal-00612",
807
+ "thermal-00615",
808
+ "thermal-00621",
809
+ "thermal-00627",
810
+ "thermal-00630",
811
+ "thermal-00633",
812
+ "thermal-00639",
813
+ "thermal-00642",
814
+ "thermal-00645",
815
+ "thermal-00648",
816
+ "thermal-00651",
817
+ "thermal-00657",
818
+ "thermal-00666",
819
+ "thermal-00675",
820
+ "thermal-00678",
821
+ "thermal-00681",
822
+ "thermal-00684",
823
+ "thermal-00687",
824
+ "thermal-00690",
825
+ "thermal-00693",
826
+ "thermal-00696",
827
+ "thermal-00699",
828
+ "thermal-00702",
829
+ "thermal-00705",
830
+ "thermal-00708",
831
+ "thermal-00714",
832
+ "thermal-00717",
833
+ "thermal-00723",
834
+ "thermal-00726",
835
+ "thermal-00729",
836
+ "thermal-00732",
837
+ "thermal-00735",
838
+ "thermal-00738",
839
+ "thermal-00741",
840
+ "thermal-00744",
841
+ "thermal-00747",
842
+ "thermal-00750",
843
+ "thermal-00753",
844
+ "thermal-00756",
845
+ "thermal-00759",
846
+ "thermal-00762",
847
+ "thermal-00765",
848
+ "thermal-00768",
849
+ "thermal-00771",
850
+ "thermal-00774",
851
+ "thermal-00777",
852
+ "thermal-00780",
853
+ "thermal-00783",
854
+ "thermal-00786",
855
+ "thermal-00789",
856
+ "thermal-00792",
857
+ "thermal-00795",
858
+ "thermal-00804",
859
+ "thermal-00807",
860
+ "thermal-00810",
861
+ "thermal-00813",
862
+ "thermal-00819",
863
+ "thermal-00828",
864
+ "thermal-00840",
865
+ "thermal-00843",
866
+ "thermal-00846",
867
+ "thermal-00849",
868
+ "thermal-00852",
869
+ "thermal-00855",
870
+ "thermal-00864",
871
+ "thermal-00867",
872
+ "thermal-00873",
873
+ "thermal-00876",
874
+ "thermal-00879",
875
+ "thermal-00882",
876
+ "thermal-00885",
877
+ "thermal-00894",
878
+ "thermal-00897",
879
+ "thermal-00906",
880
+ "thermal-00912",
881
+ "thermal-00915",
882
+ "thermal-00918",
883
+ "thermal-00921",
884
+ "thermal-00930",
885
+ "thermal-00933",
886
+ "thermal-00936",
887
+ "thermal-00954",
888
+ "thermal-00957",
889
+ "thermal-00963",
890
+ "thermal-00966",
891
+ "thermal-00969",
892
+ "thermal-00978",
893
+ "thermal-00981",
894
+ "thermal-00984",
895
+ "thermal-00987",
896
+ "thermal-00990",
897
+ "thermal-00993",
898
+ "thermal-00996",
899
+ "thermal-01002",
900
+ "thermal-01005",
901
+ "thermal-01008",
902
+ "thermal-01011",
903
+ "thermal-01014",
904
+ "thermal-01017",
905
+ "thermal-01020",
906
+ "thermal-01026",
907
+ "thermal-01029",
908
+ "thermal-01032",
909
+ "thermal-01035",
910
+ "thermal-01038",
911
+ "thermal-01041",
912
+ "thermal-01044",
913
+ "thermal-01047",
914
+ "thermal-01056",
915
+ "thermal-01059",
916
+ "thermal-01062",
917
+ "thermal-01065",
918
+ "thermal-01068",
919
+ "thermal-01071",
920
+ "thermal-01074",
921
+ "thermal-01080",
922
+ "thermal-01083",
923
+ "thermal-01086",
924
+ "thermal-01089",
925
+ "thermal-01095",
926
+ "thermal-01104",
927
+ "thermal-01107",
928
+ "thermal-01110",
929
+ "thermal-01113",
930
+ "thermal-01116",
931
+ "thermal-01122",
932
+ "thermal-01125",
933
+ "thermal-01128",
934
+ "thermal-01134",
935
+ "thermal-01140",
936
+ "thermal-01143",
937
+ "thermal-01146",
938
+ "thermal-01149",
939
+ "thermal-01152",
940
+ "thermal-01164",
941
+ "thermal-01167",
942
+ "thermal-01173",
943
+ "thermal-01176",
944
+ "thermal-01179",
945
+ "thermal-01185",
946
+ "thermal-01188",
947
+ "thermal-01191",
948
+ "thermal-01194",
949
+ "thermal-01197"
950
+ ],
951
+ "validation_selection_system_ids": [
952
+ "battery-00023",
953
+ "battery-00038",
954
+ "battery-00041",
955
+ "battery-00047",
956
+ "battery-00080",
957
+ "battery-00104",
958
+ "battery-00119",
959
+ "battery-00146",
960
+ "battery-00158",
961
+ "battery-00167",
962
+ "battery-00176",
963
+ "battery-00275",
964
+ "battery-00296",
965
+ "battery-00305",
966
+ "battery-00308",
967
+ "battery-00482",
968
+ "battery-00587",
969
+ "battery-00593",
970
+ "battery-00605",
971
+ "battery-00632",
972
+ "battery-00665",
973
+ "battery-00752",
974
+ "battery-00764",
975
+ "battery-00773",
976
+ "battery-00788",
977
+ "battery-00797",
978
+ "battery-00893",
979
+ "battery-00962",
980
+ "battery-01025",
981
+ "battery-01127",
982
+ "tank-00046",
983
+ "tank-00142",
984
+ "tank-00160",
985
+ "tank-00211",
986
+ "tank-00241",
987
+ "tank-00256",
988
+ "tank-00286",
989
+ "tank-00337",
990
+ "tank-00505",
991
+ "tank-00517",
992
+ "tank-00523",
993
+ "tank-00574",
994
+ "tank-00577",
995
+ "tank-00601",
996
+ "tank-00625",
997
+ "tank-00634",
998
+ "tank-00769",
999
+ "tank-00808",
1000
+ "tank-00874",
1001
+ "tank-00913",
1002
+ "tank-00928",
1003
+ "tank-00952",
1004
+ "tank-01078",
1005
+ "tank-01081",
1006
+ "tank-01093",
1007
+ "tank-01099",
1008
+ "tank-01102",
1009
+ "tank-01114",
1010
+ "tank-01129",
1011
+ "tank-01171",
1012
+ "thermal-00003",
1013
+ "thermal-00015",
1014
+ "thermal-00030",
1015
+ "thermal-00051",
1016
+ "thermal-00063",
1017
+ "thermal-00207",
1018
+ "thermal-00213",
1019
+ "thermal-00222",
1020
+ "thermal-00255",
1021
+ "thermal-00297",
1022
+ "thermal-00309",
1023
+ "thermal-00333",
1024
+ "thermal-00348",
1025
+ "thermal-00351",
1026
+ "thermal-00426",
1027
+ "thermal-00471",
1028
+ "thermal-00501",
1029
+ "thermal-00564",
1030
+ "thermal-00618",
1031
+ "thermal-00654",
1032
+ "thermal-00669",
1033
+ "thermal-00672",
1034
+ "thermal-00798",
1035
+ "thermal-00822",
1036
+ "thermal-00903",
1037
+ "thermal-00945",
1038
+ "thermal-01077",
1039
+ "thermal-01101",
1040
+ "thermal-01131",
1041
+ "thermal-01161"
1042
+ ],
1043
+ "test_system_ids": [
1044
+ "battery-00011",
1045
+ "battery-00017",
1046
+ "battery-00029",
1047
+ "battery-00086",
1048
+ "battery-00095",
1049
+ "battery-00128",
1050
+ "battery-00134",
1051
+ "battery-00149",
1052
+ "battery-00152",
1053
+ "battery-00161",
1054
+ "battery-00164",
1055
+ "battery-00194",
1056
+ "battery-00212",
1057
+ "battery-00224",
1058
+ "battery-00230",
1059
+ "battery-00239",
1060
+ "battery-00260",
1061
+ "battery-00263",
1062
+ "battery-00266",
1063
+ "battery-00290",
1064
+ "battery-00320",
1065
+ "battery-00341",
1066
+ "battery-00347",
1067
+ "battery-00380",
1068
+ "battery-00416",
1069
+ "battery-00446",
1070
+ "battery-00500",
1071
+ "battery-00533",
1072
+ "battery-00578",
1073
+ "battery-00581",
1074
+ "battery-00599",
1075
+ "battery-00608",
1076
+ "battery-00635",
1077
+ "battery-00659",
1078
+ "battery-00704",
1079
+ "battery-00719",
1080
+ "battery-00722",
1081
+ "battery-00737",
1082
+ "battery-00770",
1083
+ "battery-00782",
1084
+ "battery-00803",
1085
+ "battery-00806",
1086
+ "battery-00830",
1087
+ "battery-00857",
1088
+ "battery-00860",
1089
+ "battery-00878",
1090
+ "battery-00890",
1091
+ "battery-00899",
1092
+ "battery-00908",
1093
+ "battery-00947",
1094
+ "battery-00983",
1095
+ "battery-00989",
1096
+ "battery-00992",
1097
+ "battery-01052",
1098
+ "battery-01082",
1099
+ "battery-01112",
1100
+ "battery-01148",
1101
+ "battery-01154",
1102
+ "battery-01163",
1103
+ "battery-01181",
1104
+ "challenge-auditorium_3node-1041230",
1105
+ "challenge-auditorium_3node-1041231",
1106
+ "challenge-auditorium_3node-1041232",
1107
+ "challenge-auditorium_3node-1041233",
1108
+ "challenge-auditorium_3node-1041234",
1109
+ "challenge-auditorium_3node-1041235",
1110
+ "challenge-auditorium_3node-1041236",
1111
+ "challenge-auditorium_3node-1041237",
1112
+ "challenge-auditorium_3node-1041238",
1113
+ "challenge-auditorium_3node-1041239",
1114
+ "challenge-auditorium_3node-1041240",
1115
+ "challenge-auditorium_3node-1041241",
1116
+ "challenge-auditorium_3node-1041242",
1117
+ "challenge-auditorium_3node-1041243",
1118
+ "challenge-auditorium_3node-1041244",
1119
+ "challenge-auditorium_3node-1041245",
1120
+ "challenge-auditorium_3node-1041246",
1121
+ "challenge-auditorium_3node-1041247",
1122
+ "challenge-auditorium_3node-1041248",
1123
+ "challenge-auditorium_3node-1041249",
1124
+ "challenge-auditorium_3node-1041250",
1125
+ "challenge-auditorium_3node-1041251",
1126
+ "challenge-auditorium_3node-1041252",
1127
+ "challenge-auditorium_3node-1041253",
1128
+ "challenge-parameter_shift_2node-841230",
1129
+ "challenge-parameter_shift_2node-841231",
1130
+ "challenge-parameter_shift_2node-841232",
1131
+ "challenge-parameter_shift_2node-841233",
1132
+ "challenge-parameter_shift_2node-841234",
1133
+ "challenge-parameter_shift_2node-841235",
1134
+ "challenge-parameter_shift_2node-841236",
1135
+ "challenge-parameter_shift_2node-841237",
1136
+ "challenge-parameter_shift_2node-841238",
1137
+ "challenge-parameter_shift_2node-841239",
1138
+ "challenge-parameter_shift_2node-841240",
1139
+ "challenge-parameter_shift_2node-841241",
1140
+ "challenge-parameter_shift_2node-841242",
1141
+ "challenge-parameter_shift_2node-841243",
1142
+ "challenge-parameter_shift_2node-841244",
1143
+ "challenge-parameter_shift_2node-841245",
1144
+ "challenge-parameter_shift_2node-841246",
1145
+ "challenge-parameter_shift_2node-841247",
1146
+ "challenge-parameter_shift_2node-841248",
1147
+ "challenge-parameter_shift_2node-841249",
1148
+ "challenge-parameter_shift_2node-841250",
1149
+ "challenge-parameter_shift_2node-841251",
1150
+ "challenge-parameter_shift_2node-841252",
1151
+ "challenge-parameter_shift_2node-841253",
1152
+ "challenge-small_room_3node-941230",
1153
+ "challenge-small_room_3node-941231",
1154
+ "challenge-small_room_3node-941232",
1155
+ "challenge-small_room_3node-941233",
1156
+ "challenge-small_room_3node-941234",
1157
+ "challenge-small_room_3node-941235",
1158
+ "challenge-small_room_3node-941236",
1159
+ "challenge-small_room_3node-941237",
1160
+ "challenge-small_room_3node-941238",
1161
+ "challenge-small_room_3node-941239",
1162
+ "challenge-small_room_3node-941240",
1163
+ "challenge-small_room_3node-941241",
1164
+ "challenge-small_room_3node-941242",
1165
+ "challenge-small_room_3node-941243",
1166
+ "challenge-small_room_3node-941244",
1167
+ "challenge-small_room_3node-941245",
1168
+ "challenge-small_room_3node-941246",
1169
+ "challenge-small_room_3node-941247",
1170
+ "challenge-small_room_3node-941248",
1171
+ "challenge-small_room_3node-941249",
1172
+ "challenge-small_room_3node-941250",
1173
+ "challenge-small_room_3node-941251",
1174
+ "challenge-small_room_3node-941252",
1175
+ "challenge-small_room_3node-941253",
1176
+ "tank-00079",
1177
+ "tank-00097",
1178
+ "tank-00100",
1179
+ "tank-00127",
1180
+ "tank-00130",
1181
+ "tank-00148",
1182
+ "tank-00151",
1183
+ "tank-00175",
1184
+ "tank-00187",
1185
+ "tank-00190",
1186
+ "tank-00205",
1187
+ "tank-00217",
1188
+ "tank-00253",
1189
+ "tank-00271",
1190
+ "tank-00298",
1191
+ "tank-00325",
1192
+ "tank-00376",
1193
+ "tank-00391",
1194
+ "tank-00436",
1195
+ "tank-00460",
1196
+ "tank-00466",
1197
+ "tank-00481",
1198
+ "tank-00487",
1199
+ "tank-00502",
1200
+ "tank-00526",
1201
+ "tank-00532",
1202
+ "tank-00535",
1203
+ "tank-00565",
1204
+ "tank-00568",
1205
+ "tank-00571",
1206
+ "tank-00583",
1207
+ "tank-00592",
1208
+ "tank-00628",
1209
+ "tank-00631",
1210
+ "tank-00637",
1211
+ "tank-00640",
1212
+ "tank-00643",
1213
+ "tank-00661",
1214
+ "tank-00706",
1215
+ "tank-00754",
1216
+ "tank-00757",
1217
+ "tank-00772",
1218
+ "tank-00784",
1219
+ "tank-00799",
1220
+ "tank-00826",
1221
+ "tank-00841",
1222
+ "tank-00862",
1223
+ "tank-00880",
1224
+ "tank-00886",
1225
+ "tank-00943",
1226
+ "tank-00976",
1227
+ "tank-01015",
1228
+ "tank-01060",
1229
+ "tank-01072",
1230
+ "tank-01123",
1231
+ "tank-01132",
1232
+ "tank-01156",
1233
+ "tank-01183",
1234
+ "tank-01189",
1235
+ "tank-01192",
1236
+ "thermal-00018",
1237
+ "thermal-00042",
1238
+ "thermal-00045",
1239
+ "thermal-00048",
1240
+ "thermal-00066",
1241
+ "thermal-00069",
1242
+ "thermal-00093",
1243
+ "thermal-00102",
1244
+ "thermal-00105",
1245
+ "thermal-00108",
1246
+ "thermal-00123",
1247
+ "thermal-00129",
1248
+ "thermal-00147",
1249
+ "thermal-00171",
1250
+ "thermal-00228",
1251
+ "thermal-00291",
1252
+ "thermal-00306",
1253
+ "thermal-00315",
1254
+ "thermal-00327",
1255
+ "thermal-00330",
1256
+ "thermal-00339",
1257
+ "thermal-00387",
1258
+ "thermal-00408",
1259
+ "thermal-00411",
1260
+ "thermal-00417",
1261
+ "thermal-00423",
1262
+ "thermal-00480",
1263
+ "thermal-00489",
1264
+ "thermal-00525",
1265
+ "thermal-00546",
1266
+ "thermal-00570",
1267
+ "thermal-00582",
1268
+ "thermal-00585",
1269
+ "thermal-00624",
1270
+ "thermal-00636",
1271
+ "thermal-00711",
1272
+ "thermal-00720",
1273
+ "thermal-00801",
1274
+ "thermal-00816",
1275
+ "thermal-00825",
1276
+ "thermal-00837",
1277
+ "thermal-00858",
1278
+ "thermal-00861",
1279
+ "thermal-00870",
1280
+ "thermal-00888",
1281
+ "thermal-00891",
1282
+ "thermal-00900",
1283
+ "thermal-00909",
1284
+ "thermal-00927",
1285
+ "thermal-00939",
1286
+ "thermal-00942",
1287
+ "thermal-00948",
1288
+ "thermal-00951",
1289
+ "thermal-00960",
1290
+ "thermal-00975",
1291
+ "thermal-00999",
1292
+ "thermal-01053",
1293
+ "thermal-01098",
1294
+ "thermal-01119",
1295
+ "thermal-01155"
1296
+ ],
1297
+ "entries": {
1298
+ "0:15|incomplete_sensors": {
1299
+ "count": 150,
1300
+ "multiplier": 1.3273102045059204
1301
+ },
1302
+ "0:15|low_observed_action_variation+incomplete_sensors": {
1303
+ "count": 1,
1304
+ "multiplier": null
1305
+ },
1306
+ "0:30|incomplete_sensors": {
1307
+ "count": 152,
1308
+ "multiplier": 1.4217875003814697
1309
+ },
1310
+ "0:30|low_observed_action_variation+incomplete_sensors": {
1311
+ "count": 1,
1312
+ "multiplier": null
1313
+ },
1314
+ "0:60|incomplete_sensors": {
1315
+ "count": 150,
1316
+ "multiplier": 1.387830138206482
1317
+ },
1318
+ "0:60|low_observed_action_variation+incomplete_sensors": {
1319
+ "count": 1,
1320
+ "multiplier": null
1321
+ },
1322
+ "0:120|incomplete_sensors": {
1323
+ "count": 153,
1324
+ "multiplier": 1.444566011428833
1325
+ },
1326
+ "0:120|low_observed_action_variation+incomplete_sensors": {
1327
+ "count": 1,
1328
+ "multiplier": null
1329
+ },
1330
+ "3:15|incomplete_sensors": {
1331
+ "count": 139,
1332
+ "multiplier": 1.602124810218811
1333
+ },
1334
+ "3:15|incomplete_sensors+outside_training_support": {
1335
+ "count": 1,
1336
+ "multiplier": null
1337
+ },
1338
+ "3:15|observed_context": {
1339
+ "count": 23,
1340
+ "multiplier": null
1341
+ },
1342
+ "3:30|incomplete_sensors": {
1343
+ "count": 138,
1344
+ "multiplier": 1.6081340312957764
1345
+ },
1346
+ "3:30|incomplete_sensors+outside_training_support": {
1347
+ "count": 1,
1348
+ "multiplier": null
1349
+ },
1350
+ "3:30|observed_context": {
1351
+ "count": 22,
1352
+ "multiplier": null
1353
+ },
1354
+ "3:60|incomplete_sensors": {
1355
+ "count": 140,
1356
+ "multiplier": 1.7659860849380493
1357
+ },
1358
+ "3:60|incomplete_sensors+outside_training_support": {
1359
+ "count": 1,
1360
+ "multiplier": null
1361
+ },
1362
+ "3:60|observed_context": {
1363
+ "count": 23,
1364
+ "multiplier": null
1365
+ },
1366
+ "3:120|incomplete_sensors": {
1367
+ "count": 136,
1368
+ "multiplier": 1.4678348302841187
1369
+ },
1370
+ "3:120|observed_context": {
1371
+ "count": 23,
1372
+ "multiplier": null
1373
+ },
1374
+ "4:15|incomplete_sensors": {
1375
+ "count": 137,
1376
+ "multiplier": 1.9247430562973022
1377
+ },
1378
+ "4:15|incomplete_sensors+outside_training_support": {
1379
+ "count": 1,
1380
+ "multiplier": null
1381
+ },
1382
+ "4:15|observed_context": {
1383
+ "count": 23,
1384
+ "multiplier": null
1385
+ },
1386
+ "4:30|incomplete_sensors": {
1387
+ "count": 136,
1388
+ "multiplier": 1.9186359643936157
1389
+ },
1390
+ "4:30|incomplete_sensors+outside_training_support": {
1391
+ "count": 1,
1392
+ "multiplier": null
1393
+ },
1394
+ "4:30|observed_context": {
1395
+ "count": 23,
1396
+ "multiplier": null
1397
+ },
1398
+ "4:60|incomplete_sensors": {
1399
+ "count": 137,
1400
+ "multiplier": 1.886003851890564
1401
+ },
1402
+ "4:60|incomplete_sensors+outside_training_support": {
1403
+ "count": 1,
1404
+ "multiplier": null
1405
+ },
1406
+ "4:60|observed_context": {
1407
+ "count": 23,
1408
+ "multiplier": null
1409
+ },
1410
+ "4:120|incomplete_sensors": {
1411
+ "count": 137,
1412
+ "multiplier": 1.8352510929107666
1413
+ },
1414
+ "4:120|incomplete_sensors+outside_training_support": {
1415
+ "count": 1,
1416
+ "multiplier": null
1417
+ },
1418
+ "4:120|observed_context": {
1419
+ "count": 22,
1420
+ "multiplier": null
1421
+ },
1422
+ "5:15|incomplete_sensors": {
1423
+ "count": 142,
1424
+ "multiplier": 1.2845878601074219
1425
+ },
1426
+ "5:15|incomplete_sensors+outside_training_support": {
1427
+ "count": 1,
1428
+ "multiplier": null
1429
+ },
1430
+ "5:15|observed_context": {
1431
+ "count": 22,
1432
+ "multiplier": null
1433
+ },
1434
+ "5:30|incomplete_sensors": {
1435
+ "count": 142,
1436
+ "multiplier": 1.1014809608459473
1437
+ },
1438
+ "5:30|incomplete_sensors+outside_training_support": {
1439
+ "count": 1,
1440
+ "multiplier": null
1441
+ },
1442
+ "5:30|observed_context": {
1443
+ "count": 21,
1444
+ "multiplier": null
1445
+ },
1446
+ "5:60|incomplete_sensors": {
1447
+ "count": 142,
1448
+ "multiplier": 1.1266205310821533
1449
+ },
1450
+ "5:60|incomplete_sensors+outside_training_support": {
1451
+ "count": 1,
1452
+ "multiplier": null
1453
+ },
1454
+ "5:60|observed_context": {
1455
+ "count": 23,
1456
+ "multiplier": null
1457
+ },
1458
+ "5:120|incomplete_sensors": {
1459
+ "count": 141,
1460
+ "multiplier": 1.3680959939956665
1461
+ },
1462
+ "5:120|incomplete_sensors+outside_training_support": {
1463
+ "count": 1,
1464
+ "multiplier": null
1465
+ },
1466
+ "5:120|observed_context": {
1467
+ "count": 22,
1468
+ "multiplier": null
1469
+ },
1470
+ "6:15|incomplete_sensors": {
1471
+ "count": 140,
1472
+ "multiplier": 1.5263735055923462
1473
+ },
1474
+ "6:15|incomplete_sensors+outside_training_support": {
1475
+ "count": 1,
1476
+ "multiplier": null
1477
+ },
1478
+ "6:15|observed_context": {
1479
+ "count": 23,
1480
+ "multiplier": null
1481
+ },
1482
+ "6:30|incomplete_sensors": {
1483
+ "count": 142,
1484
+ "multiplier": 1.5905961990356445
1485
+ },
1486
+ "6:30|incomplete_sensors+outside_training_support": {
1487
+ "count": 1,
1488
+ "multiplier": null
1489
+ },
1490
+ "6:30|observed_context": {
1491
+ "count": 22,
1492
+ "multiplier": null
1493
+ },
1494
+ "6:60|incomplete_sensors": {
1495
+ "count": 138,
1496
+ "multiplier": 1.6447550058364868
1497
+ },
1498
+ "6:60|incomplete_sensors+outside_training_support": {
1499
+ "count": 1,
1500
+ "multiplier": null
1501
+ },
1502
+ "6:60|observed_context": {
1503
+ "count": 23,
1504
+ "multiplier": null
1505
+ },
1506
+ "6:120|incomplete_sensors": {
1507
+ "count": 141,
1508
+ "multiplier": 1.6278109550476074
1509
+ },
1510
+ "6:120|incomplete_sensors+outside_training_support": {
1511
+ "count": 1,
1512
+ "multiplier": null
1513
+ },
1514
+ "6:120|observed_context": {
1515
+ "count": 22,
1516
+ "multiplier": null
1517
+ },
1518
+ "7:15|incomplete_sensors": {
1519
+ "count": 132,
1520
+ "multiplier": 1.4662518501281738
1521
+ },
1522
+ "7:15|incomplete_sensors+outside_training_support": {
1523
+ "count": 2,
1524
+ "multiplier": null
1525
+ },
1526
+ "7:15|low_observed_action_variation+incomplete_sensors": {
1527
+ "count": 1,
1528
+ "multiplier": null
1529
+ },
1530
+ "7:15|observed_context": {
1531
+ "count": 38,
1532
+ "multiplier": 1.448378086090088
1533
+ },
1534
+ "7:30|incomplete_sensors": {
1535
+ "count": 130,
1536
+ "multiplier": 1.5992043018341064
1537
+ },
1538
+ "7:30|incomplete_sensors+outside_training_support": {
1539
+ "count": 2,
1540
+ "multiplier": null
1541
+ },
1542
+ "7:30|low_observed_action_variation+incomplete_sensors": {
1543
+ "count": 1,
1544
+ "multiplier": null
1545
+ },
1546
+ "7:30|observed_context": {
1547
+ "count": 38,
1548
+ "multiplier": 1.6324751377105713
1549
+ },
1550
+ "7:60|incomplete_sensors": {
1551
+ "count": 131,
1552
+ "multiplier": 1.5418221950531006
1553
+ },
1554
+ "7:60|incomplete_sensors+outside_training_support": {
1555
+ "count": 2,
1556
+ "multiplier": null
1557
+ },
1558
+ "7:60|low_observed_action_variation+incomplete_sensors": {
1559
+ "count": 1,
1560
+ "multiplier": null
1561
+ },
1562
+ "7:60|observed_context": {
1563
+ "count": 38,
1564
+ "multiplier": 1.7413384914398193
1565
+ },
1566
+ "7:120|incomplete_sensors": {
1567
+ "count": 125,
1568
+ "multiplier": 1.5066261291503906
1569
+ },
1570
+ "7:120|incomplete_sensors+outside_training_support": {
1571
+ "count": 1,
1572
+ "multiplier": null
1573
+ },
1574
+ "7:120|low_observed_action_variation+incomplete_sensors": {
1575
+ "count": 1,
1576
+ "multiplier": null
1577
+ },
1578
+ "7:120|observed_context": {
1579
+ "count": 38,
1580
+ "multiplier": 2.2300097942352295
1581
+ },
1582
+ "8:15|incomplete_sensors": {
1583
+ "count": 136,
1584
+ "multiplier": 2.743947744369507
1585
+ },
1586
+ "8:15|incomplete_sensors+outside_training_support": {
1587
+ "count": 2,
1588
+ "multiplier": null
1589
+ },
1590
+ "8:15|low_observed_action_variation+incomplete_sensors": {
1591
+ "count": 1,
1592
+ "multiplier": null
1593
+ },
1594
+ "8:15|observed_context": {
1595
+ "count": 37,
1596
+ "multiplier": 2.638808012008667
1597
+ },
1598
+ "8:30|incomplete_sensors": {
1599
+ "count": 136,
1600
+ "multiplier": 2.544153928756714
1601
+ },
1602
+ "8:30|incomplete_sensors+outside_training_support": {
1603
+ "count": 2,
1604
+ "multiplier": null
1605
+ },
1606
+ "8:30|low_observed_action_variation+incomplete_sensors": {
1607
+ "count": 1,
1608
+ "multiplier": null
1609
+ },
1610
+ "8:30|observed_context": {
1611
+ "count": 35,
1612
+ "multiplier": 2.0292863845825195
1613
+ },
1614
+ "8:60|incomplete_sensors": {
1615
+ "count": 138,
1616
+ "multiplier": 2.4357502460479736
1617
+ },
1618
+ "8:60|incomplete_sensors+outside_training_support": {
1619
+ "count": 2,
1620
+ "multiplier": null
1621
+ },
1622
+ "8:60|low_observed_action_variation+incomplete_sensors": {
1623
+ "count": 1,
1624
+ "multiplier": null
1625
+ },
1626
+ "8:60|observed_context": {
1627
+ "count": 37,
1628
+ "multiplier": 1.743776798248291
1629
+ },
1630
+ "8:120|incomplete_sensors": {
1631
+ "count": 135,
1632
+ "multiplier": 2.4752357006073
1633
+ },
1634
+ "8:120|incomplete_sensors+outside_training_support": {
1635
+ "count": 2,
1636
+ "multiplier": null
1637
+ },
1638
+ "8:120|low_observed_action_variation+incomplete_sensors": {
1639
+ "count": 1,
1640
+ "multiplier": null
1641
+ },
1642
+ "8:120|observed_context": {
1643
+ "count": 36,
1644
+ "multiplier": 1.53176748752594
1645
+ },
1646
+ "9:15|incomplete_sensors": {
1647
+ "count": 130,
1648
+ "multiplier": 1.7122114896774292
1649
+ },
1650
+ "9:15|incomplete_sensors+outside_training_support": {
1651
+ "count": 2,
1652
+ "multiplier": null
1653
+ },
1654
+ "9:15|low_observed_action_variation+incomplete_sensors": {
1655
+ "count": 1,
1656
+ "multiplier": null
1657
+ },
1658
+ "9:15|observed_context": {
1659
+ "count": 37,
1660
+ "multiplier": 1.9634885787963867
1661
+ },
1662
+ "9:30|incomplete_sensors": {
1663
+ "count": 130,
1664
+ "multiplier": 1.6662906408309937
1665
+ },
1666
+ "9:30|incomplete_sensors+outside_training_support": {
1667
+ "count": 2,
1668
+ "multiplier": null
1669
+ },
1670
+ "9:30|low_observed_action_variation+incomplete_sensors": {
1671
+ "count": 1,
1672
+ "multiplier": null
1673
+ },
1674
+ "9:30|observed_context": {
1675
+ "count": 38,
1676
+ "multiplier": 1.8528791666030884
1677
+ },
1678
+ "9:60|incomplete_sensors": {
1679
+ "count": 130,
1680
+ "multiplier": 1.7820677757263184
1681
+ },
1682
+ "9:60|incomplete_sensors+outside_training_support": {
1683
+ "count": 2,
1684
+ "multiplier": null
1685
+ },
1686
+ "9:60|low_observed_action_variation+incomplete_sensors": {
1687
+ "count": 1,
1688
+ "multiplier": null
1689
+ },
1690
+ "9:60|observed_context": {
1691
+ "count": 36,
1692
+ "multiplier": 1.8004403114318848
1693
+ },
1694
+ "9:120|incomplete_sensors": {
1695
+ "count": 132,
1696
+ "multiplier": 1.935365915298462
1697
+ },
1698
+ "9:120|incomplete_sensors+outside_training_support": {
1699
+ "count": 2,
1700
+ "multiplier": null
1701
+ },
1702
+ "9:120|low_observed_action_variation+incomplete_sensors": {
1703
+ "count": 1,
1704
+ "multiplier": null
1705
+ },
1706
+ "9:120|observed_context": {
1707
+ "count": 37,
1708
+ "multiplier": 2.4109134674072266
1709
+ }
1710
+ },
1711
+ "pooled_entries": {
1712
+ "0:15": {
1713
+ "count": 151,
1714
+ "multiplier": 1.3273102045059204
1715
+ },
1716
+ "0:30": {
1717
+ "count": 153,
1718
+ "multiplier": 1.4217875003814697
1719
+ },
1720
+ "0:60": {
1721
+ "count": 151,
1722
+ "multiplier": 1.387830138206482
1723
+ },
1724
+ "0:120": {
1725
+ "count": 154,
1726
+ "multiplier": 1.4806492328643799
1727
+ },
1728
+ "3:15": {
1729
+ "count": 163,
1730
+ "multiplier": 1.6815403699874878
1731
+ },
1732
+ "3:30": {
1733
+ "count": 161,
1734
+ "multiplier": 1.6081340312957764
1735
+ },
1736
+ "3:60": {
1737
+ "count": 164,
1738
+ "multiplier": 1.7659860849380493
1739
+ },
1740
+ "3:120": {
1741
+ "count": 159,
1742
+ "multiplier": 1.4390878677368164
1743
+ },
1744
+ "4:15": {
1745
+ "count": 161,
1746
+ "multiplier": 1.9247430562973022
1747
+ },
1748
+ "4:30": {
1749
+ "count": 160,
1750
+ "multiplier": 1.8637503385543823
1751
+ },
1752
+ "4:60": {
1753
+ "count": 161,
1754
+ "multiplier": 1.8226490020751953
1755
+ },
1756
+ "4:120": {
1757
+ "count": 160,
1758
+ "multiplier": 1.8352510929107666
1759
+ },
1760
+ "5:15": {
1761
+ "count": 165,
1762
+ "multiplier": 1.2845878601074219
1763
+ },
1764
+ "5:30": {
1765
+ "count": 164,
1766
+ "multiplier": 1.1519505977630615
1767
+ },
1768
+ "5:60": {
1769
+ "count": 166,
1770
+ "multiplier": 1.1387457847595215
1771
+ },
1772
+ "5:120": {
1773
+ "count": 164,
1774
+ "multiplier": 1.3353688716888428
1775
+ },
1776
+ "6:15": {
1777
+ "count": 164,
1778
+ "multiplier": 1.3798800706863403
1779
+ },
1780
+ "6:30": {
1781
+ "count": 165,
1782
+ "multiplier": 1.5757362842559814
1783
+ },
1784
+ "6:60": {
1785
+ "count": 162,
1786
+ "multiplier": 1.58315908908844
1787
+ },
1788
+ "6:120": {
1789
+ "count": 164,
1790
+ "multiplier": 1.5772513151168823
1791
+ },
1792
+ "7:15": {
1793
+ "count": 173,
1794
+ "multiplier": 1.4304349422454834
1795
+ },
1796
+ "7:30": {
1797
+ "count": 171,
1798
+ "multiplier": 1.5780900716781616
1799
+ },
1800
+ "7:60": {
1801
+ "count": 172,
1802
+ "multiplier": 1.5485385656356812
1803
+ },
1804
+ "7:120": {
1805
+ "count": 165,
1806
+ "multiplier": 1.5068869590759277
1807
+ },
1808
+ "8:15": {
1809
+ "count": 176,
1810
+ "multiplier": 2.663092851638794
1811
+ },
1812
+ "8:30": {
1813
+ "count": 174,
1814
+ "multiplier": 2.3664445877075195
1815
+ },
1816
+ "8:60": {
1817
+ "count": 178,
1818
+ "multiplier": 2.1575307846069336
1819
+ },
1820
+ "8:120": {
1821
+ "count": 174,
1822
+ "multiplier": 2.195991277694702
1823
+ },
1824
+ "9:15": {
1825
+ "count": 170,
1826
+ "multiplier": 1.6777907609939575
1827
+ },
1828
+ "9:30": {
1829
+ "count": 171,
1830
+ "multiplier": 1.6662906408309937
1831
+ },
1832
+ "9:60": {
1833
+ "count": 169,
1834
+ "multiplier": 1.7003358602523804
1835
+ },
1836
+ "9:120": {
1837
+ "count": 172,
1838
+ "multiplier": 1.8664488792419434
1839
+ }
1840
+ },
1841
+ "limitation": "Sampled windows are correlated. Finite-sample coverage requires exchangeability; no system-level or shift guarantee. Pooled fallback has no stratum-conditional coverage claim."
1842
+ }
studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/thermal_temperature_only/parameter_shift_2node-context.json ADDED
The diff for this file is too large to render. See raw diff
 
studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/thermal_temperature_only/parameter_shift_2node-predictions.npz ADDED
Binary file (87.9 kB). View file
 
studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/thermal_temperature_only/small_room_3node-context.json ADDED
The diff for this file is too large to render. See raw diff
 
studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/thermal_temperature_only/small_room_3node-predictions.npz ADDED
Binary file (89.3 kB). View file
 
studies/robustness-seed42-20261011/artifacts/robustness/thermal-evaluation/thermal_temperature_only/support.json ADDED
@@ -0,0 +1,1237 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "version": "tfm-context-uncertainty-v1",
3
+ "training_system_ids": [
4
+ "battery-00002",
5
+ "battery-00005",
6
+ "battery-00008",
7
+ "battery-00014",
8
+ "battery-00020",
9
+ "battery-00026",
10
+ "battery-00032",
11
+ "battery-00035",
12
+ "battery-00044",
13
+ "battery-00050",
14
+ "battery-00053",
15
+ "battery-00056",
16
+ "battery-00059",
17
+ "battery-00062",
18
+ "battery-00065",
19
+ "battery-00071",
20
+ "battery-00077",
21
+ "battery-00089",
22
+ "battery-00092",
23
+ "battery-00098",
24
+ "battery-00101",
25
+ "battery-00110",
26
+ "battery-00113",
27
+ "battery-00116",
28
+ "battery-00122",
29
+ "battery-00125",
30
+ "battery-00131",
31
+ "battery-00137",
32
+ "battery-00140",
33
+ "battery-00143",
34
+ "battery-00155",
35
+ "battery-00170",
36
+ "battery-00173",
37
+ "battery-00179",
38
+ "battery-00182",
39
+ "battery-00185",
40
+ "battery-00191",
41
+ "battery-00197",
42
+ "battery-00200",
43
+ "battery-00203",
44
+ "battery-00206",
45
+ "battery-00209",
46
+ "battery-00215",
47
+ "battery-00218",
48
+ "battery-00221",
49
+ "battery-00227",
50
+ "battery-00233",
51
+ "battery-00236",
52
+ "battery-00242",
53
+ "battery-00245",
54
+ "battery-00248",
55
+ "battery-00251",
56
+ "battery-00254",
57
+ "battery-00257",
58
+ "battery-00269",
59
+ "battery-00272",
60
+ "battery-00278",
61
+ "battery-00284",
62
+ "battery-00293",
63
+ "battery-00302",
64
+ "battery-00311",
65
+ "battery-00314",
66
+ "battery-00323",
67
+ "battery-00326",
68
+ "battery-00329",
69
+ "battery-00332",
70
+ "battery-00335",
71
+ "battery-00338",
72
+ "battery-00344",
73
+ "battery-00350",
74
+ "battery-00356",
75
+ "battery-00359",
76
+ "battery-00362",
77
+ "battery-00368",
78
+ "battery-00371",
79
+ "battery-00374",
80
+ "battery-00383",
81
+ "battery-00386",
82
+ "battery-00389",
83
+ "battery-00392",
84
+ "battery-00395",
85
+ "battery-00398",
86
+ "battery-00401",
87
+ "battery-00404",
88
+ "battery-00407",
89
+ "battery-00410",
90
+ "battery-00419",
91
+ "battery-00422",
92
+ "battery-00425",
93
+ "battery-00428",
94
+ "battery-00431",
95
+ "battery-00434",
96
+ "battery-00437",
97
+ "battery-00440",
98
+ "battery-00443",
99
+ "battery-00449",
100
+ "battery-00452",
101
+ "battery-00455",
102
+ "battery-00458",
103
+ "battery-00461",
104
+ "battery-00467",
105
+ "battery-00470",
106
+ "battery-00473",
107
+ "battery-00476",
108
+ "battery-00485",
109
+ "battery-00488",
110
+ "battery-00491",
111
+ "battery-00494",
112
+ "battery-00497",
113
+ "battery-00503",
114
+ "battery-00506",
115
+ "battery-00509",
116
+ "battery-00512",
117
+ "battery-00515",
118
+ "battery-00518",
119
+ "battery-00521",
120
+ "battery-00524",
121
+ "battery-00530",
122
+ "battery-00536",
123
+ "battery-00539",
124
+ "battery-00542",
125
+ "battery-00545",
126
+ "battery-00548",
127
+ "battery-00551",
128
+ "battery-00554",
129
+ "battery-00557",
130
+ "battery-00560",
131
+ "battery-00563",
132
+ "battery-00569",
133
+ "battery-00572",
134
+ "battery-00575",
135
+ "battery-00584",
136
+ "battery-00590",
137
+ "battery-00596",
138
+ "battery-00602",
139
+ "battery-00611",
140
+ "battery-00614",
141
+ "battery-00617",
142
+ "battery-00620",
143
+ "battery-00623",
144
+ "battery-00626",
145
+ "battery-00629",
146
+ "battery-00638",
147
+ "battery-00641",
148
+ "battery-00644",
149
+ "battery-00647",
150
+ "battery-00650",
151
+ "battery-00656",
152
+ "battery-00662",
153
+ "battery-00668",
154
+ "battery-00671",
155
+ "battery-00680",
156
+ "battery-00683",
157
+ "battery-00686",
158
+ "battery-00689",
159
+ "battery-00692",
160
+ "battery-00695",
161
+ "battery-00698",
162
+ "battery-00701",
163
+ "battery-00707",
164
+ "battery-00710",
165
+ "battery-00713",
166
+ "battery-00716",
167
+ "battery-00725",
168
+ "battery-00728",
169
+ "battery-00731",
170
+ "battery-00734",
171
+ "battery-00740",
172
+ "battery-00743",
173
+ "battery-00746",
174
+ "battery-00749",
175
+ "battery-00755",
176
+ "battery-00758",
177
+ "battery-00776",
178
+ "battery-00779",
179
+ "battery-00785",
180
+ "battery-00791",
181
+ "battery-00794",
182
+ "battery-00800",
183
+ "battery-00809",
184
+ "battery-00812",
185
+ "battery-00815",
186
+ "battery-00818",
187
+ "battery-00821",
188
+ "battery-00824",
189
+ "battery-00827",
190
+ "battery-00836",
191
+ "battery-00842",
192
+ "battery-00845",
193
+ "battery-00848",
194
+ "battery-00851",
195
+ "battery-00854",
196
+ "battery-00863",
197
+ "battery-00866",
198
+ "battery-00869",
199
+ "battery-00872",
200
+ "battery-00875",
201
+ "battery-00884",
202
+ "battery-00887",
203
+ "battery-00896",
204
+ "battery-00902",
205
+ "battery-00905",
206
+ "battery-00911",
207
+ "battery-00914",
208
+ "battery-00917",
209
+ "battery-00920",
210
+ "battery-00923",
211
+ "battery-00926",
212
+ "battery-00929",
213
+ "battery-00932",
214
+ "battery-00935",
215
+ "battery-00938",
216
+ "battery-00941",
217
+ "battery-00944",
218
+ "battery-00953",
219
+ "battery-00956",
220
+ "battery-00959",
221
+ "battery-00965",
222
+ "battery-00968",
223
+ "battery-00971",
224
+ "battery-00974",
225
+ "battery-00977",
226
+ "battery-00980",
227
+ "battery-00986",
228
+ "battery-00995",
229
+ "battery-00998",
230
+ "battery-01001",
231
+ "battery-01004",
232
+ "battery-01007",
233
+ "battery-01013",
234
+ "battery-01019",
235
+ "battery-01022",
236
+ "battery-01028",
237
+ "battery-01031",
238
+ "battery-01034",
239
+ "battery-01037",
240
+ "battery-01040",
241
+ "battery-01043",
242
+ "battery-01046",
243
+ "battery-01049",
244
+ "battery-01055",
245
+ "battery-01058",
246
+ "battery-01061",
247
+ "battery-01064",
248
+ "battery-01067",
249
+ "battery-01070",
250
+ "battery-01073",
251
+ "battery-01076",
252
+ "battery-01079",
253
+ "battery-01085",
254
+ "battery-01088",
255
+ "battery-01091",
256
+ "battery-01094",
257
+ "battery-01097",
258
+ "battery-01103",
259
+ "battery-01106",
260
+ "battery-01109",
261
+ "battery-01115",
262
+ "battery-01118",
263
+ "battery-01121",
264
+ "battery-01124",
265
+ "battery-01130",
266
+ "battery-01133",
267
+ "battery-01136",
268
+ "battery-01139",
269
+ "battery-01142",
270
+ "battery-01145",
271
+ "battery-01151",
272
+ "battery-01157",
273
+ "battery-01160",
274
+ "battery-01166",
275
+ "battery-01169",
276
+ "battery-01172",
277
+ "battery-01175",
278
+ "battery-01178",
279
+ "battery-01184",
280
+ "battery-01187",
281
+ "battery-01193",
282
+ "battery-01196",
283
+ "battery-01199",
284
+ "tank-00001",
285
+ "tank-00004",
286
+ "tank-00007",
287
+ "tank-00013",
288
+ "tank-00016",
289
+ "tank-00019",
290
+ "tank-00022",
291
+ "tank-00025",
292
+ "tank-00028",
293
+ "tank-00031",
294
+ "tank-00034",
295
+ "tank-00037",
296
+ "tank-00040",
297
+ "tank-00043",
298
+ "tank-00049",
299
+ "tank-00055",
300
+ "tank-00058",
301
+ "tank-00061",
302
+ "tank-00064",
303
+ "tank-00067",
304
+ "tank-00070",
305
+ "tank-00073",
306
+ "tank-00076",
307
+ "tank-00082",
308
+ "tank-00085",
309
+ "tank-00088",
310
+ "tank-00091",
311
+ "tank-00094",
312
+ "tank-00103",
313
+ "tank-00106",
314
+ "tank-00109",
315
+ "tank-00115",
316
+ "tank-00118",
317
+ "tank-00121",
318
+ "tank-00124",
319
+ "tank-00133",
320
+ "tank-00136",
321
+ "tank-00145",
322
+ "tank-00154",
323
+ "tank-00157",
324
+ "tank-00163",
325
+ "tank-00166",
326
+ "tank-00169",
327
+ "tank-00172",
328
+ "tank-00178",
329
+ "tank-00181",
330
+ "tank-00184",
331
+ "tank-00193",
332
+ "tank-00196",
333
+ "tank-00199",
334
+ "tank-00202",
335
+ "tank-00208",
336
+ "tank-00214",
337
+ "tank-00220",
338
+ "tank-00223",
339
+ "tank-00226",
340
+ "tank-00229",
341
+ "tank-00232",
342
+ "tank-00235",
343
+ "tank-00238",
344
+ "tank-00244",
345
+ "tank-00247",
346
+ "tank-00250",
347
+ "tank-00259",
348
+ "tank-00262",
349
+ "tank-00265",
350
+ "tank-00268",
351
+ "tank-00274",
352
+ "tank-00277",
353
+ "tank-00280",
354
+ "tank-00283",
355
+ "tank-00289",
356
+ "tank-00295",
357
+ "tank-00301",
358
+ "tank-00304",
359
+ "tank-00307",
360
+ "tank-00310",
361
+ "tank-00316",
362
+ "tank-00319",
363
+ "tank-00322",
364
+ "tank-00328",
365
+ "tank-00334",
366
+ "tank-00340",
367
+ "tank-00343",
368
+ "tank-00346",
369
+ "tank-00349",
370
+ "tank-00352",
371
+ "tank-00355",
372
+ "tank-00358",
373
+ "tank-00361",
374
+ "tank-00364",
375
+ "tank-00367",
376
+ "tank-00370",
377
+ "tank-00373",
378
+ "tank-00379",
379
+ "tank-00382",
380
+ "tank-00385",
381
+ "tank-00388",
382
+ "tank-00394",
383
+ "tank-00397",
384
+ "tank-00403",
385
+ "tank-00406",
386
+ "tank-00409",
387
+ "tank-00415",
388
+ "tank-00418",
389
+ "tank-00421",
390
+ "tank-00424",
391
+ "tank-00427",
392
+ "tank-00430",
393
+ "tank-00439",
394
+ "tank-00442",
395
+ "tank-00445",
396
+ "tank-00451",
397
+ "tank-00454",
398
+ "tank-00457",
399
+ "tank-00463",
400
+ "tank-00469",
401
+ "tank-00472",
402
+ "tank-00478",
403
+ "tank-00490",
404
+ "tank-00493",
405
+ "tank-00496",
406
+ "tank-00499",
407
+ "tank-00508",
408
+ "tank-00511",
409
+ "tank-00514",
410
+ "tank-00520",
411
+ "tank-00529",
412
+ "tank-00538",
413
+ "tank-00541",
414
+ "tank-00547",
415
+ "tank-00550",
416
+ "tank-00556",
417
+ "tank-00559",
418
+ "tank-00562",
419
+ "tank-00580",
420
+ "tank-00589",
421
+ "tank-00595",
422
+ "tank-00598",
423
+ "tank-00604",
424
+ "tank-00607",
425
+ "tank-00610",
426
+ "tank-00613",
427
+ "tank-00619",
428
+ "tank-00622",
429
+ "tank-00646",
430
+ "tank-00649",
431
+ "tank-00652",
432
+ "tank-00655",
433
+ "tank-00658",
434
+ "tank-00664",
435
+ "tank-00667",
436
+ "tank-00670",
437
+ "tank-00673",
438
+ "tank-00676",
439
+ "tank-00685",
440
+ "tank-00688",
441
+ "tank-00691",
442
+ "tank-00694",
443
+ "tank-00697",
444
+ "tank-00700",
445
+ "tank-00703",
446
+ "tank-00709",
447
+ "tank-00712",
448
+ "tank-00715",
449
+ "tank-00718",
450
+ "tank-00721",
451
+ "tank-00724",
452
+ "tank-00727",
453
+ "tank-00730",
454
+ "tank-00733",
455
+ "tank-00736",
456
+ "tank-00739",
457
+ "tank-00742",
458
+ "tank-00745",
459
+ "tank-00748",
460
+ "tank-00751",
461
+ "tank-00760",
462
+ "tank-00763",
463
+ "tank-00766",
464
+ "tank-00778",
465
+ "tank-00781",
466
+ "tank-00787",
467
+ "tank-00790",
468
+ "tank-00793",
469
+ "tank-00802",
470
+ "tank-00805",
471
+ "tank-00811",
472
+ "tank-00814",
473
+ "tank-00817",
474
+ "tank-00820",
475
+ "tank-00823",
476
+ "tank-00829",
477
+ "tank-00832",
478
+ "tank-00835",
479
+ "tank-00838",
480
+ "tank-00844",
481
+ "tank-00847",
482
+ "tank-00850",
483
+ "tank-00853",
484
+ "tank-00856",
485
+ "tank-00859",
486
+ "tank-00865",
487
+ "tank-00868",
488
+ "tank-00871",
489
+ "tank-00877",
490
+ "tank-00889",
491
+ "tank-00892",
492
+ "tank-00895",
493
+ "tank-00898",
494
+ "tank-00901",
495
+ "tank-00904",
496
+ "tank-00907",
497
+ "tank-00910",
498
+ "tank-00916",
499
+ "tank-00919",
500
+ "tank-00922",
501
+ "tank-00925",
502
+ "tank-00931",
503
+ "tank-00937",
504
+ "tank-00946",
505
+ "tank-00949",
506
+ "tank-00955",
507
+ "tank-00958",
508
+ "tank-00961",
509
+ "tank-00964",
510
+ "tank-00970",
511
+ "tank-00973",
512
+ "tank-00979",
513
+ "tank-00982",
514
+ "tank-00985",
515
+ "tank-00988",
516
+ "tank-00991",
517
+ "tank-00997",
518
+ "tank-01000",
519
+ "tank-01003",
520
+ "tank-01006",
521
+ "tank-01012",
522
+ "tank-01018",
523
+ "tank-01021",
524
+ "tank-01024",
525
+ "tank-01027",
526
+ "tank-01030",
527
+ "tank-01033",
528
+ "tank-01036",
529
+ "tank-01039",
530
+ "tank-01042",
531
+ "tank-01045",
532
+ "tank-01048",
533
+ "tank-01054",
534
+ "tank-01057",
535
+ "tank-01066",
536
+ "tank-01069",
537
+ "tank-01075",
538
+ "tank-01084",
539
+ "tank-01087",
540
+ "tank-01090",
541
+ "tank-01096",
542
+ "tank-01105",
543
+ "tank-01108",
544
+ "tank-01111",
545
+ "tank-01117",
546
+ "tank-01120",
547
+ "tank-01126",
548
+ "tank-01135",
549
+ "tank-01138",
550
+ "tank-01141",
551
+ "tank-01144",
552
+ "tank-01147",
553
+ "tank-01150",
554
+ "tank-01153",
555
+ "tank-01159",
556
+ "tank-01162",
557
+ "tank-01165",
558
+ "tank-01168",
559
+ "tank-01174",
560
+ "tank-01177",
561
+ "tank-01180",
562
+ "tank-01186",
563
+ "tank-01195",
564
+ "thermal-00000",
565
+ "thermal-00006",
566
+ "thermal-00009",
567
+ "thermal-00012",
568
+ "thermal-00021",
569
+ "thermal-00024",
570
+ "thermal-00027",
571
+ "thermal-00033",
572
+ "thermal-00036",
573
+ "thermal-00054",
574
+ "thermal-00057",
575
+ "thermal-00060",
576
+ "thermal-00072",
577
+ "thermal-00075",
578
+ "thermal-00078",
579
+ "thermal-00081",
580
+ "thermal-00084",
581
+ "thermal-00087",
582
+ "thermal-00090",
583
+ "thermal-00096",
584
+ "thermal-00099",
585
+ "thermal-00111",
586
+ "thermal-00114",
587
+ "thermal-00120",
588
+ "thermal-00132",
589
+ "thermal-00135",
590
+ "thermal-00138",
591
+ "thermal-00141",
592
+ "thermal-00144",
593
+ "thermal-00150",
594
+ "thermal-00153",
595
+ "thermal-00156",
596
+ "thermal-00159",
597
+ "thermal-00162",
598
+ "thermal-00165",
599
+ "thermal-00168",
600
+ "thermal-00174",
601
+ "thermal-00177",
602
+ "thermal-00180",
603
+ "thermal-00183",
604
+ "thermal-00186",
605
+ "thermal-00192",
606
+ "thermal-00198",
607
+ "thermal-00201",
608
+ "thermal-00204",
609
+ "thermal-00210",
610
+ "thermal-00216",
611
+ "thermal-00219",
612
+ "thermal-00225",
613
+ "thermal-00231",
614
+ "thermal-00237",
615
+ "thermal-00240",
616
+ "thermal-00243",
617
+ "thermal-00246",
618
+ "thermal-00249",
619
+ "thermal-00252",
620
+ "thermal-00258",
621
+ "thermal-00264",
622
+ "thermal-00267",
623
+ "thermal-00273",
624
+ "thermal-00276",
625
+ "thermal-00279",
626
+ "thermal-00282",
627
+ "thermal-00285",
628
+ "thermal-00288",
629
+ "thermal-00294",
630
+ "thermal-00300",
631
+ "thermal-00303",
632
+ "thermal-00312",
633
+ "thermal-00318",
634
+ "thermal-00321",
635
+ "thermal-00324",
636
+ "thermal-00336",
637
+ "thermal-00345",
638
+ "thermal-00354",
639
+ "thermal-00357",
640
+ "thermal-00360",
641
+ "thermal-00363",
642
+ "thermal-00372",
643
+ "thermal-00378",
644
+ "thermal-00381",
645
+ "thermal-00384",
646
+ "thermal-00390",
647
+ "thermal-00393",
648
+ "thermal-00396",
649
+ "thermal-00399",
650
+ "thermal-00402",
651
+ "thermal-00405",
652
+ "thermal-00414",
653
+ "thermal-00420",
654
+ "thermal-00429",
655
+ "thermal-00432",
656
+ "thermal-00435",
657
+ "thermal-00438",
658
+ "thermal-00441",
659
+ "thermal-00444",
660
+ "thermal-00447",
661
+ "thermal-00450",
662
+ "thermal-00453",
663
+ "thermal-00456",
664
+ "thermal-00459",
665
+ "thermal-00462",
666
+ "thermal-00465",
667
+ "thermal-00468",
668
+ "thermal-00474",
669
+ "thermal-00477",
670
+ "thermal-00483",
671
+ "thermal-00486",
672
+ "thermal-00492",
673
+ "thermal-00495",
674
+ "thermal-00498",
675
+ "thermal-00504",
676
+ "thermal-00507",
677
+ "thermal-00513",
678
+ "thermal-00516",
679
+ "thermal-00519",
680
+ "thermal-00522",
681
+ "thermal-00528",
682
+ "thermal-00534",
683
+ "thermal-00537",
684
+ "thermal-00540",
685
+ "thermal-00543",
686
+ "thermal-00549",
687
+ "thermal-00552",
688
+ "thermal-00555",
689
+ "thermal-00558",
690
+ "thermal-00573",
691
+ "thermal-00576",
692
+ "thermal-00579",
693
+ "thermal-00588",
694
+ "thermal-00591",
695
+ "thermal-00594",
696
+ "thermal-00600",
697
+ "thermal-00603",
698
+ "thermal-00606",
699
+ "thermal-00609",
700
+ "thermal-00612",
701
+ "thermal-00615",
702
+ "thermal-00621",
703
+ "thermal-00627",
704
+ "thermal-00630",
705
+ "thermal-00633",
706
+ "thermal-00639",
707
+ "thermal-00642",
708
+ "thermal-00645",
709
+ "thermal-00648",
710
+ "thermal-00651",
711
+ "thermal-00657",
712
+ "thermal-00666",
713
+ "thermal-00675",
714
+ "thermal-00678",
715
+ "thermal-00681",
716
+ "thermal-00684",
717
+ "thermal-00687",
718
+ "thermal-00690",
719
+ "thermal-00693",
720
+ "thermal-00696",
721
+ "thermal-00699",
722
+ "thermal-00702",
723
+ "thermal-00705",
724
+ "thermal-00708",
725
+ "thermal-00714",
726
+ "thermal-00717",
727
+ "thermal-00723",
728
+ "thermal-00726",
729
+ "thermal-00729",
730
+ "thermal-00732",
731
+ "thermal-00735",
732
+ "thermal-00738",
733
+ "thermal-00741",
734
+ "thermal-00744",
735
+ "thermal-00747",
736
+ "thermal-00750",
737
+ "thermal-00753",
738
+ "thermal-00756",
739
+ "thermal-00759",
740
+ "thermal-00762",
741
+ "thermal-00765",
742
+ "thermal-00768",
743
+ "thermal-00771",
744
+ "thermal-00774",
745
+ "thermal-00777",
746
+ "thermal-00780",
747
+ "thermal-00783",
748
+ "thermal-00786",
749
+ "thermal-00789",
750
+ "thermal-00792",
751
+ "thermal-00795",
752
+ "thermal-00804",
753
+ "thermal-00807",
754
+ "thermal-00810",
755
+ "thermal-00813",
756
+ "thermal-00819",
757
+ "thermal-00828",
758
+ "thermal-00840",
759
+ "thermal-00843",
760
+ "thermal-00846",
761
+ "thermal-00849",
762
+ "thermal-00852",
763
+ "thermal-00855",
764
+ "thermal-00864",
765
+ "thermal-00867",
766
+ "thermal-00873",
767
+ "thermal-00876",
768
+ "thermal-00879",
769
+ "thermal-00882",
770
+ "thermal-00885",
771
+ "thermal-00894",
772
+ "thermal-00897",
773
+ "thermal-00906",
774
+ "thermal-00912",
775
+ "thermal-00915",
776
+ "thermal-00918",
777
+ "thermal-00921",
778
+ "thermal-00930",
779
+ "thermal-00933",
780
+ "thermal-00936",
781
+ "thermal-00954",
782
+ "thermal-00957",
783
+ "thermal-00963",
784
+ "thermal-00966",
785
+ "thermal-00969",
786
+ "thermal-00978",
787
+ "thermal-00981",
788
+ "thermal-00984",
789
+ "thermal-00987",
790
+ "thermal-00990",
791
+ "thermal-00993",
792
+ "thermal-00996",
793
+ "thermal-01002",
794
+ "thermal-01005",
795
+ "thermal-01008",
796
+ "thermal-01011",
797
+ "thermal-01014",
798
+ "thermal-01017",
799
+ "thermal-01020",
800
+ "thermal-01026",
801
+ "thermal-01029",
802
+ "thermal-01032",
803
+ "thermal-01035",
804
+ "thermal-01038",
805
+ "thermal-01041",
806
+ "thermal-01044",
807
+ "thermal-01047",
808
+ "thermal-01056",
809
+ "thermal-01059",
810
+ "thermal-01062",
811
+ "thermal-01065",
812
+ "thermal-01068",
813
+ "thermal-01071",
814
+ "thermal-01074",
815
+ "thermal-01080",
816
+ "thermal-01083",
817
+ "thermal-01086",
818
+ "thermal-01089",
819
+ "thermal-01095",
820
+ "thermal-01104",
821
+ "thermal-01107",
822
+ "thermal-01110",
823
+ "thermal-01113",
824
+ "thermal-01116",
825
+ "thermal-01122",
826
+ "thermal-01125",
827
+ "thermal-01128",
828
+ "thermal-01134",
829
+ "thermal-01140",
830
+ "thermal-01143",
831
+ "thermal-01146",
832
+ "thermal-01149",
833
+ "thermal-01152",
834
+ "thermal-01164",
835
+ "thermal-01167",
836
+ "thermal-01173",
837
+ "thermal-01176",
838
+ "thermal-01179",
839
+ "thermal-01185",
840
+ "thermal-01188",
841
+ "thermal-01191",
842
+ "thermal-01194",
843
+ "thermal-01197"
844
+ ],
845
+ "normalization_sha256": "33b890560ea8c2a881a2a4e125f19e2beb7a69029ec406d4b7e6a13b9880f3e4",
846
+ "transform_id": "history-diagnostics-v1:thermal_temperature_only",
847
+ "policy": {
848
+ "low_action_std": 0.05,
849
+ "weak_driver_std": 0.05,
850
+ "support_std_margin": 0.5,
851
+ "support_min_count": 10
852
+ },
853
+ "entries": {
854
+ "7:history_mean": {
855
+ "min": -1.7290941794445034,
856
+ "max": 1.8246395905544277,
857
+ "count": 166
858
+ },
859
+ "7:history_std": {
860
+ "min": -0.14753900378313156,
861
+ "max": 1.3250237545308123,
862
+ "count": 166
863
+ },
864
+ "7:current": {
865
+ "min": -1.8345006415060774,
866
+ "max": 1.9300461718252913,
867
+ "count": 161
868
+ },
869
+ "8:history_mean": {
870
+ "min": -2.7116945162741843,
871
+ "max": 5.300033368202895,
872
+ "count": 166
873
+ },
874
+ "8:history_std": {
875
+ "min": -0.05179354612176362,
876
+ "max": 1.5166016823465103,
877
+ "count": 166
878
+ },
879
+ "8:current": {
880
+ "min": -2.6074347951387455,
881
+ "max": 7.818217561480909,
882
+ "count": 161
883
+ },
884
+ "9:history_mean": {
885
+ "min": -2.056825511125415,
886
+ "max": 2.2097568437136106,
887
+ "count": 166
888
+ },
889
+ "9:history_std": {
890
+ "min": -0.17823313635116123,
891
+ "max": 1.9088055984330703,
892
+ "count": 166
893
+ },
894
+ "13:history_mean": {
895
+ "min": -2.0936771567892185,
896
+ "max": 1.7503260310720552,
897
+ "count": 166
898
+ },
899
+ "13:history_std": {
900
+ "min": -0.1447562429817589,
901
+ "max": 1.683897765702215,
902
+ "count": 166
903
+ },
904
+ "13:future_mean": {
905
+ "min": -2.238292363341394,
906
+ "max": 2.211721089537683,
907
+ "count": 166
908
+ },
909
+ "13:future_min": {
910
+ "min": -2.2697974004622576,
911
+ "max": 2.2432261266585467,
912
+ "count": 166
913
+ },
914
+ "13:future_max": {
915
+ "min": -2.297491891162325,
916
+ "max": 2.2709206173586143,
917
+ "count": 166
918
+ },
919
+ "19:history_mean": {
920
+ "min": -2.512724211988596,
921
+ "max": 2.6421154053784885,
922
+ "count": 166
923
+ },
924
+ "19:history_std": {
925
+ "min": -0.01629896148844199,
926
+ "max": 0.6177024400214716,
927
+ "count": 166
928
+ },
929
+ "19:future_mean": {
930
+ "min": -2.550761560451955,
931
+ "max": 2.681303361904592,
932
+ "count": 166
933
+ },
934
+ "19:future_min": {
935
+ "min": -2.5944152936715605,
936
+ "max": 2.55719195459075,
937
+ "count": 166
938
+ },
939
+ "19:future_max": {
940
+ "min": -2.4807157780203952,
941
+ "max": 2.7571237350974216,
942
+ "count": 166
943
+ },
944
+ "history_dt:span": {
945
+ "min": 344.999999,
946
+ "max": 345.000001,
947
+ "count": 512
948
+ },
949
+ "future_dt:span": {
950
+ "min": 104.999999,
951
+ "max": 105.000001,
952
+ "count": 512
953
+ },
954
+ "0:history_mean": {
955
+ "min": -6.85931327257949,
956
+ "max": 3.8700592747767555,
957
+ "count": 164
958
+ },
959
+ "0:history_std": {
960
+ "min": -0.14812077248330266,
961
+ "max": 3.648339465650274,
962
+ "count": 164
963
+ },
964
+ "0:current": {
965
+ "min": -6.9520964500721725,
966
+ "max": 5.824430453520659,
967
+ "count": 162
968
+ },
969
+ "10:history_mean": {
970
+ "min": -1.740817883887489,
971
+ "max": 1.853309133925636,
972
+ "count": 164
973
+ },
974
+ "10:history_std": {
975
+ "min": -0.0890423621527541,
976
+ "max": 1.6749548133007872,
977
+ "count": 164
978
+ },
979
+ "10:future_mean": {
980
+ "min": -2.1179169128433077,
981
+ "max": 2.063192219544968,
982
+ "count": 164
983
+ },
984
+ "10:future_min": {
985
+ "min": -2.1401726013446933,
986
+ "max": 2.0854479080463535,
987
+ "count": 164
988
+ },
989
+ "10:future_max": {
990
+ "min": -2.173832215238251,
991
+ "max": 2.1191075219399114,
992
+ "count": 164
993
+ },
994
+ "14:history_mean": {
995
+ "min": -2.689308242364716,
996
+ "max": 3.0338695807893985,
997
+ "count": 164
998
+ },
999
+ "14:history_std": {
1000
+ "min": -0.018989615475842492,
1001
+ "max": 0.7289486307260484,
1002
+ "count": 164
1003
+ },
1004
+ "14:future_mean": {
1005
+ "min": -2.7651439204986037,
1006
+ "max": 2.859376241284126,
1007
+ "count": 164
1008
+ },
1009
+ "14:future_min": {
1010
+ "min": -2.858736697072939,
1011
+ "max": 2.7898151310634174,
1012
+ "count": 164
1013
+ },
1014
+ "14:future_max": {
1015
+ "min": -2.602246621295409,
1016
+ "max": 2.8956323653902576,
1017
+ "count": 164
1018
+ },
1019
+ "15:history_mean": {
1020
+ "min": -1.1876729007038567,
1021
+ "max": 3.167044984844634,
1022
+ "count": 164
1023
+ },
1024
+ "15:history_std": {
1025
+ "min": -0.1774676965720896,
1026
+ "max": 1.407122425938392,
1027
+ "count": 164
1028
+ },
1029
+ "15:future_mean": {
1030
+ "min": -1.2472839059341898,
1031
+ "max": 3.519802898358201,
1032
+ "count": 164
1033
+ },
1034
+ "15:future_min": {
1035
+ "min": -1.1967750033853077,
1036
+ "max": 3.271746560048821,
1037
+ "count": 164
1038
+ },
1039
+ "15:future_max": {
1040
+ "min": -1.2916961727275031,
1041
+ "max": 3.6574930963648935,
1042
+ "count": 164
1043
+ },
1044
+ "16:history_mean": {
1045
+ "min": -1.4585570580650746,
1046
+ "max": 2.278922486989635,
1047
+ "count": 164
1048
+ },
1049
+ "16:history_std": {
1050
+ "min": -0.20281806972541155,
1051
+ "max": 1.6377656796268398,
1052
+ "count": 164
1053
+ },
1054
+ "16:future_mean": {
1055
+ "min": -1.508917923957451,
1056
+ "max": 2.3405938110653004,
1057
+ "count": 164
1058
+ },
1059
+ "16:future_min": {
1060
+ "min": -1.530090790360272,
1061
+ "max": 2.361766677468121,
1062
+ "count": 164
1063
+ },
1064
+ "16:future_max": {
1065
+ "min": -1.5293852632568663,
1066
+ "max": 2.3854243104980775,
1067
+ "count": 164
1068
+ },
1069
+ "9:current": {
1070
+ "min": -2.2023716167498044,
1071
+ "max": 2.350365538983386,
1072
+ "count": 163
1073
+ },
1074
+ "3:history_mean": {
1075
+ "min": -2.2399926701024677,
1076
+ "max": 2.2418771305516865,
1077
+ "count": 182
1078
+ },
1079
+ "3:history_std": {
1080
+ "min": -0.05676209614502155,
1081
+ "max": 0.5641803710244772,
1082
+ "count": 182
1083
+ },
1084
+ "3:current": {
1085
+ "min": -2.563212187898668,
1086
+ "max": 2.2561561185245833,
1087
+ "count": 176
1088
+ },
1089
+ "4:history_mean": {
1090
+ "min": -3.2996764706891586,
1091
+ "max": 3.761055760316234,
1092
+ "count": 182
1093
+ },
1094
+ "4:history_std": {
1095
+ "min": -0.04884014944697704,
1096
+ "max": 0.9120465918924936,
1097
+ "count": 182
1098
+ },
1099
+ "4:current": {
1100
+ "min": -3.1131484877348123,
1101
+ "max": 2.602466858553904,
1102
+ "count": 182
1103
+ },
1104
+ "5:history_mean": {
1105
+ "min": -1.1944803404030346,
1106
+ "max": 2.729539375227406,
1107
+ "count": 182
1108
+ },
1109
+ "5:history_std": {
1110
+ "min": -0.08287015431103012,
1111
+ "max": 1.9152921351092747,
1112
+ "count": 182
1113
+ },
1114
+ "5:current": {
1115
+ "min": -1.480774138844447,
1116
+ "max": 3.880235885060267,
1117
+ "count": 180
1118
+ },
1119
+ "6:history_mean": {
1120
+ "min": -1.1502409290800177,
1121
+ "max": 2.670196266127595,
1122
+ "count": 182
1123
+ },
1124
+ "6:history_std": {
1125
+ "min": -0.19670381183905647,
1126
+ "max": 2.0513433881214085,
1127
+ "count": 182
1128
+ },
1129
+ "6:current": {
1130
+ "min": -1.5452842631124117,
1131
+ "max": 3.9898985543035126,
1132
+ "count": 181
1133
+ },
1134
+ "11:history_mean": {
1135
+ "min": -1.6990408255271952,
1136
+ "max": 1.4302197052650492,
1137
+ "count": 182
1138
+ },
1139
+ "11:history_std": {
1140
+ "min": -0.16567403674837314,
1141
+ "max": 1.5934025704931851,
1142
+ "count": 182
1143
+ },
1144
+ "11:future_mean": {
1145
+ "min": -2.2535837142599267,
1146
+ "max": 1.642554327739541,
1147
+ "count": 182
1148
+ },
1149
+ "11:future_min": {
1150
+ "min": -2.368155647330075,
1151
+ "max": 1.7571262608096894,
1152
+ "count": 182
1153
+ },
1154
+ "11:future_max": {
1155
+ "min": -2.225419785163575,
1156
+ "max": 1.6143903986431893,
1157
+ "count": 182
1158
+ },
1159
+ "12:history_mean": {
1160
+ "min": -1.4981196618844557,
1161
+ "max": 2.06382686861053,
1162
+ "count": 182
1163
+ },
1164
+ "12:history_std": {
1165
+ "min": -0.15577708269347942,
1166
+ "max": 1.5672527411102275,
1167
+ "count": 182
1168
+ },
1169
+ "12:future_mean": {
1170
+ "min": -1.6069942850034193,
1171
+ "max": 2.228464140311952,
1172
+ "count": 182
1173
+ },
1174
+ "12:future_min": {
1175
+ "min": -1.6163920905112403,
1176
+ "max": 2.237861945819773,
1177
+ "count": 182
1178
+ },
1179
+ "12:future_max": {
1180
+ "min": -1.6778703322386568,
1181
+ "max": 2.2993401875471893,
1182
+ "count": 182
1183
+ },
1184
+ "17:history_mean": {
1185
+ "min": -1.326923140336929,
1186
+ "max": 1.4096587980288846,
1187
+ "count": 182
1188
+ },
1189
+ "17:history_std": {
1190
+ "min": 0.06630291095130086,
1191
+ "max": 1.2772367740596091,
1192
+ "count": 182
1193
+ },
1194
+ "17:future_mean": {
1195
+ "min": -1.6140851258041038,
1196
+ "max": 2.1835690020324363,
1197
+ "count": 182
1198
+ },
1199
+ "17:future_min": {
1200
+ "min": -1.991628161844835,
1201
+ "max": 1.8608503259988878,
1202
+ "count": 182
1203
+ },
1204
+ "17:future_max": {
1205
+ "min": -1.4533015714685136,
1206
+ "max": 2.676748834709328,
1207
+ "count": 182
1208
+ },
1209
+ "18:history_mean": {
1210
+ "min": -2.8351903908626053,
1211
+ "max": 2.4837180130855057,
1212
+ "count": 182
1213
+ },
1214
+ "18:history_std": {
1215
+ "min": -0.02398234947422148,
1216
+ "max": 0.6540905643890202,
1217
+ "count": 182
1218
+ },
1219
+ "18:future_mean": {
1220
+ "min": -2.522122064472071,
1221
+ "max": 2.864874044299952,
1222
+ "count": 182
1223
+ },
1224
+ "18:future_min": {
1225
+ "min": -2.7156598317395617,
1226
+ "max": 2.8018741357098986,
1227
+ "count": 182
1228
+ },
1229
+ "18:future_max": {
1230
+ "min": -2.4448202826306225,
1231
+ "max": 2.98024523036927,
1232
+ "count": 182
1233
+ }
1234
+ },
1235
+ "limitation": "Marginal feature-envelope heuristic; no guarantee of detecting hidden physical shift.",
1236
+ "sha256": "fea344dd6ad5066c3c0328c18bb31d57e34bc6704bc2ea7bed51df6d90f7b992"
1237
+ }
studies/robustness-seed42-20261011/configs/full-study.json ADDED
@@ -0,0 +1,44 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "version": "tfm-full-study-v1",
3
+ "data": "artifacts/poc/data",
4
+ "manifest_sha256": "bfa4012aa501c8645851c8bf03095b8eacaa63da8f7ae4b2a559ac01d3cd41b7",
5
+ "output": "artifacts/full",
6
+ "shared_transformer_seeds": [42, 43, 44],
7
+ "learned_baseline_seeds": [42],
8
+ "specialist_seeds": [42],
9
+ "specialist_models": ["transformer", "mlp"],
10
+ "families": ["thermal", "tank", "battery"],
11
+ "training": {
12
+ "epochs": 50,
13
+ "shared_windows": 150000,
14
+ "specialist_windows": 50000,
15
+ "batch_size": 256,
16
+ "shared_validation_windows": 4096,
17
+ "specialist_validation_windows": 1365,
18
+ "history_steps": 24,
19
+ "split_seed": 42,
20
+ "device": "mps",
21
+ "learning_rate": 0.0003,
22
+ "weight_decay": 0.01,
23
+ "patience": 8,
24
+ "warmup_fraction": 0.05
25
+ },
26
+ "evaluation": {
27
+ "windows": 4096,
28
+ "calibration_windows": 4096,
29
+ "counterfactual_snapshots": 300,
30
+ "seed": 2026,
31
+ "batch_size": 64,
32
+ "catboost_training_windows": 150000,
33
+ "catboost_iterations": 200,
34
+ "catboost_depth": 6,
35
+ "catboost_threads": 4
36
+ },
37
+ "fairness": {
38
+ "specialist_partition_rule": "Filter parent train/test and already established parent selection/calibration IDs; never repartition by family",
39
+ "specialist_exposure_rule": "50000 family windows versus approximately50000 of each family in150000 shared windows per epoch",
40
+ "early_stopping_rule": "Validation NLL only; up to50 epochs, patience8; actual optimizer budgets may differ due to early stopping",
41
+ "replication_limitation": "Three shared transformer seeds; MLP, GRU and each specialist initially one seed",
42
+ "calibration_limitation": "Nominal90% coverage is measured per field/horizon with widths; correlated windows do not guarantee system-level coverage"
43
+ }
44
+ }
studies/robustness-seed42-20261011/configs/poc.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "version": "tfm-experiment-v1",
3
+ "generation": {"systems": 1200, "days": 7, "dt_minutes": 15, "seed": 42, "families": ["thermal", "tank", "battery"], "policy_weights": [0.35, 0.30, 0.25, 0.10]},
4
+ "windows": {"history_steps": 24, "long_history_steps": 96, "future_steps": 8, "horizons_minutes": [15, 30, 60, 120], "samples": 150000},
5
+ "model": {"d_model": 128, "layers": 4, "heads": 4, "ff_dim": 512},
6
+ "training": {"epochs": 50, "batch_size": 256, "learning_rate": 0.0003, "weight_decay": 0.01, "patience": 8, "gradient_clip": 1.0, "seeds": [42, 43, 44], "precision": "float32"},
7
+ "success_gates": {"normalized_mae_improvement": 0.10, "nominal_90_coverage_range": [0.85, 0.95], "single_query_ms": 20, "intervention_improvement_required": true},
8
+ "status": "configuration only; success gates require completed held-out experiments"
9
+ }
studies/robustness-seed42-20261011/configs/robustness-study.json ADDED
@@ -0,0 +1,38 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "version": "tfm-robustness-study-v1",
3
+ "data": "artifacts/poc/data",
4
+ "manifest_sha256": "bfa4012aa501c8645851c8bf03095b8eacaa63da8f7ae4b2a559ac01d3cd41b7",
5
+ "output": "artifacts/robustness",
6
+ "wait_for_pipeline": "artifacts/full/pipeline.json",
7
+ "reference_checkpoint": "artifacts/full/transformer/seed-42/checkpoint.pt",
8
+ "scaling_status": "deferred_until_robustness_assessment",
9
+ "training": {
10
+ "seed": 42,
11
+ "split_seed": 42,
12
+ "epochs": 50,
13
+ "samples": 150000,
14
+ "validation_samples": 4096,
15
+ "batch_size": 256,
16
+ "device": "mps",
17
+ "learning_rate": 0.0003,
18
+ "weight_decay": 0.01,
19
+ "warmup_fraction": 0.05,
20
+ "patience": 8
21
+ },
22
+ "controls": [
23
+ {"name": "no-history", "history_steps": 24, "no_history": true},
24
+ {"name": "history96", "history_steps": 96, "no_history": false},
25
+ {"name": "leave-thermal-out", "history_steps": 24, "families": ["tank", "battery"], "excluded_family": "thermal", "samples": 100000},
26
+ {"name": "leave-tank-out", "history_steps": 24, "families": ["thermal", "battery"], "excluded_family": "tank", "samples": 100000},
27
+ {"name": "leave-battery-out", "history_steps": 24, "families": ["thermal", "tank"], "excluded_family": "battery", "samples": 100000}
28
+ ],
29
+ "evaluation": {"samples": 4096, "calibration_samples": 4096, "counterfactual_samples": 300, "batch_size": 64, "seed": 2026},
30
+ "limitations": [
31
+ "Single-seed controls; early stopping may produce different optimizer budgets.",
32
+ "History96 changes temporal-encoder parameter count and training-anchor eligibility; record both.",
33
+ "Leave-family-out introduces field/control IDs absent during training; this tests the current closed-vocabulary architecture, not arbitrary-schema support.",
34
+ "Same-checkpoint history masking is a distribution-shift diagnostic; fair no-history comparison requires retraining.",
35
+ "Independently implemented thermal dynamics remain synthetic. Third-party simulator/real-world validation must be reported separately.",
36
+ "Context flags cannot detect all latent differences. Calibrated coverage is conditional on the evaluated population, not universal."
37
+ ]
38
+ }
studies/robustness-seed42-20261011/configs/smoke.json ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ {
2
+ "generation": {"systems": 90, "days": 2, "seed": 42},
3
+ "training": {"epochs": 5, "samples": 2048, "validation_samples": 256, "batch_size": 64, "seed": 42},
4
+ "evaluation": {"samples": 256, "seed": 2026},
5
+ "purpose": "End-to-end implementation validation, not research gate evidence"
6
+ }
studies/robustness-seed42-20261011/docs/context-uncertainty.md ADDED
@@ -0,0 +1,125 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Useful predictions with limited context
2
+
3
+ `scripts/context_uncertainty.py` adds observation quality flags and optional
4
+ residual intervals to the frozen model. It never changes the model or starts
5
+ training. Missing history, historical actions or disturbances, weak observed
6
+ excitation, and incomplete sensors preserve eligibility when at least one state
7
+ observation is available. Eligibility means a provisional estimate is possible;
8
+ it does not certify accuracy or identify the physical system.
9
+
10
+ Only the explicit model observation inputs are assessed: masked history, current
11
+ state, planned actions/exogenous inputs, times, active fields and role IDs. The
12
+ assessor never reads targets, simulator parameters, family labels, actual future
13
+ observations or hidden room size. A small room and an auditorium can have the same
14
+ temperature: no flag can identify their hidden size from that observation alone.
15
+ Low action variation and weak excitation are fixed heuristics, not proofs that a
16
+ parameter is unidentifiable. Missing support violations also cannot certify that
17
+ an unseen system is in distribution.
18
+
19
+ ## Importable helpers
20
+
21
+ ```python
22
+ from scripts.context_uncertainty import (
23
+ fit_context_support, assess_context, fit_context_calibration,
24
+ apply_context_calibration, context_interval_metrics,
25
+ )
26
+
27
+ # Support statistics are fit only on TRAIN observation batches. IDs are caller
28
+ # declarations and must correspond to the actual windows, not merely filenames.
29
+ support = fit_context_support(
30
+ train_batches,
31
+ training_system_ids=train_ids,
32
+ forbidden_system_ids=cal_ids + val_ids + test_ids,
33
+ normalization_sha256=norm_hash,
34
+ transform_id="training-observations-identity-v1",
35
+ )
36
+ cal_context = assess_context(cal_batch, support)
37
+ test_context = assess_context(test_batch, support)
38
+ # Convert the frozen model's log standard deviations to standard deviations.
39
+ artifact = fit_context_calibration(
40
+ cal_mean, np.exp(cal_log_std), cal_target, cal_fields, cal_horizons,
41
+ cal_mask, cal_context,
42
+ calibration_system_ids=cal_ids, training_system_ids=train_ids,
43
+ validation_selection_system_ids=val_ids, test_system_ids=test_ids,
44
+ checkpoint_sha256=[checkpoint_hash], normalization_sha256=norm_hash,
45
+ transform_id="mask-history-except-anchor-v1", min_count=30,
46
+ )
47
+ intervals = apply_context_calibration(
48
+ test_mean, np.exp(test_log_std), test_fields, test_horizons, test_mask,
49
+ test_context, artifact, checkpoint_sha256=[checkpoint_hash],
50
+ normalization_sha256=norm_hash, transform_id="mask-history-except-anchor-v1",
51
+ evaluation_system_ids=test_ids,
52
+ )
53
+ metrics = context_interval_metrics(
54
+ test_mean, test_target, test_mask, test_context, intervals,
55
+ )
56
+ ```
57
+
58
+ All prediction arrays are `[sample, query]`. Assessment inputs are normal model
59
+ batches, including batched role IDs; support accepts a batch or an iterable of
60
+ batches. Means, standard deviations, targets and bounds use normalized state
61
+ deltas. Convert bounds with the existing per-field scale and observed anchor only
62
+ when that anchor is available. The helpers do not invent an unknown anchor.
63
+
64
+ Calibration and evaluation model predictions must be generated after applying the
65
+ same masking transformation. Use a separate artifact for each transformation or
66
+ checkpoint/normalization change. The helpers require matching fingerprints and
67
+ transform identifiers and reject calibration-system overlap with training,
68
+ selection or evaluation systems. Calibration and evaluation assessments must use
69
+ the same support fingerprint. A transform identifier records provenance; it does
70
+ not inspect the prediction-generation code, so the runner must actually perform
71
+ the declared transform before inference.
72
+
73
+ ## Interval interpretation
74
+
75
+ Each field/horizon/flag stratum uses a finite-sample quantile of
76
+ `abs(target - mean) / std` from held-out calibration systems. The default minimum
77
+ is 30 residuals per entry. Insufficient strata fall back to a field/horizon pooled
78
+ entry, with `pooled_fallback_no_conditional_coverage` status. If the pool also has
79
+ too few residuals or the requested finite-sample rank exceeds its count, bounds
80
+ are unavailable (`NaN` arrays and an explicit status), rather than an invented
81
+ broad interval. Counts refer to residuals, not independent systems; overlapping
82
+ windows remain correlated.
83
+
84
+ An observed support shift appends `shift_no_coverage_claim` even when numerical
85
+ intervals are available. No flag or interval provides a generic scalar confidence
86
+ in a room's physical identity. Empirical evaluation reports coverage together
87
+ with width, calibration availability, and errors for flagged and unflagged groups.
88
+ Thresholds are fixed before evaluation; test errors do not select them.
89
+
90
+ Conformal coverage depends on exchangeability of calibration and deployment
91
+ examples. A pooled interval is not a guarantee conditional on a sparse-context
92
+ stratum, and ordinary residual calibration does not establish coverage under
93
+ distribution shift. These restrictions follow the primary treatments in
94
+ [Angelopoulos and Bates](https://arxiv.org/abs/2107.07511) and
95
+ [Tibshirani et al.](https://arxiv.org/abs/1904.06019). This module implements no
96
+ likelihood-ratio weighting or guarantee for unknown physical shift.
97
+
98
+ ## Retrospective behavior assay
99
+
100
+ `scripts/behavior_support.py` complements marginal feature envelopes with an
101
+ observation-only check of past input-to-state behavior. It backtests the frozen
102
+ H24 model from history index 15 to the already observed current value at index 23.
103
+ The prefix contains 16 observations and eight masked padding rows; timestamps are
104
+ shifted to that past anchor. Historical actions and exogenous values at indices
105
+ 15 through 22 supply its eight-transition plan. Request future plans, request
106
+ future targets, hidden parameters and actual room size never enter the assay.
107
+
108
+ The policy is fixed: a field-specific residual divided by predicted standard
109
+ deviation is compared with the 99th percentile of scores from training-system
110
+ histories, with at least 30 scores. Thresholds require matching active schemas
111
+ and complete or partial history availability. In particular, source battery
112
+ histories are transformed to the same SOC-and-requested-power-only schema as
113
+ CityLearn before fitting their thresholds. Missing past observations or drivers
114
+ make the assay unavailable or limited; constant observed drivers retain a low
115
+ excitation limitation even when the past fit is good.
116
+
117
+ The 99th percentile is a training threshold, not a 99 percent confidence or
118
+ coverage claim. Thresholds fitted on the model's own training systems can be
119
+ optimistic, flags across fields increase the any-field false-alarm rate, and
120
+ overlapping windows are correlated. Always report actual held-out source false
121
+ alarms alongside external flag rates and availability. An unflagged window can
122
+ still have shifted physics or a poor future forecast. Future forecast errors may
123
+ be compared after flags are fixed for descriptive analysis; they never select
124
+ the thresholds. The bounded synthetic/CityLearn assay still misses many shifted
125
+ windows and does not certify a new room or battery as safe.
studies/robustness-seed42-20261011/docs/contracts.md ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Internal v0.1 contracts
2
+
3
+ Canonical padded vocabulary F fields, identified by schema FIELD_SPECS (list of FieldSpec).
4
+ FieldSpec: name, unit, role ('state'/'action'/'exogenous'), minimum, maximum; type continuous.
5
+ SCHEMAS maps family string to ordered field names; indices into FIELD_SPECS never change.
6
+ DT_MINUTES = 15; HORIZONS = (1,2,4,8) step offsets.
7
+ Episode dataclass: system_id, family_id, schema (list field names), timestamps [T+1], values [T+1,F], observation_mask [T+1,F], actions [T,F] and exogenous [T,F] (zero padded), hidden_parameters dict, seed. states property alias values. Field padding represented by field_mask [F] from schema; action/exogenous observation values may be inserted into values rows for history, at t corresponding to next applied transition. Explicit arrays source of truth.
8
+ Episode save/load via save_episode(episode,path), load_episode(path). generate_dataset(output,systems,days,seed,families) writes episodes and manifest.json: episodes [{path,system_id,family_id,split}], generator_version, seed, dt_minutes. Relative episode paths. Split train/validation/test by system with stratification. Simulator: make_simulator(family, seed), .parameters dict, .state, .schema list names, .step(action compact or canonical F-vector, exogenous compact or canonical F-vector, dt_minutes=15) returns compact state, .snapshot(), .restore(snapshot). Exogenous trajectory generator documented by data agent.
9
+ Normalization.fit(episodes) only called on training episodes; mean, scale [F] physical values; save(path), load(path), transform(values, mask=None), inverse(values); field deltas standardized using scale (no subtraction of mean). Versioned JSON must include training_system_ids. Bounds are metadata, never hidden capacity.
10
+ WindowDataset(episodes, normalization, history_steps=24, future_steps=8, samples=..., seed=...): deterministic torch Dataset sampled windows without hidden metadata. __getitem__ returns float32 tensors:
11
+ history [H,F] normalized historical values, history_mask [H,F], history_dt [H] minutes relative to current (ends at 0), current [F] normalized current values, current_mask [F], future_actions [K,F] normalized role-specific values with non-role slots zero, future_exogenous [K,F], future_dt [K] minutes since current (15..120), field_mask [F] bool, field_ids [F] int64, unit_ids [F] int64, role_ids [F] int64, query_fields [Q] int64, query_horizons [Q] float minutes, target [Q] normalized delta, target_mask [Q] float/bool. Q all active state fields x offsets 1,2,4,8. history includes current t and preceding H-1 states, actions t-H+1 through t-1; action/exogenous current unavailable, masked at current. Padding query mask false. Fixed Q = max state fields * 4 across families; padded query_fields 0. No system/family/hidden parameters in model input. Collation default.
12
+ models.build_model(kind,config) -> nn.Module. Model forward(batch: dict tensor)-> (delta_mean [B,Q], log_std [B,Q]); supports metadata ids and masks; ignores target keys. config dict history_steps,future_steps,num_fields,d_model,layers,heads,ff_dim,dropout (defaults vocabulary / 128 / 4 / 4 / 512 / 0.0). Kinds transformer/mlp/gru. gaussian_nll(mean,log_std,target,mask).
13
+ training.train CLI: --data path --output path --model transformer|mlp|gru --epochs --batch-size --samples --history-steps --seed --device auto|cpu|mps. Save checkpoint.pt with model_kind, model_config, model_state, seed, epoch; normalization.json and run.json with exact split IDs. load_checkpoint(path,device='cpu')->(model,checkpoint). data path manifest.json. Early stopping uses validation only; test never selected on. Validation split can be subdivided by system for disjoint calibration.
14
+ Evaluation loads same artifacts and independently sampled test windows. evaluation.benchmark CLI --data --checkpoint --output --samples --device. calibration JSON residual quantile multipliers per field/horizon fitted only to calibration systems; API may only call interval calibrated if present and compatible. Baselines persistence/CatBoost/fitted linear system ID all use identical normalization and partitions. Counterfactual snapshots never passed into neural features.
studies/robustness-seed42-20261011/docs/robustness-study.md ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Robustness before scaling
2
+
3
+ The user requested history ablations, harder holdouts and better uncertainty handling before increasing model size. The active full study remains frozen. Scaling is deferred while this evidence is pending, even if the original transformer wins its baseline comparisons.
4
+
5
+ ## Immediate completed-checkpoint diagnostics
6
+
7
+ Use the seed-42 checkpoint without selecting a seed based on test results. `scripts/history_diagnostics.py` evaluates identical test anchors under intact, removed, shortened, shuffled and different-system histories, historical-state-only inputs and thermal-temperature-only sensing. Future action/exogenous plans and targets are protected. Temporal and donor interventions are sensitivity diagnostics, not proof of causal identification. Perturbation-specific calibration uses only the checkpoint's reserved calibration systems; test targets never fit intervals or choose flag thresholds.
8
+
9
+ `scripts/thermal_challenges.py` independently integrates thermal systems with SciPy, including shifted parameter regimes and changed three-node dynamics, ventilation changes and actuator lag. Small-room and auditorium cohorts test effective capacities outside the training range while keeping a meaningful public input schema. Test systems are entirely new. A new integrator alone is not a new physical model; reports distinguish numerical implementation changes, parameter shifts and changed equations.
10
+
11
+ `scripts/evaluate_thermal_challenges.py` compares transformer, persistence and history-fitted linear dynamics on those challenges. Report air-temperature errors and action sensitivity alongside the full field/horizon breakdown. Original-distribution intervals can be measured on challenge data but must be labeled unvalidated under shift. Independently implemented synthetic dynamics are not third-party or real-world validation. BOPTEST or CityLearn validation requires actual enabled actions, validated units and causal input mappings; missing measurements must not be invented.
12
+
13
+ ## Retrained controls
14
+
15
+ `configs/robustness-study.json` records five new serial training controls: no history, 24-hour history, and each of the three leave-family-out models. They retain the same source dataset and established parent system partitions. The two-family runs receive 100,000 windows per epoch to retain approximate 50,000-window exposure per included family. All controls use the existing small architecture, selection-based early stopping and separate output directories. H96 necessarily changes temporal-encoder size and eligible anchors; disclose both. The current family's unseen field embeddings make leave-family-out a challenging test of this implementation, not a claim of unrestricted schema generality.
16
+
17
+ The independent `scripts/robustness_study.py` continuation waits for the existing training/evaluation pipeline to complete, then runs new accelerator jobs serially, records durable receipts and resumes valid epoch-boundary checkpoints. It must not compete with the active MPS queue. Context comparisons use common H96-eligible test anchors. Every leave-family-out evaluation verifies the excluded family in checkpoint provenance.
18
+
19
+ ## Useful uncertainty with partial evidence
20
+
21
+ Keep a mean estimate and a prediction interval when supported, accompanied by observation-only reason codes: missing/short history, absent historical drivers, weak excitation, missing sensors and observed inputs outside training support. Use training observations alone for support statistics. Normal room temperature cannot reveal hidden auditorium-scale capacity by itself; absence of an OOD flag must not be treated as proof that latent dynamics are familiar.
22
+
23
+ `scripts/context_uncertainty.py` assesses evidence without reading targets or hidden simulator metadata. It supports residual calibration by context stratum on disjoint calibration systems with exact transformation/checkpoint/normalization provenance, explicit pooled fallback and insufficient-sample status. Measure coverage and interval width together, plus error for flagged/unflagged groups. Do not manufacture broad intervals or relabel unvalidated shifted intervals as calibrated. The frozen API continues to expose its existing basic flags until a separately versioned integration has been validated; the new assessor is experimental.
24
+
25
+ The user approved intervals plus evidence/context flags rather than a generic confidence percentage. Any later probability-of-accuracy score must define an event, such as error within 1 degree, and have separate empirical calibration.
26
+
27
+ ## Interpretation
28
+
29
+ No single ablation establishes universal generalization. Assess matched-history benefit, intervention accuracy under changed regimes, performance against simpler models, coverage-width tradeoffs, and concrete external validation. Preserve failures as research evidence. Report which tests remain unrun. Increasing width alone does not address failures caused by missing information, schema restrictions or narrow data.
studies/robustness-seed42-20261011/licenses/CityLearn-3.0.2-LICENSE ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ MIT License
2
+
3
+ Copyright (c) 2020 Jose Ramon Vazquez-Canteli, Intelligent Environments Laboratory
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the "Software"), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE.
studies/robustness-seed42-20261011/pyproject.toml ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [build-system]
2
+ requires = ["setuptools>=68"]
3
+ build-backend = "setuptools.build_meta"
4
+
5
+ [project]
6
+ name = "transition-foundation-model"
7
+ version = "0.1.0"
8
+ description = "Schema-aware probabilistic physical-system transitions"
9
+ requires-python = ">=3.11"
10
+ dependencies = ["numpy>=1.26,<3", "scipy>=1.12", "torch>=2.5,<3", "pyarrow>=17", "fastapi>=0.115", "uvicorn>=0.30", "pydantic>=2.9"]
11
+
12
+ [project.optional-dependencies]
13
+ dev = ["pytest>=8", "httpx>=0.27", "ruff>=0.9"]
14
+ baselines = ["catboost>=1.2"]
15
+ export = ["onnx>=1.17", "onnxruntime>=1.20", "onnxscript>=0.2"]
16
+ external = ["minari>=0.5"]
17
+
18
+ [tool.setuptools.packages.find]
19
+ include = ["tfm*"]
20
+
21
+ [tool.pytest.ini_options]
22
+ testpaths = ["tests"]
23
+
24
+ [tool.ruff]
25
+ line-length = 110
26
+
27
+ [tool.ruff.lint]
28
+ select = ["E4", "E7", "E9", "F", "I"]
studies/robustness-seed42-20261011/release.json ADDED
@@ -0,0 +1,1313 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "checkpoint_sha256": "41485d6dd0c72f09a9ca67ced83bc4a0725f9cde731276d1672902304675ebec",
3
+ "files": {
4
+ "LICENSE": {
5
+ "bytes": 1068,
6
+ "sha256": "ba2f860a3c2cba35f85f21ac65fea46c603900544f90ddad503e3ba9e44dce43"
7
+ },
8
+ "README.md": {
9
+ "bytes": 1847,
10
+ "sha256": "2fc0b0828168bfc79931e93ee8f8154963f011f9a83eba3585b358aaa66e747e"
11
+ },
12
+ "artifacts/full/transformer/seed-42/checkpoint.pt": {
13
+ "bytes": 6474292,
14
+ "sha256": "41485d6dd0c72f09a9ca67ced83bc4a0725f9cde731276d1672902304675ebec"
15
+ },
16
+ "artifacts/full/transformer/seed-42/normalization.json": {
17
+ "bytes": 19191,
18
+ "sha256": "33b890560ea8c2a881a2a4e125f19e2beb7a69029ec406d4b7e6a13b9880f3e4"
19
+ },
20
+ "artifacts/full/transformer/seed-42/run.json": {
21
+ "bytes": 190137,
22
+ "sha256": "13c74ef434c51be324c208a534cc31a93374dff44927bafd6a7a9c3d02e2ae42"
23
+ },
24
+ "artifacts/poc/data/manifest.json": {
25
+ "bytes": 295069,
26
+ "sha256": "bfa4012aa501c8645851c8bf03095b8eacaa63da8f7ae4b2a559ac01d3cd41b7"
27
+ },
28
+ "artifacts/robustness/behavior/battery/citylearn_battery_common_scale-assessments.json": {
29
+ "bytes": 354486,
30
+ "sha256": "7696f63f59bfb0bb5e1c36cc405a1685b229984d3d8613e2a79c604dd4dd84f5"
31
+ },
32
+ "artifacts/robustness/behavior/battery/citylearn_battery_scale_shift-assessments.json": {
33
+ "bytes": 355577,
34
+ "sha256": "6a26ca1d2ce0dc6054223dc327ace2c83e79b01cd5614c8d7f81369cd4ef969b"
35
+ },
36
+ "artifacts/robustness/behavior/battery/original_heldout-assessments.json": {
37
+ "bytes": 355134,
38
+ "sha256": "fa8644d2c4a3bafc758ec69fcea9185207ca2b47d8dccebd8d8b7c4a98acb179"
39
+ },
40
+ "artifacts/robustness/behavior/battery/thresholds.json": {
41
+ "bytes": 8851,
42
+ "sha256": "86e2379b21dbf77cb77c703d489f9e7fe3a3c4fb7d836f92190cbc28761144cb"
43
+ },
44
+ "artifacts/robustness/behavior/behavior-diagnostic.json": {
45
+ "bytes": 129773,
46
+ "sha256": "2fd1d7dae2538f18ce9bf74c46846eb468a2a88a09c2fb164ba0154cfbaccf2a"
47
+ },
48
+ "artifacts/robustness/behavior/thermal/auditorium_3node-assessments.json": {
49
+ "bytes": 607713,
50
+ "sha256": "46c9a4d90377deac53c802906de4ae4e9813c236fb09a5ef65fa196a30f5e885"
51
+ },
52
+ "artifacts/robustness/behavior/thermal/original_heldout-assessments.json": {
53
+ "bytes": 551825,
54
+ "sha256": "d37fc14e47a7488e3dc5b18d18246da4aa29025b64c203ec8e27ad1139d8408b"
55
+ },
56
+ "artifacts/robustness/behavior/thermal/parameter_shift_2node-assessments.json": {
57
+ "bytes": 615919,
58
+ "sha256": "edf78b3dbcc92504d135e08a33d4a1b48f1c04b5f983fdc29c43a566b56342bc"
59
+ },
60
+ "artifacts/robustness/behavior/thermal/small_room_3node-assessments.json": {
61
+ "bytes": 616783,
62
+ "sha256": "51a2ab552220624aabc40fb2f71e143eac7b2667b992ce1294ed20c5849f1ecd"
63
+ },
64
+ "artifacts/robustness/behavior/thermal/thresholds.json": {
65
+ "bytes": 9943,
66
+ "sha256": "9fcada67a04b02a7d650e323ac22173bdece47080250f5488ea865dcb444bc35"
67
+ },
68
+ "artifacts/robustness/challenge-interventions/interventions.json": {
69
+ "bytes": 606603,
70
+ "sha256": "c558198983e314b8d9fb3b860c9031dabdc91797a2806349bb66b3d19433a7b0"
71
+ },
72
+ "artifacts/robustness/challenge-interventions/provenance.json": {
73
+ "bytes": 52098,
74
+ "sha256": "35977663cd38d3ca6dbd3edf7c0768174d6ddbf2288797dbcdea13e541de92c6"
75
+ },
76
+ "artifacts/robustness/challenge-interventions/summary.json": {
77
+ "bytes": 35521,
78
+ "sha256": "1580aacc75ecafabd59a2990eaf69d7964a6323bdd16989a79dc56c72c71a452"
79
+ },
80
+ "artifacts/robustness/challenge-interventions/verification.json": {
81
+ "bytes": 562,
82
+ "sha256": "23bbd3238463308df0d3c9569c3d3834c67e858b70f97efe08d77e5c70155886"
83
+ },
84
+ "artifacts/robustness/citylearn/README.md": {
85
+ "bytes": 1961,
86
+ "sha256": "690b26d7c5656d5e23667ea62c943b3d07d38956f8f7c2e45bd4db24bfe01f4c"
87
+ },
88
+ "artifacts/robustness/citylearn/config.json": {
89
+ "bytes": 171,
90
+ "sha256": "703f38e13b5b93ecf2cbeaaeac8d274a30d795976d3f666ec1f6ddd0a851cffc"
91
+ },
92
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_common_scale-283127.npz": {
93
+ "bytes": 6848,
94
+ "sha256": "3a656d7b5a3fb3c0dc6034af2b7ef330e27f7b20f2df3c9022381950a45386e0"
95
+ },
96
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_common_scale-283128.npz": {
97
+ "bytes": 7418,
98
+ "sha256": "0b906e82b0dec454ee92242e96997d9049f0ba74cad88fbbe5d486dbba1e523f"
99
+ },
100
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_common_scale-283129.npz": {
101
+ "bytes": 7640,
102
+ "sha256": "84816b3c049cb99a415efd554aa96a717e41836a22db810c32cb5b53030e71c4"
103
+ },
104
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_common_scale-283130.npz": {
105
+ "bytes": 7247,
106
+ "sha256": "74edf4f38e8f23b2b28a1129505066af6cc67cced5bdc9ca693286461a243b9d"
107
+ },
108
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_common_scale-283131.npz": {
109
+ "bytes": 7366,
110
+ "sha256": "330845003fb6266d69649e765979ccd2ab0ee45723bc9f4c5167cbae2e8b3092"
111
+ },
112
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_common_scale-283132.npz": {
113
+ "bytes": 7312,
114
+ "sha256": "5bb02061fb1b0762f8802cf30b0160980a177bc437e9e186c12cbb00b0ebba9b"
115
+ },
116
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_common_scale-283133.npz": {
117
+ "bytes": 7287,
118
+ "sha256": "eeebd8797fdef8e8ea6423c218b90f3a0e52630464f517fa0d18f1a2a21b1eb1"
119
+ },
120
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_common_scale-283134.npz": {
121
+ "bytes": 6920,
122
+ "sha256": "af5339d1f033dc3f4adc26377d03f1b81c3679c7e23767d1bd4e7046ccd65ca9"
123
+ },
124
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_common_scale-283135.npz": {
125
+ "bytes": 7404,
126
+ "sha256": "8738f1cdb1fc3ee858fc77ee660b2336280a37d2a2f9a502f2fa2401bd583dc0"
127
+ },
128
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_common_scale-283136.npz": {
129
+ "bytes": 7458,
130
+ "sha256": "e522e9223d3999c13b852b237b161ac56b4d5414845a519a6106385ee1deb81b"
131
+ },
132
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_common_scale-283137.npz": {
133
+ "bytes": 7263,
134
+ "sha256": "9ee6c1953f967d938b0318e226024685d80ed94fc0266bab6b181439b4b7c630"
135
+ },
136
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_common_scale-283138.npz": {
137
+ "bytes": 7161,
138
+ "sha256": "d3e3ac191ffa148656dce742e073536d5c1eb8998c394f47d5667aa97ab956c2"
139
+ },
140
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_common_scale-283139.npz": {
141
+ "bytes": 7542,
142
+ "sha256": "985a537ef3cb5b25e3ca6802dafec0d02919b2e7bf0c84c4b611b318c989fe6c"
143
+ },
144
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_common_scale-283140.npz": {
145
+ "bytes": 7526,
146
+ "sha256": "c14dd583ae98e783b6445978e315644bc2ddd145fcbf1d802e577068cf37021a"
147
+ },
148
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_common_scale-283141.npz": {
149
+ "bytes": 7148,
150
+ "sha256": "761fc5439b5e961499541a41196e3be80fd16a1960576e6da74246cf4793c08e"
151
+ },
152
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_common_scale-283142.npz": {
153
+ "bytes": 7264,
154
+ "sha256": "7f96abafde3becae97c2809520227c56e5f7349f4b91b162a68949fe9aaddcd0"
155
+ },
156
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_common_scale-283143.npz": {
157
+ "bytes": 7447,
158
+ "sha256": "d01baabdcc924d66411a4db06335b57419386e270d0e262da50fbb29bed4dc79"
159
+ },
160
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_common_scale-283144.npz": {
161
+ "bytes": 7062,
162
+ "sha256": "0902e10dd7b87eb8b04d9270cfda5faa540eb28463d5b017191c7569cf06022e"
163
+ },
164
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_common_scale-283145.npz": {
165
+ "bytes": 7167,
166
+ "sha256": "e8488ea88f55b569e9e122d43c1cbbc4da214ada235577ee0b2763e2bc0ed45e"
167
+ },
168
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_common_scale-283146.npz": {
169
+ "bytes": 7219,
170
+ "sha256": "f887e01b7eb6975d2bd758292820fd154df1761891cef9be9af902ca0e989cfb"
171
+ },
172
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_common_scale-283147.npz": {
173
+ "bytes": 7201,
174
+ "sha256": "2c53bbb86877cee9340e2da00dd92387c9e252d603f267e7582fbaf24c2c9aba"
175
+ },
176
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_common_scale-283148.npz": {
177
+ "bytes": 7353,
178
+ "sha256": "ef443224210275826550ae2dedfd0ce4f7117a2f1c0c5ca19114d858f73eb83f"
179
+ },
180
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_common_scale-283149.npz": {
181
+ "bytes": 7066,
182
+ "sha256": "3b5fe4e6abd5bda6f34830444f68c185d4d87fce395382100a1b73796038174b"
183
+ },
184
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_common_scale-283150.npz": {
185
+ "bytes": 7283,
186
+ "sha256": "e5b912ee964ffb305b07bbc1bec4ac6954c80ad66f6bbd035450883f2afb9e61"
187
+ },
188
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_scale_shift-383127.npz": {
189
+ "bytes": 7158,
190
+ "sha256": "6412d6baf0d8dd319101fa79b51580b47e58864ce0275d5ff13da73447871122"
191
+ },
192
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_scale_shift-383128.npz": {
193
+ "bytes": 7193,
194
+ "sha256": "7c4d6e551e7ab9c7ab58013cc7fe73d16f465a70caf39f2db69f26e73a1b9518"
195
+ },
196
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_scale_shift-383129.npz": {
197
+ "bytes": 6845,
198
+ "sha256": "5b9850cc0f6083a37f5133d12a985527ca6e9403830f63b9d555c76176ec19f8"
199
+ },
200
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_scale_shift-383130.npz": {
201
+ "bytes": 7132,
202
+ "sha256": "07e1b198dbadee4571cf9ae8bb462083becb45a77e88ca08ef00e7302c8025f1"
203
+ },
204
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_scale_shift-383131.npz": {
205
+ "bytes": 7083,
206
+ "sha256": "a46c8bd1160e440566f53c109663db779e75b7ffc7392ab7739c0788dfdc146f"
207
+ },
208
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_scale_shift-383132.npz": {
209
+ "bytes": 7308,
210
+ "sha256": "6ecd37a48e1067feee56e068cf488f1fb943e175b096a1639e2150906fb63186"
211
+ },
212
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_scale_shift-383133.npz": {
213
+ "bytes": 6833,
214
+ "sha256": "df4d61f69ce47cb019befb40168a1f2d798cec79bb3f6cdb5b3eadeaed048a1f"
215
+ },
216
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_scale_shift-383134.npz": {
217
+ "bytes": 7176,
218
+ "sha256": "0be8416fbc74db11968a8ee3297ddd679652d41d250019f0d3e7c7ca597755f3"
219
+ },
220
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_scale_shift-383135.npz": {
221
+ "bytes": 7004,
222
+ "sha256": "12f0ec3e5da009ad5fcd0b9dfb03498a9a04dd28a9e5857cd1a1503126dcecab"
223
+ },
224
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_scale_shift-383136.npz": {
225
+ "bytes": 7382,
226
+ "sha256": "da317f6280925b26081ed1ca16480b64939e8e3ac456120acf7ea0cb0f0508e9"
227
+ },
228
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_scale_shift-383137.npz": {
229
+ "bytes": 7084,
230
+ "sha256": "e3cccafe45147443aad53046dd790839b5c6e16cba8080f502a16222a98de3e0"
231
+ },
232
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_scale_shift-383138.npz": {
233
+ "bytes": 7309,
234
+ "sha256": "389029c45375c86472856388ae015cc47881dab29107001eee29db958fa8ce73"
235
+ },
236
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_scale_shift-383139.npz": {
237
+ "bytes": 7209,
238
+ "sha256": "4123544a9b719bf31e226ea087d19eb12a35d0d7a1a3fe23dcfe9a0f7632f6c4"
239
+ },
240
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_scale_shift-383140.npz": {
241
+ "bytes": 7271,
242
+ "sha256": "5afa22cfa22164836b9af26741171e7be27fd9cc112928a93d6797d4fb3e4209"
243
+ },
244
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_scale_shift-383141.npz": {
245
+ "bytes": 7365,
246
+ "sha256": "ce3e1e8354fe9162722ca156be6ea8a95e2346f5c251e27043f8af8355ad2a6e"
247
+ },
248
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_scale_shift-383142.npz": {
249
+ "bytes": 7321,
250
+ "sha256": "568b1350bd88b461e29012d83fc9349aec01dfb47c6d554bdd87d2883186daa7"
251
+ },
252
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_scale_shift-383143.npz": {
253
+ "bytes": 7294,
254
+ "sha256": "546b338e261bcf2358ebf1d2d44ba17b3f17fdb6225fad01f6659418bfc06559"
255
+ },
256
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_scale_shift-383144.npz": {
257
+ "bytes": 7091,
258
+ "sha256": "218a807ad448a45cc5de0b8fa5d834962c647f6f54c93633c87f9615b1554db9"
259
+ },
260
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_scale_shift-383145.npz": {
261
+ "bytes": 6878,
262
+ "sha256": "49ba8f48f8bb878778eafe1067878a4ee29937261cbf256b0a40711d32f1b4e5"
263
+ },
264
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_scale_shift-383146.npz": {
265
+ "bytes": 7301,
266
+ "sha256": "8b4e2e86eb7d1c5315aa71badd11a71918280e299dd39e6a2efd32996e52ff07"
267
+ },
268
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_scale_shift-383147.npz": {
269
+ "bytes": 6961,
270
+ "sha256": "183da4c43014589a9a888d7373cddf2399440503bbce1fd03f0c63098ae53614"
271
+ },
272
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_scale_shift-383148.npz": {
273
+ "bytes": 7493,
274
+ "sha256": "d8bbf04c719018620fb9e4046e274491de49db7e6f608a5488a934400b21e18a"
275
+ },
276
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_scale_shift-383149.npz": {
277
+ "bytes": 7038,
278
+ "sha256": "3e98521775823223184bad88edab14461bba596ebbfbda00bae0dc298465e236"
279
+ },
280
+ "artifacts/robustness/citylearn/episodes/citylearn-device-citylearn_battery_scale_shift-383150.npz": {
281
+ "bytes": 7431,
282
+ "sha256": "90bf17f620b9ef55ad2c90932a993df0b8553c1aca758264df66f28fae39cc37"
283
+ },
284
+ "artifacts/robustness/citylearn/evaluation/citylearn-evaluation.json": {
285
+ "bytes": 100896,
286
+ "sha256": "37d242cc2dbe62e9ad3edc9c201643add9cc1c1a69814e5a4473cf3f56669552"
287
+ },
288
+ "artifacts/robustness/citylearn/evaluation/citylearn_battery_common_scale-context.json": {
289
+ "bytes": 219796,
290
+ "sha256": "fe8e8114a221c30e31c65f48a97dc46c3f5373efe1d81ac766b4049df586b538"
291
+ },
292
+ "artifacts/robustness/citylearn/evaluation/citylearn_battery_common_scale-history_linear-predictions.npz": {
293
+ "bytes": 92929,
294
+ "sha256": "5ddbed3757f16cb0f85bc400548a4cf912658521f4be9b1c0e94153f87304e19"
295
+ },
296
+ "artifacts/robustness/citylearn/evaluation/citylearn_battery_common_scale-persistence-predictions.npz": {
297
+ "bytes": 67436,
298
+ "sha256": "d0c788759996e8d8539086a9c2127087d07528a876a90f1e2effca6ab6d2a23f"
299
+ },
300
+ "artifacts/robustness/citylearn/evaluation/citylearn_battery_common_scale-transformer-predictions.npz": {
301
+ "bytes": 98890,
302
+ "sha256": "16a031a5dd0757b0d42f884f9f06337a028a226655c49b0db9b23bcc340e0e8b"
303
+ },
304
+ "artifacts/robustness/citylearn/evaluation/citylearn_battery_scale_shift-context.json": {
305
+ "bytes": 219674,
306
+ "sha256": "3220d9234b201463ec500677c0531294a23abf36a6488a8e8ba196ea803dc854"
307
+ },
308
+ "artifacts/robustness/citylearn/evaluation/citylearn_battery_scale_shift-history_linear-predictions.npz": {
309
+ "bytes": 91587,
310
+ "sha256": "b30c0efd747a5e9b73ea766dda84cec48b3ebaae939dd7733dceeed0909a2ce3"
311
+ },
312
+ "artifacts/robustness/citylearn/evaluation/citylearn_battery_scale_shift-persistence-predictions.npz": {
313
+ "bytes": 60371,
314
+ "sha256": "84ad3863f5e8fa2345c3e20ed8a74aebfc623fc47aef8a68f2a60b16d431816c"
315
+ },
316
+ "artifacts/robustness/citylearn/evaluation/citylearn_battery_scale_shift-transformer-predictions.npz": {
317
+ "bytes": 95744,
318
+ "sha256": "a7a3f2e3edb305e35a21ff73b14d5686c607078cf2a4c66919e7883c12a96b33"
319
+ },
320
+ "artifacts/robustness/citylearn/evaluation/history_linear-sparse-calibration.json": {
321
+ "bytes": 26853,
322
+ "sha256": "8363d6b49c4a1999e8a733217cbbacbe0356e39a8298691cef0b7ec06209ec61"
323
+ },
324
+ "artifacts/robustness/citylearn/evaluation/persistence-sparse-calibration.json": {
325
+ "bytes": 26850,
326
+ "sha256": "e87830f0a2545b552c9d51c85275416f4bf4a1f81e571d0c508828e9725e10e3"
327
+ },
328
+ "artifacts/robustness/citylearn/evaluation/support.json": {
329
+ "bytes": 7576,
330
+ "sha256": "dbdbc4ba688e19ca3f4367a35034cb79dbd29466bfac419587dc1852ff2cf5ec"
331
+ },
332
+ "artifacts/robustness/citylearn/evaluation/transformer-sparse-calibration.json": {
333
+ "bytes": 26847,
334
+ "sha256": "c0fea2ca00b24c37a8a60a6b6eac29e7da3f38c4d343b9a9f675ff3b1d12247d"
335
+ },
336
+ "artifacts/robustness/citylearn/manifest.json": {
337
+ "bytes": 67375,
338
+ "sha256": "2f18ff1700da91594e1fdc1c59c88715c4397ac18523e1c7b8e271c74291aba0"
339
+ },
340
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_common_scale-283127.npz": {
341
+ "bytes": 8091,
342
+ "sha256": "2165efc49ba2d648faf1532c347fb1d36619ba5209d140f8bcf75eb0548cba19"
343
+ },
344
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_common_scale-283128.npz": {
345
+ "bytes": 9255,
346
+ "sha256": "cc37bc707c3c5b1b6638a265d8ac98fd630cb68fe908569a582d875b5890cc4d"
347
+ },
348
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_common_scale-283129.npz": {
349
+ "bytes": 9653,
350
+ "sha256": "7c603724a1b9a30535846ba4e0be170914fb340c6a9ab8a14e6c410364924f07"
351
+ },
352
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_common_scale-283130.npz": {
353
+ "bytes": 9090,
354
+ "sha256": "dfdaa62f61bdec0a78a194a599ebd924728981580519e50e6103130bad211d74"
355
+ },
356
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_common_scale-283131.npz": {
357
+ "bytes": 9275,
358
+ "sha256": "a31e597ec6b34117ee0e8e39a174dbf9336a40e5bb7245050f319314bb694414"
359
+ },
360
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_common_scale-283132.npz": {
361
+ "bytes": 9055,
362
+ "sha256": "b79acce7fcc241d1a23159763fa1039222492553cfa4b6290107c574dfe3157f"
363
+ },
364
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_common_scale-283133.npz": {
365
+ "bytes": 9296,
366
+ "sha256": "afcfa02601b328b771bfcf79a4a172fae18ba2d7906b3d4c7713b4d403d96778"
367
+ },
368
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_common_scale-283134.npz": {
369
+ "bytes": 7849,
370
+ "sha256": "1962aa6e09889bccf0a9ca21e5700d9eb57da0577e63b5420b12e42391ec4d43"
371
+ },
372
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_common_scale-283135.npz": {
373
+ "bytes": 9387,
374
+ "sha256": "f8b7922854465029e6b5944c26e8b098af2436a565c858a86f8f3f2710c55b31"
375
+ },
376
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_common_scale-283136.npz": {
377
+ "bytes": 9458,
378
+ "sha256": "98c0fcc0074741d8b53dd93b6508e9099302472e871f7b3a7e790dab021dae57"
379
+ },
380
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_common_scale-283137.npz": {
381
+ "bytes": 9205,
382
+ "sha256": "74ae97840e37974662e4b5e0afad0705e03a882e22724df200d1b8fcb4058e33"
383
+ },
384
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_common_scale-283138.npz": {
385
+ "bytes": 9028,
386
+ "sha256": "5fdd20680cb157bad663f91f7693c8971e42b9d52ef72a9c96de579eea1f0236"
387
+ },
388
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_common_scale-283139.npz": {
389
+ "bytes": 9474,
390
+ "sha256": "59bf748e766a62fe414df823f8623a288f6789b1191d6b229b20eb0d436df0ae"
391
+ },
392
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_common_scale-283140.npz": {
393
+ "bytes": 9369,
394
+ "sha256": "0527035e40e53f049b1120fbe71f090a3832e0b26b492f345d068cfc7675e8e0"
395
+ },
396
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_common_scale-283141.npz": {
397
+ "bytes": 8442,
398
+ "sha256": "0c186830fa66b17f044c6338491eb9f5d4a8e63a27ee39ba9d52c102ecb966de"
399
+ },
400
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_common_scale-283142.npz": {
401
+ "bytes": 8710,
402
+ "sha256": "9a152c1b7eefe69435338ba748d413d16a84d3bdbc9843a0ee30de8491b80eb1"
403
+ },
404
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_common_scale-283143.npz": {
405
+ "bytes": 9431,
406
+ "sha256": "8878aeceada539d08705be287bb71a4efde8d73c52dfd20afc792a85bde70ecc"
407
+ },
408
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_common_scale-283144.npz": {
409
+ "bytes": 9050,
410
+ "sha256": "56ba723b42400b7d84be3513cd531fdfc830ac2c28d0d1eb8fe092c8d9255320"
411
+ },
412
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_common_scale-283145.npz": {
413
+ "bytes": 9010,
414
+ "sha256": "835f4460d3e1475f914b6721f72b6f5e2ffd1073b2d700135f851213c91c86ae"
415
+ },
416
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_common_scale-283146.npz": {
417
+ "bytes": 8649,
418
+ "sha256": "041ff8353433c6f966def54bd542a4f93723fcba8530d90684c47b42a8874576"
419
+ },
420
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_common_scale-283147.npz": {
421
+ "bytes": 9107,
422
+ "sha256": "29ef659c6f96910bcfbf8f1666c14d3adad186e6486f0d3aefbe7562aa6a0575"
423
+ },
424
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_common_scale-283148.npz": {
425
+ "bytes": 8900,
426
+ "sha256": "2b4823d9f7e916528b4bf0327d0c5869f92cc60ce67993b7083cb1e4635ef484"
427
+ },
428
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_common_scale-283149.npz": {
429
+ "bytes": 8564,
430
+ "sha256": "3d5311a66bbef60f349f8b8b0b0a6ee59867b5ab9457bed19ecac3b1278f0785"
431
+ },
432
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_common_scale-283150.npz": {
433
+ "bytes": 8861,
434
+ "sha256": "fab2face65b42d7689d22f7a141f21a966125dbebd7395f51e20642575a2db53"
435
+ },
436
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_scale_shift-383127.npz": {
437
+ "bytes": 7592,
438
+ "sha256": "85120699ca7618de8e028d925b7d563b11673090426d6308c9d0cb0a8d1051c3"
439
+ },
440
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_scale_shift-383128.npz": {
441
+ "bytes": 8700,
442
+ "sha256": "b1278fe2525b65f74097f4cedd99c34dda03347e03b5d0df4d32c2304b85952f"
443
+ },
444
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_scale_shift-383129.npz": {
445
+ "bytes": 7646,
446
+ "sha256": "cd32591eba15230b0d09349d2fec39931ed6df724a46d07bcfba8fabc96967d1"
447
+ },
448
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_scale_shift-383130.npz": {
449
+ "bytes": 8319,
450
+ "sha256": "053a1bc8f9413ce72514413744028c71caeca48713dcfbad3a9dc9d0dd5abc88"
451
+ },
452
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_scale_shift-383131.npz": {
453
+ "bytes": 8480,
454
+ "sha256": "f4bdf4b595efab8b2638afc76404bfceaede8cf376f07513a23c58404aa7305a"
455
+ },
456
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_scale_shift-383132.npz": {
457
+ "bytes": 8982,
458
+ "sha256": "a10f4b29563d1ee4644d5cde8ad5e59bd14ab15c3280106d368d0ac7a52492a2"
459
+ },
460
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_scale_shift-383133.npz": {
461
+ "bytes": 7580,
462
+ "sha256": "d7b00e4a45276b32770952a976b8eb0282d55a632011e4ddfa17985e2436ae61"
463
+ },
464
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_scale_shift-383134.npz": {
465
+ "bytes": 8827,
466
+ "sha256": "e8a58c8e4f404962c7d4863d535dc071634ba7291283c89bad7f5b99882be760"
467
+ },
468
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_scale_shift-383135.npz": {
469
+ "bytes": 8575,
470
+ "sha256": "fae09169da6085792d576ddced74c0fd8a476075ae92f8b314d118bf7052b239"
471
+ },
472
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_scale_shift-383136.npz": {
473
+ "bytes": 9162,
474
+ "sha256": "592fbaa417c5bc22f7276a5631b8881987ce2488ff8c4fc8d4e15b7aaab6cdb8"
475
+ },
476
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_scale_shift-383137.npz": {
477
+ "bytes": 8344,
478
+ "sha256": "eb5768551660060b74b18e0ff020d4af9a3798aaf1dced98d522ace64645d705"
479
+ },
480
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_scale_shift-383138.npz": {
481
+ "bytes": 8855,
482
+ "sha256": "e3239c9daca3239feefa41dd74cb46aa2db36cb1df611b7c0178d85086a517a6"
483
+ },
484
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_scale_shift-383139.npz": {
485
+ "bytes": 7776,
486
+ "sha256": "22f17f1cd26d5c1079e0e648f2afe30e122fb8f35984b90578861a178f9f8e26"
487
+ },
488
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_scale_shift-383140.npz": {
489
+ "bytes": 9029,
490
+ "sha256": "f420935b2af62c7babfa64140937c6ddf3be375a35a471d9862e70ba480d990b"
491
+ },
492
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_scale_shift-383141.npz": {
493
+ "bytes": 8872,
494
+ "sha256": "723d39d540cb4653158c7c45e494a7d0bb4293c680c10a48094e537264adcbc6"
495
+ },
496
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_scale_shift-383142.npz": {
497
+ "bytes": 9042,
498
+ "sha256": "530dfd260d90182de3344dd3bfd4ab139314b77f005111f22bec47063cb84c37"
499
+ },
500
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_scale_shift-383143.npz": {
501
+ "bytes": 8139,
502
+ "sha256": "2fcd6030d66bf540d399f8d43590f96452dd0526c7f19fc5d8026c152ccab4d0"
503
+ },
504
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_scale_shift-383144.npz": {
505
+ "bytes": 8274,
506
+ "sha256": "9fa0a4bed08946b76860c505fabba56323d7ca99e40f6fbf87e352c7992e200f"
507
+ },
508
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_scale_shift-383145.npz": {
509
+ "bytes": 7933,
510
+ "sha256": "bd1d1bf801f33be637a5e28495f84bb67fc56581f7ac26dc141c1086b2a18474"
511
+ },
512
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_scale_shift-383146.npz": {
513
+ "bytes": 9133,
514
+ "sha256": "d644952648233c51997e0d804cbadb2632710fd71a3d1091838e0b1551117115"
515
+ },
516
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_scale_shift-383147.npz": {
517
+ "bytes": 7805,
518
+ "sha256": "18049abb942329f7425e136e3e136421b80334fbb0362b12674382a7d51cb736"
519
+ },
520
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_scale_shift-383148.npz": {
521
+ "bytes": 9214,
522
+ "sha256": "8f2e77fe40ca2b77f09e7a75cac0ebdcfc7d359903aab8a90c305495d15fc7e8"
523
+ },
524
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_scale_shift-383149.npz": {
525
+ "bytes": 7757,
526
+ "sha256": "161ca0e4eb80a0b1967067d9ab85f1f4045e67ae231d612057e35f10b54e5a49"
527
+ },
528
+ "artifacts/robustness/citylearn/native-audit/citylearn-device-citylearn_battery_scale_shift-383150.npz": {
529
+ "bytes": 8950,
530
+ "sha256": "400afa7160a10701e3ee4d0031e37c6c92de662cb814752181f6a170f61f55ad"
531
+ },
532
+ "artifacts/robustness/citylearn/verification.json": {
533
+ "bytes": 339,
534
+ "sha256": "b10367d242f6ca0bafe11a6b1d068778d4064c3a2e28d5a3d86e56bcf8d4f169"
535
+ },
536
+ "artifacts/robustness/history/calibration-current_only.json": {
537
+ "bytes": 27811,
538
+ "sha256": "4b1fda0fb4d38a7c7306c93c84fef6b17fcf38850d495fd1d5257b0bae86216d"
539
+ },
540
+ "artifacts/robustness/history/calibration-full_history.json": {
541
+ "bytes": 27822,
542
+ "sha256": "83fe0d8bf2142383c9086fdef3e2f508094e60b078aab9df26239a06eb4e104e"
543
+ },
544
+ "artifacts/robustness/history/calibration-history_branch_disabled.json": {
545
+ "bytes": 27811,
546
+ "sha256": "0f4154262c04696d40aa6b7a828874893f2732f1981f93ea5e8fabfedddb1869"
547
+ },
548
+ "artifacts/robustness/history/calibration-matched_family_swap.json": {
549
+ "bytes": 27806,
550
+ "sha256": "f55a8bb058b612b7c0b94e95de68b0b15cabe306a8d5ccf63d5b796f261b024f"
551
+ },
552
+ "artifacts/robustness/history/calibration-past_state_only.json": {
553
+ "bytes": 27812,
554
+ "sha256": "a777fdd83b5547963a28e4cc0d4f604feb325035c26c3d527fcf4e1ac5176fce"
555
+ },
556
+ "artifacts/robustness/history/calibration-temporal_shuffle.json": {
557
+ "bytes": 27805,
558
+ "sha256": "43a25d9e34a024ff059b0d5365c87c59a280a28e275a27e32c13d235401408b7"
559
+ },
560
+ "artifacts/robustness/history/calibration-thermal_temperature_only.json": {
561
+ "bytes": 27187,
562
+ "sha256": "9fc985eec8e4cf2a14bf5f5d1ae914096b064a22e9787432506cf40c7ba4658d"
563
+ },
564
+ "artifacts/robustness/history/calibration-truncate_4.json": {
565
+ "bytes": 27816,
566
+ "sha256": "efe0df5815eeb180509db240981193e08e3e7cb2a072061d9df570fe9a169555"
567
+ },
568
+ "artifacts/robustness/history/calibration-truncate_8.json": {
569
+ "bytes": 27815,
570
+ "sha256": "5e95f2bc33633df3371254ed31bae814d1fb18aec7c8044fff1f3b44f223086d"
571
+ },
572
+ "artifacts/robustness/history/current_only-calibration.npz": {
573
+ "bytes": 82153,
574
+ "sha256": "3a094d50ab0dddc70a919f31a294c6283fe375be19f00279688333f1cbc27de4"
575
+ },
576
+ "artifacts/robustness/history/current_only-test.npz": {
577
+ "bytes": 160503,
578
+ "sha256": "788b5f5f17fa38785bf5dfe0c90a3569f8e6afd6882b1c74fbaaac04f215e10b"
579
+ },
580
+ "artifacts/robustness/history/full_history-calibration.npz": {
581
+ "bytes": 89257,
582
+ "sha256": "6383f9263fc371bb716f643137e26b734a2d7c937e39042357b394062f9e4c88"
583
+ },
584
+ "artifacts/robustness/history/full_history-test.npz": {
585
+ "bytes": 174850,
586
+ "sha256": "fbc1501f289b8bb8bf5f439a54a96dbf6121cb243dcb621d4224f8af0364c844"
587
+ },
588
+ "artifacts/robustness/history/history-diagnostics.json": {
589
+ "bytes": 676430,
590
+ "sha256": "2a2a44d85306f3f560593e240995d7861c0e44ada24549152f9d06f944524cd2"
591
+ },
592
+ "artifacts/robustness/history/history_branch_disabled-calibration.npz": {
593
+ "bytes": 81889,
594
+ "sha256": "60bb54cb16d74de879f90e205f5d84aba0bed9980f95b4219eda743862edeeb8"
595
+ },
596
+ "artifacts/robustness/history/history_branch_disabled-test.npz": {
597
+ "bytes": 160051,
598
+ "sha256": "ea0b575fdb03e38f8bc41a60ad46aa78a38d76c6da4bd3875ca983f2aa7d9780"
599
+ },
600
+ "artifacts/robustness/history/matched_family_swap-calibration.npz": {
601
+ "bytes": 89646,
602
+ "sha256": "db6e8eca61ebc39d4e4a35be219a2a69e13dacc50657ac2d412652668c7c92f2"
603
+ },
604
+ "artifacts/robustness/history/matched_family_swap-test.npz": {
605
+ "bytes": 175341,
606
+ "sha256": "d2c7f2e7e387effbe4f9685752335a00e28e74a2d1618c0b35bff00b6895b06d"
607
+ },
608
+ "artifacts/robustness/history/past_state_only-calibration.npz": {
609
+ "bytes": 87630,
610
+ "sha256": "cd8bcf85d23895f7a01c55257a93e591af424d1e7ca776d218be2e67d507dd5d"
611
+ },
612
+ "artifacts/robustness/history/past_state_only-test.npz": {
613
+ "bytes": 171575,
614
+ "sha256": "bea6faaaa7e7064d004a5a2704304c25ca2366767e22957bfa4de3ad18638810"
615
+ },
616
+ "artifacts/robustness/history/provenance.json": {
617
+ "bytes": 427728,
618
+ "sha256": "6ce68f1f54cff30fd0c9699bac235eb77b6e9326d9ff7d4ac93128af626d8c18"
619
+ },
620
+ "artifacts/robustness/history/summary.json": {
621
+ "bytes": 4002,
622
+ "sha256": "0144bf198cdfd4a66fad46967c7ad10b1024a796eede274ebb2e6908e4f05d83"
623
+ },
624
+ "artifacts/robustness/history/temporal_shuffle-calibration.npz": {
625
+ "bytes": 89869,
626
+ "sha256": "d8440855974c3dc767c0354a912bab5b961374ebd586863b145538701211db25"
627
+ },
628
+ "artifacts/robustness/history/temporal_shuffle-test.npz": {
629
+ "bytes": 176276,
630
+ "sha256": "51a093233d877c57c35bb964c20ea913243456aa2ad3c2ad6a2fd972a761bcc7"
631
+ },
632
+ "artifacts/robustness/history/thermal_temperature_only-calibration.npz": {
633
+ "bytes": 88553,
634
+ "sha256": "7f3da8c6427eff9e88f7375dd01f7e203b346ba513e741456e05cb0d8949b60b"
635
+ },
636
+ "artifacts/robustness/history/thermal_temperature_only-test.npz": {
637
+ "bytes": 173390,
638
+ "sha256": "fff0f371ed9e9446b895bcab23769075e6dde35122e8273bc227b28267336052"
639
+ },
640
+ "artifacts/robustness/history/truncate_4-calibration.npz": {
641
+ "bytes": 83994,
642
+ "sha256": "eb5d97a9fe51af808c42b07e28e7f7fb826440eb795e15d95ccbc06081a6900e"
643
+ },
644
+ "artifacts/robustness/history/truncate_4-test.npz": {
645
+ "bytes": 164343,
646
+ "sha256": "546125f1d16ab9887b47cb52a735e402c338959a25a508c4019d501e45ef40aa"
647
+ },
648
+ "artifacts/robustness/history/truncate_8-calibration.npz": {
649
+ "bytes": 85435,
650
+ "sha256": "c8dabdee185dc72eda7f4db1dd94cbd224a50a936016c5477ea1f65ca230998e"
651
+ },
652
+ "artifacts/robustness/history/truncate_8-test.npz": {
653
+ "bytes": 167113,
654
+ "sha256": "e052b15d95f30b80d9bb353dc2031b769fcc666eaefe18d08d192c11385d48c6"
655
+ },
656
+ "artifacts/robustness/preliminary-results.md": {
657
+ "bytes": 6627,
658
+ "sha256": "0d9ac2969f909500dbd43998b8169c820acb9d5fc5c8aff5faa256891c3c3122"
659
+ },
660
+ "artifacts/robustness/preliminary-verification.json": {
661
+ "bytes": 2110,
662
+ "sha256": "971df9a8db923a00dbf49cd2b4de31f336c421712dc06cee94bd1b7eb99472ee"
663
+ },
664
+ "artifacts/robustness/thermal-challenges/README.md": {
665
+ "bytes": 1308,
666
+ "sha256": "728d82cbc917ebf05012116ad25897ac4398c0cb94caf20dc7f08eb4a7547aff"
667
+ },
668
+ "artifacts/robustness/thermal-challenges/config.json": {
669
+ "bytes": 258,
670
+ "sha256": "8d884e7b9249df93d30e0ebdc82155960b33e8b21483a9954388b64efcb07162"
671
+ },
672
+ "artifacts/robustness/thermal-challenges/episodes/challenge-auditorium_3node-1041230.npz": {
673
+ "bytes": 15490,
674
+ "sha256": "909c4df4c3e16f8c9d6927b7b538cdd0bdc3773e46b648e31ab35fe233d2ba95"
675
+ },
676
+ "artifacts/robustness/thermal-challenges/episodes/challenge-auditorium_3node-1041231.npz": {
677
+ "bytes": 15601,
678
+ "sha256": "cb101b90a5649efc479bb5d8a812a20e814a420c12129d81b225d307b191d825"
679
+ },
680
+ "artifacts/robustness/thermal-challenges/episodes/challenge-auditorium_3node-1041232.npz": {
681
+ "bytes": 15481,
682
+ "sha256": "797163c7811deedcbbda4f8434512b1e9712a77b9dddacd3f2d28f7e02b0449c"
683
+ },
684
+ "artifacts/robustness/thermal-challenges/episodes/challenge-auditorium_3node-1041233.npz": {
685
+ "bytes": 15233,
686
+ "sha256": "08b48742612b07b0a97b2bad7f4c03e0745c78404808b04469b1f47e46caff8a"
687
+ },
688
+ "artifacts/robustness/thermal-challenges/episodes/challenge-auditorium_3node-1041234.npz": {
689
+ "bytes": 15318,
690
+ "sha256": "3b1e7873e7db1b4a1191c85a82461b75040967f73962e3b05cb8d53d150a8d20"
691
+ },
692
+ "artifacts/robustness/thermal-challenges/episodes/challenge-auditorium_3node-1041235.npz": {
693
+ "bytes": 15489,
694
+ "sha256": "d75e02840e80cc8fbfd4488662a0db5c610847f4e57aeca447714ab6b1ba3766"
695
+ },
696
+ "artifacts/robustness/thermal-challenges/episodes/challenge-auditorium_3node-1041236.npz": {
697
+ "bytes": 15037,
698
+ "sha256": "7802684e0b5608ad658b6daafcca4aada0eb3bf68dc9c11239d9f1ebd4cc3438"
699
+ },
700
+ "artifacts/robustness/thermal-challenges/episodes/challenge-auditorium_3node-1041237.npz": {
701
+ "bytes": 15399,
702
+ "sha256": "7c920dada0b947d5c419ec3b8bfd945e9fbbcd96100bcaf029ac303e17a85eda"
703
+ },
704
+ "artifacts/robustness/thermal-challenges/episodes/challenge-auditorium_3node-1041238.npz": {
705
+ "bytes": 15563,
706
+ "sha256": "aa2e43363343f3f825705ddf61d0d491ee830ed87f9fefb6b360e17534468bc7"
707
+ },
708
+ "artifacts/robustness/thermal-challenges/episodes/challenge-auditorium_3node-1041239.npz": {
709
+ "bytes": 15333,
710
+ "sha256": "a3381546d34eaba15c1556b78771ab3ddfe4701cc266e24c92c780f859ccb616"
711
+ },
712
+ "artifacts/robustness/thermal-challenges/episodes/challenge-auditorium_3node-1041240.npz": {
713
+ "bytes": 15398,
714
+ "sha256": "7fda0071578634c230c7b0075b20a40939c688183688d3bdd4766a450b86b799"
715
+ },
716
+ "artifacts/robustness/thermal-challenges/episodes/challenge-auditorium_3node-1041241.npz": {
717
+ "bytes": 15344,
718
+ "sha256": "bf5e5024f9d6c677ee46665af75b09ca093944396a0729814059d29bdcdc9e38"
719
+ },
720
+ "artifacts/robustness/thermal-challenges/episodes/challenge-auditorium_3node-1041242.npz": {
721
+ "bytes": 15181,
722
+ "sha256": "a307e864df7e4a04711a150be16600647d78f4e7b7f6ad472803100652ca2620"
723
+ },
724
+ "artifacts/robustness/thermal-challenges/episodes/challenge-auditorium_3node-1041243.npz": {
725
+ "bytes": 15107,
726
+ "sha256": "bbadc4730e9f34e2c4dfd05c1b453489a687ee4f8f8cbd6440d1f27051cd4fe8"
727
+ },
728
+ "artifacts/robustness/thermal-challenges/episodes/challenge-auditorium_3node-1041244.npz": {
729
+ "bytes": 15244,
730
+ "sha256": "98a6a98da2590f92f5d07aa5c17f515b0d55f25286943d3798f52ce83d0022b1"
731
+ },
732
+ "artifacts/robustness/thermal-challenges/episodes/challenge-auditorium_3node-1041245.npz": {
733
+ "bytes": 15451,
734
+ "sha256": "20c19df7e264082055d9fc64e12d7076e631e777d2d5160e05711b529318bdb6"
735
+ },
736
+ "artifacts/robustness/thermal-challenges/episodes/challenge-auditorium_3node-1041246.npz": {
737
+ "bytes": 15249,
738
+ "sha256": "412c2843571614356c67318322f30ffabec7043fc8b01f79735930784cb141ca"
739
+ },
740
+ "artifacts/robustness/thermal-challenges/episodes/challenge-auditorium_3node-1041247.npz": {
741
+ "bytes": 15379,
742
+ "sha256": "04cba1063d4f7ff83ad34cd4ccb7b9a49c1e9edbdf3fb9ff108fe73fe2a2100a"
743
+ },
744
+ "artifacts/robustness/thermal-challenges/episodes/challenge-auditorium_3node-1041248.npz": {
745
+ "bytes": 15138,
746
+ "sha256": "76353c885a993fff08524ef8516367875d74d79be905f4d4e5705eb8700765f5"
747
+ },
748
+ "artifacts/robustness/thermal-challenges/episodes/challenge-auditorium_3node-1041249.npz": {
749
+ "bytes": 15319,
750
+ "sha256": "6f1ecd89b215404fdeb2c395fcc21a374b7dcce15d46aeb09a1a8348aea451b5"
751
+ },
752
+ "artifacts/robustness/thermal-challenges/episodes/challenge-auditorium_3node-1041250.npz": {
753
+ "bytes": 15365,
754
+ "sha256": "47ac85908002c115eaacc257c84a3dfb614a31d4c54c69238c4cde0cfefda159"
755
+ },
756
+ "artifacts/robustness/thermal-challenges/episodes/challenge-auditorium_3node-1041251.npz": {
757
+ "bytes": 15239,
758
+ "sha256": "12b3f1d1f776e348bcc59a91b9aeef7376fe08cd53f4695efcc43ccf0d654e0e"
759
+ },
760
+ "artifacts/robustness/thermal-challenges/episodes/challenge-auditorium_3node-1041252.npz": {
761
+ "bytes": 15537,
762
+ "sha256": "8f9e82273cf94eb7ef675fbbfdb0bf18c42dabca995c061cae02c9160c688fb6"
763
+ },
764
+ "artifacts/robustness/thermal-challenges/episodes/challenge-auditorium_3node-1041253.npz": {
765
+ "bytes": 15437,
766
+ "sha256": "84d907e9ba3111bb87ff9bcacafac6ec24f2c822585bbf0d233bba87ebf2a318"
767
+ },
768
+ "artifacts/robustness/thermal-challenges/episodes/challenge-parameter_shift_2node-841230.npz": {
769
+ "bytes": 15317,
770
+ "sha256": "6f39bab4317626cb46e49588877a361fec560a65727c5f7b227f24a396db8dbf"
771
+ },
772
+ "artifacts/robustness/thermal-challenges/episodes/challenge-parameter_shift_2node-841231.npz": {
773
+ "bytes": 15259,
774
+ "sha256": "9748e865dd8c3511b479217e550006c51ce0775f2c86a8722d68206e50c77c15"
775
+ },
776
+ "artifacts/robustness/thermal-challenges/episodes/challenge-parameter_shift_2node-841232.npz": {
777
+ "bytes": 15234,
778
+ "sha256": "612e3e996bd4bef2b270abe04047780a9d52eb7e663453c52d1acc62f09ab507"
779
+ },
780
+ "artifacts/robustness/thermal-challenges/episodes/challenge-parameter_shift_2node-841233.npz": {
781
+ "bytes": 15304,
782
+ "sha256": "a6dbbde1e006cf8a63611b0e626d8880198c3cb68eb7f4cf6df356fced9fa025"
783
+ },
784
+ "artifacts/robustness/thermal-challenges/episodes/challenge-parameter_shift_2node-841234.npz": {
785
+ "bytes": 15138,
786
+ "sha256": "35112a75b7a84d8feae255e0bdba38393852f70a379ad8b2be950c363bb54077"
787
+ },
788
+ "artifacts/robustness/thermal-challenges/episodes/challenge-parameter_shift_2node-841235.npz": {
789
+ "bytes": 15172,
790
+ "sha256": "a55856798c69008ca186e6ecbfa72e5f1008b2313886682a0abea6b5f6f6bb04"
791
+ },
792
+ "artifacts/robustness/thermal-challenges/episodes/challenge-parameter_shift_2node-841236.npz": {
793
+ "bytes": 15124,
794
+ "sha256": "6f1a7accdd6716b081974151d632b6d108e9965d3aefc9023e0deb4966d723d5"
795
+ },
796
+ "artifacts/robustness/thermal-challenges/episodes/challenge-parameter_shift_2node-841237.npz": {
797
+ "bytes": 15147,
798
+ "sha256": "8800e013e1e92fff833c38ef71120f5bb6ca9e4c8e186248e741e5da1e59290d"
799
+ },
800
+ "artifacts/robustness/thermal-challenges/episodes/challenge-parameter_shift_2node-841238.npz": {
801
+ "bytes": 15028,
802
+ "sha256": "0ccf54ae0c2fa75e4e5e5bd4f54ddba92a62984fd018c80ab5e5187b52cc4bb0"
803
+ },
804
+ "artifacts/robustness/thermal-challenges/episodes/challenge-parameter_shift_2node-841239.npz": {
805
+ "bytes": 15078,
806
+ "sha256": "ed88464398eee594dfc1454c42f77abd14042b237295f7d721c16d59c1f99a55"
807
+ },
808
+ "artifacts/robustness/thermal-challenges/episodes/challenge-parameter_shift_2node-841240.npz": {
809
+ "bytes": 15176,
810
+ "sha256": "6ef701a2f2fe539c0ac12ad37f2bfd3d7f5b99cd30c5b106d4fddede16db98b0"
811
+ },
812
+ "artifacts/robustness/thermal-challenges/episodes/challenge-parameter_shift_2node-841241.npz": {
813
+ "bytes": 15390,
814
+ "sha256": "93fdd052163100a2e77a42284c7d6ac64a777622b18c88407e62e88f85daf85b"
815
+ },
816
+ "artifacts/robustness/thermal-challenges/episodes/challenge-parameter_shift_2node-841242.npz": {
817
+ "bytes": 14976,
818
+ "sha256": "0c2bde725f7b553eee8dfd4e203bb3e72643398d80e354edef5e960c89937028"
819
+ },
820
+ "artifacts/robustness/thermal-challenges/episodes/challenge-parameter_shift_2node-841243.npz": {
821
+ "bytes": 15281,
822
+ "sha256": "18f98668832e8ccbd1b2f779287d1395f40f8d58621cda285137666b1b93083d"
823
+ },
824
+ "artifacts/robustness/thermal-challenges/episodes/challenge-parameter_shift_2node-841244.npz": {
825
+ "bytes": 15477,
826
+ "sha256": "2f2c54cb541044d3062422c086ec54b2061257bd6616d06f3afb01f9a773cd33"
827
+ },
828
+ "artifacts/robustness/thermal-challenges/episodes/challenge-parameter_shift_2node-841245.npz": {
829
+ "bytes": 15191,
830
+ "sha256": "5e78f8f28c110165e230bd3ad4c1b7c2af164881fdb1f966ad95e0d6fd2b5156"
831
+ },
832
+ "artifacts/robustness/thermal-challenges/episodes/challenge-parameter_shift_2node-841246.npz": {
833
+ "bytes": 15155,
834
+ "sha256": "ec3b33d6367463721e34cb4c57efa7bee626de9a43b5368eb1bd518e53644adf"
835
+ },
836
+ "artifacts/robustness/thermal-challenges/episodes/challenge-parameter_shift_2node-841247.npz": {
837
+ "bytes": 15236,
838
+ "sha256": "309a99fa1882dfe77abec497c511e60280ea0deb3efa9137c28bf6b2d6c3d037"
839
+ },
840
+ "artifacts/robustness/thermal-challenges/episodes/challenge-parameter_shift_2node-841248.npz": {
841
+ "bytes": 15232,
842
+ "sha256": "db9138d5e2893fbbb068ad7a4abf09f630ebf977f219ff86891d667b6c8d7673"
843
+ },
844
+ "artifacts/robustness/thermal-challenges/episodes/challenge-parameter_shift_2node-841249.npz": {
845
+ "bytes": 14966,
846
+ "sha256": "55b3b2ef27c54859f216fdd0f30edb113d57dfb27a2c3e0ff919b91a72baf6a7"
847
+ },
848
+ "artifacts/robustness/thermal-challenges/episodes/challenge-parameter_shift_2node-841250.npz": {
849
+ "bytes": 14972,
850
+ "sha256": "e82eb18a63bcecaa07655f8bfabc119eab4aee397aad3f274f1088cb2985ddae"
851
+ },
852
+ "artifacts/robustness/thermal-challenges/episodes/challenge-parameter_shift_2node-841251.npz": {
853
+ "bytes": 15241,
854
+ "sha256": "59c8f9dab9e8b94ec8ac4f31af34755a51f140c649e907705971e7eb4843e88d"
855
+ },
856
+ "artifacts/robustness/thermal-challenges/episodes/challenge-parameter_shift_2node-841252.npz": {
857
+ "bytes": 15050,
858
+ "sha256": "8687142c157c1973d5ba13a1871f5774591b2cbf4c4648d11eb5131da21f8191"
859
+ },
860
+ "artifacts/robustness/thermal-challenges/episodes/challenge-parameter_shift_2node-841253.npz": {
861
+ "bytes": 15238,
862
+ "sha256": "460f8ef7608c9b3bc8ffeec3017128d256e7bd2f6b02ce642306bc55fa0e5406"
863
+ },
864
+ "artifacts/robustness/thermal-challenges/episodes/challenge-small_room_3node-941230.npz": {
865
+ "bytes": 15321,
866
+ "sha256": "eff1fdbf357934387d9c86f61b4e448576e303382a8c5bcfd55885c688e4c89f"
867
+ },
868
+ "artifacts/robustness/thermal-challenges/episodes/challenge-small_room_3node-941231.npz": {
869
+ "bytes": 15484,
870
+ "sha256": "7df8a10f5859612d7b948b25ee05a7e7d426e338605ec8d16f6c083b8ec98b1c"
871
+ },
872
+ "artifacts/robustness/thermal-challenges/episodes/challenge-small_room_3node-941232.npz": {
873
+ "bytes": 15480,
874
+ "sha256": "7f0d2d0e22fea02d388a1c7309bda3a6e1b96eff4b02b81194ddf6e18c6b5ac9"
875
+ },
876
+ "artifacts/robustness/thermal-challenges/episodes/challenge-small_room_3node-941233.npz": {
877
+ "bytes": 15303,
878
+ "sha256": "b92dc8b4fafce39c4b63b9defd136b4bd1f895c822f67255ef48eaf94a5bf4fa"
879
+ },
880
+ "artifacts/robustness/thermal-challenges/episodes/challenge-small_room_3node-941234.npz": {
881
+ "bytes": 15626,
882
+ "sha256": "6ff72d6f58bd2ea1b0119e87eb036c56bafebb48a360a7b62e69de30820f4726"
883
+ },
884
+ "artifacts/robustness/thermal-challenges/episodes/challenge-small_room_3node-941235.npz": {
885
+ "bytes": 15172,
886
+ "sha256": "bc62810f057c326d68b28497d618438bd3853aa1f9964dded6e538390f4ee63d"
887
+ },
888
+ "artifacts/robustness/thermal-challenges/episodes/challenge-small_room_3node-941236.npz": {
889
+ "bytes": 15106,
890
+ "sha256": "bbff3ad44986c27ef4dc94de5bbf1f6d556cd0fea08a55ef47a6979925690f2d"
891
+ },
892
+ "artifacts/robustness/thermal-challenges/episodes/challenge-small_room_3node-941237.npz": {
893
+ "bytes": 15388,
894
+ "sha256": "a3bf02e113176c1b9d201f643cb23e5ef7f675a5a4209672d3bf472fc4d2d535"
895
+ },
896
+ "artifacts/robustness/thermal-challenges/episodes/challenge-small_room_3node-941238.npz": {
897
+ "bytes": 15389,
898
+ "sha256": "ecf10d4247e7ee09b9850f85e6da7f2f698241c834b6adcfbb3b38c44addaad2"
899
+ },
900
+ "artifacts/robustness/thermal-challenges/episodes/challenge-small_room_3node-941239.npz": {
901
+ "bytes": 15433,
902
+ "sha256": "f1a5ecea1c670170ccf129a7ce960a674f4c805f49afce358c7787fb334dcecc"
903
+ },
904
+ "artifacts/robustness/thermal-challenges/episodes/challenge-small_room_3node-941240.npz": {
905
+ "bytes": 15406,
906
+ "sha256": "08da29ee77eb6c1428cbd64e7a5dbd1d1011aba96d334acb773421b03d34af85"
907
+ },
908
+ "artifacts/robustness/thermal-challenges/episodes/challenge-small_room_3node-941241.npz": {
909
+ "bytes": 15608,
910
+ "sha256": "2f39fbdadd72c4b54867bdb16aa5363b76c0688ded7e7d60f358c233c709ddda"
911
+ },
912
+ "artifacts/robustness/thermal-challenges/episodes/challenge-small_room_3node-941242.npz": {
913
+ "bytes": 15486,
914
+ "sha256": "0c15d7d2963085871a778ff723df58f1dacf2f78bb1448d00019dec6768d7cba"
915
+ },
916
+ "artifacts/robustness/thermal-challenges/episodes/challenge-small_room_3node-941243.npz": {
917
+ "bytes": 15554,
918
+ "sha256": "459cd4d44b839d99111f9181e32e472e1e4994126a890fe38018d3bb9d5b8c62"
919
+ },
920
+ "artifacts/robustness/thermal-challenges/episodes/challenge-small_room_3node-941244.npz": {
921
+ "bytes": 15280,
922
+ "sha256": "f32b1cd4cc1f3a92c2498302a7aa1de363db2b6f95286f8285a1556fb7ac2b9e"
923
+ },
924
+ "artifacts/robustness/thermal-challenges/episodes/challenge-small_room_3node-941245.npz": {
925
+ "bytes": 15281,
926
+ "sha256": "0a3f4c9e209134a12dd486ac9196b5037be9db9020a3ce3f9a2ec6a681e058f8"
927
+ },
928
+ "artifacts/robustness/thermal-challenges/episodes/challenge-small_room_3node-941246.npz": {
929
+ "bytes": 15244,
930
+ "sha256": "ad81061d314250c36028184084db1b375549526389833fe254a6618a47d9c30a"
931
+ },
932
+ "artifacts/robustness/thermal-challenges/episodes/challenge-small_room_3node-941247.npz": {
933
+ "bytes": 15304,
934
+ "sha256": "d43fc56346191c048ddd07e606cb566937cce73d964887d63d063e2e343ce45d"
935
+ },
936
+ "artifacts/robustness/thermal-challenges/episodes/challenge-small_room_3node-941248.npz": {
937
+ "bytes": 15214,
938
+ "sha256": "8bce37f0e3bf02097cbf830d5b62520c524ca3e4e61239bad3a124dc6c0dfcf9"
939
+ },
940
+ "artifacts/robustness/thermal-challenges/episodes/challenge-small_room_3node-941249.npz": {
941
+ "bytes": 15220,
942
+ "sha256": "c65bd39ef584f0908fb2061be0d8a4b2617dbf6b40544175218a4d08b7dbf22c"
943
+ },
944
+ "artifacts/robustness/thermal-challenges/episodes/challenge-small_room_3node-941250.npz": {
945
+ "bytes": 15390,
946
+ "sha256": "4cc062ad3832f9c5d84ba57d33bd6820b604dabd35e5e162f7a0e8b8751edbec"
947
+ },
948
+ "artifacts/robustness/thermal-challenges/episodes/challenge-small_room_3node-941251.npz": {
949
+ "bytes": 15607,
950
+ "sha256": "478876d37fe92ab306ef0b57c95013f1592f7b563a0c28896ad334d2b9e23b80"
951
+ },
952
+ "artifacts/robustness/thermal-challenges/episodes/challenge-small_room_3node-941252.npz": {
953
+ "bytes": 15384,
954
+ "sha256": "ebf7c7726ba56b9eff24d17df9ad5db68a0772e14eb55f3d18d6aa3157aee7e1"
955
+ },
956
+ "artifacts/robustness/thermal-challenges/episodes/challenge-small_room_3node-941253.npz": {
957
+ "bytes": 15331,
958
+ "sha256": "9ab95a58eaab18d39ad2bde69ddbdb48c95f65a499df19a1e6ff3dea6f33c1f4"
959
+ },
960
+ "artifacts/robustness/thermal-challenges/manifest.json": {
961
+ "bytes": 101401,
962
+ "sha256": "a3b9d662ff085243832ff1cc17d6cd5122e7179791f0d712069d3358ac389d9f"
963
+ },
964
+ "artifacts/robustness/thermal-challenges/verification.json": {
965
+ "bytes": 371,
966
+ "sha256": "8ea5fe8970bf88fa9a7241caabac218870c9d3459941332920fc1fb14f7a23e9"
967
+ },
968
+ "artifacts/robustness/thermal-evaluation/current_only/auditorium_3node-context.json": {
969
+ "bytes": 475518,
970
+ "sha256": "40f3dd62cf84535793263e8101e3198f481b5e41a1a49e0246410f6e177fa510"
971
+ },
972
+ "artifacts/robustness/thermal-evaluation/current_only/auditorium_3node-predictions.npz": {
973
+ "bytes": 123767,
974
+ "sha256": "a34a38ff9264d315106aa98b33a02c32f18cad7fe057f51336ea528a161b186f"
975
+ },
976
+ "artifacts/robustness/thermal-evaluation/current_only/calibration.json": {
977
+ "bytes": 43239,
978
+ "sha256": "e6f787063eacbce062b24025b9772d8baba53a13764f4442ba0a8db7966be079"
979
+ },
980
+ "artifacts/robustness/thermal-evaluation/current_only/parameter_shift_2node-context.json": {
981
+ "bytes": 467074,
982
+ "sha256": "1f099edadf69ae75b1c39c350a4c0d9c2431a4ba3d386d045cc018b57255e9c1"
983
+ },
984
+ "artifacts/robustness/thermal-evaluation/current_only/parameter_shift_2node-predictions.npz": {
985
+ "bytes": 121579,
986
+ "sha256": "5345e1042a1c9d89604671b8892b9b3b9d34b9fc2a835e097f42d2512ce65ea0"
987
+ },
988
+ "artifacts/robustness/thermal-evaluation/current_only/small_room_3node-context.json": {
989
+ "bytes": 475364,
990
+ "sha256": "065a2223901c5089e8217859a5004be9a4e3e30a3d545984310ecdae6873b22c"
991
+ },
992
+ "artifacts/robustness/thermal-evaluation/current_only/small_room_3node-predictions.npz": {
993
+ "bytes": 122659,
994
+ "sha256": "f119ea07d493e7efcec5a79acd21bee815fbe7d1c5a47f0211d5d73b80c24019"
995
+ },
996
+ "artifacts/robustness/thermal-evaluation/current_only/support.json": {
997
+ "bytes": 24200,
998
+ "sha256": "17629e4c0582e9803b8c942158335b915e347cff4858dff8cd6f15829f5e895b"
999
+ },
1000
+ "artifacts/robustness/thermal-evaluation/full_history/auditorium_3node-context.json": {
1001
+ "bytes": 304654,
1002
+ "sha256": "4a8b32aef8dde9ccd1a22d091edd292066157fda55b38e1320219ffd7b78d428"
1003
+ },
1004
+ "artifacts/robustness/thermal-evaluation/full_history/auditorium_3node-predictions.npz": {
1005
+ "bytes": 123912,
1006
+ "sha256": "bcb5acbc99ea8f73b65abfb9495c253154ae5e3be7806bf79740c964a77e6079"
1007
+ },
1008
+ "artifacts/robustness/thermal-evaluation/full_history/calibration.json": {
1009
+ "bytes": 46552,
1010
+ "sha256": "83edbd7344fb5f767384a22fdc3e70144b1520003b52db2e4e8b67bf0ed4456a"
1011
+ },
1012
+ "artifacts/robustness/thermal-evaluation/full_history/parameter_shift_2node-context.json": {
1013
+ "bytes": 293530,
1014
+ "sha256": "47c2318d6b155bbcca054f67d0d0171174ce44a570e278fc204add04f07dd624"
1015
+ },
1016
+ "artifacts/robustness/thermal-evaluation/full_history/parameter_shift_2node-predictions.npz": {
1017
+ "bytes": 122761,
1018
+ "sha256": "01aee73627cde532e0e973b80a1d97b837684e6cae6878756b7d55a1e642b0c0"
1019
+ },
1020
+ "artifacts/robustness/thermal-evaluation/full_history/small_room_3node-context.json": {
1021
+ "bytes": 300433,
1022
+ "sha256": "1550a63d836c9138d73641ff8243e106b5f92e703d290f0e4bd001a053877501"
1023
+ },
1024
+ "artifacts/robustness/thermal-evaluation/full_history/small_room_3node-predictions.npz": {
1025
+ "bytes": 124068,
1026
+ "sha256": "9052b77ab6661143e5688260ea8cdd84b5e52bc12b4d1d7d3b751d1e85f34b9e"
1027
+ },
1028
+ "artifacts/robustness/thermal-evaluation/full_history/support.json": {
1029
+ "bytes": 26818,
1030
+ "sha256": "e8459445a41b3ab862d2ad942d9af57f84d5596715628d5fbf80907607fe27ba"
1031
+ },
1032
+ "artifacts/robustness/thermal-evaluation/thermal-evaluation.json": {
1033
+ "bytes": 257262,
1034
+ "sha256": "19d9b9c4116509e5cc7f1124808a5c5f2c0691a30af53dc53581b279d71c62ec"
1035
+ },
1036
+ "artifacts/robustness/thermal-evaluation/thermal_temperature_only/auditorium_3node-context.json": {
1037
+ "bytes": 329706,
1038
+ "sha256": "cb828f093dfd6f8b57ef7e7814f1c459ec06f5248de63e9ceb892c00ea328367"
1039
+ },
1040
+ "artifacts/robustness/thermal-evaluation/thermal_temperature_only/auditorium_3node-predictions.npz": {
1041
+ "bytes": 89530,
1042
+ "sha256": "6a4002c042779a1bbab8709c0c89b11428d2f1e4aa64e6756cce6d0e0d17aa72"
1043
+ },
1044
+ "artifacts/robustness/thermal-evaluation/thermal_temperature_only/calibration.json": {
1045
+ "bytes": 40987,
1046
+ "sha256": "9dc836aa7e47538b04d31c641d198aea37b634098964fe32e5e2e77f8090013d"
1047
+ },
1048
+ "artifacts/robustness/thermal-evaluation/thermal_temperature_only/parameter_shift_2node-context.json": {
1049
+ "bytes": 318594,
1050
+ "sha256": "faa41a69c307447c613981f5e5cdedade72730e95446a2612e4864ab86f24b99"
1051
+ },
1052
+ "artifacts/robustness/thermal-evaluation/thermal_temperature_only/parameter_shift_2node-predictions.npz": {
1053
+ "bytes": 87944,
1054
+ "sha256": "b984d04ffef460e34270c8d850a0337f3a59c00cc52ce0a5ae8fc48cd941e30a"
1055
+ },
1056
+ "artifacts/robustness/thermal-evaluation/thermal_temperature_only/small_room_3node-context.json": {
1057
+ "bytes": 317105,
1058
+ "sha256": "4bfdbb4be5159a58cecfba8dd1f9e3f0ec3d6fd259ed3b96f64367dfb4299a76"
1059
+ },
1060
+ "artifacts/robustness/thermal-evaluation/thermal_temperature_only/small_room_3node-predictions.npz": {
1061
+ "bytes": 89255,
1062
+ "sha256": "e45f4434fd6c625969ecb9b4bf00b2a88ad2930fc4fb0361e6a77d84e8f3ca34"
1063
+ },
1064
+ "artifacts/robustness/thermal-evaluation/thermal_temperature_only/support.json": {
1065
+ "bytes": 26141,
1066
+ "sha256": "edafa68742986afb6abc3bc1c325ca21d13ee4aa6b80a4da15415ab20bab2dbe"
1067
+ },
1068
+ "configs/full-study.json": {
1069
+ "bytes": 1743,
1070
+ "sha256": "c3988bca613ff29bab16a80ab12b7e58e385085b82bdca6d7a2ae9d6a1b8c326"
1071
+ },
1072
+ "configs/poc.json": {
1073
+ "bytes": 843,
1074
+ "sha256": "ad133d98de7bccc0130e11077b7503c4d40ac67633e6765496fdfc3d5bdf65fc"
1075
+ },
1076
+ "configs/robustness-study.json": {
1077
+ "bytes": 2143,
1078
+ "sha256": "210c4425b9c931c612e852e2402021b20399fac368c2c3840c499288caa52cb1"
1079
+ },
1080
+ "configs/smoke.json": {
1081
+ "bytes": 291,
1082
+ "sha256": "1913ec4925af9029694c4e1ffc8cb51b1a29ea53f850f1f0f8f68c610f0bc52e"
1083
+ },
1084
+ "docs/context-uncertainty.md": {
1085
+ "bytes": 7087,
1086
+ "sha256": "7feab5d6e5a479dd23cd55af3926cb430902940b55d62bc8882a6f337dea06c1"
1087
+ },
1088
+ "docs/contracts.md": {
1089
+ "bytes": 4177,
1090
+ "sha256": "3029bf21f52c59bd67460abb162a19b174b8597085382c67334777e128eebc98"
1091
+ },
1092
+ "docs/robustness-study.md": {
1093
+ "bytes": 5008,
1094
+ "sha256": "cd3349693d1fcc8ae8371a1cf951706881dea2b1efd6080ed76f1b4799421fdc"
1095
+ },
1096
+ "licenses/CityLearn-3.0.2-LICENSE": {
1097
+ "bytes": 1141,
1098
+ "sha256": "5a136b692e5288cfc83099df5f21d4dc6ebbb20303ceaf7116f231158c333ea3"
1099
+ },
1100
+ "pyproject.toml": {
1101
+ "bytes": 772,
1102
+ "sha256": "254144cbe213232223dc274f607069f715ca3844f67d6799dbfe93d5e4cbfcd8"
1103
+ },
1104
+ "scripts/behavior_support.py": {
1105
+ "bytes": 26207,
1106
+ "sha256": "fde731bd86dbd989031b2ba34d31dfa483c34b476753d09bfd5575110480b27d"
1107
+ },
1108
+ "scripts/build_robustness_release.py": {
1109
+ "bytes": 10602,
1110
+ "sha256": "ab2a1aba89abf527757a839c12978554ed3ea6956b0472d790a0c0e863bdfd24"
1111
+ },
1112
+ "scripts/challenge_interventions.py": {
1113
+ "bytes": 17189,
1114
+ "sha256": "cf7f85ff76891fea7a759b17f8d38a6062cfe6e69154376f3b0fcbbb20f0443f"
1115
+ },
1116
+ "scripts/citylearn_validation.py": {
1117
+ "bytes": 15135,
1118
+ "sha256": "7593a9a71f77e835b06cec23de4a4d4c169aed82c2508f777557b26e3394f2d3"
1119
+ },
1120
+ "scripts/context_uncertainty.py": {
1121
+ "bytes": 19640,
1122
+ "sha256": "1e734bb91f0c34e56010d5f8665e3f05d4d349166baf852947e21e541d6cc989"
1123
+ },
1124
+ "scripts/evaluate_citylearn.py": {
1125
+ "bytes": 15845,
1126
+ "sha256": "bc2085006b4fb12893ad5e6e839f7f47c4288f5f18badc550f35b721e0f50422"
1127
+ },
1128
+ "scripts/evaluate_thermal_challenges.py": {
1129
+ "bytes": 14414,
1130
+ "sha256": "f10b5c12c182361780bdf2548d98aca103b2f3568045a7aaebfbb6da038db9a5"
1131
+ },
1132
+ "scripts/history_diagnostics.py": {
1133
+ "bytes": 17072,
1134
+ "sha256": "bce3097a1271cd787a01be728ab104fae840832681ac512132066c52f3ba45fa"
1135
+ },
1136
+ "scripts/thermal_challenges.py": {
1137
+ "bytes": 17632,
1138
+ "sha256": "4dc1b47589232fcc2619ddc4c5a93e7cce4270cda376d29af2c236fbc67b4162"
1139
+ },
1140
+ "tests/test_behavior_support.py": {
1141
+ "bytes": 7239,
1142
+ "sha256": "ace1139d3866614b34ee73598bfd3cd190c035d446ebd86518ae7bc5d73dea2c"
1143
+ },
1144
+ "tests/test_challenge_interventions.py": {
1145
+ "bytes": 5291,
1146
+ "sha256": "135f07e9ccd3ab92c340813d22e8e2d5e53654c2b6c3502d31feb6abe18fa54c"
1147
+ },
1148
+ "tests/test_citylearn_validation.py": {
1149
+ "bytes": 5800,
1150
+ "sha256": "4a95a1d17b88445d5580f0a952e11eb39d2b7a02835532998a8a2c138f0664c5"
1151
+ },
1152
+ "tests/test_context_uncertainty.py": {
1153
+ "bytes": 6758,
1154
+ "sha256": "c54717a329d3b6ab3e67643f9befd5d74bbb76077640b13cdeb8d3ce3ace0552"
1155
+ },
1156
+ "tests/test_evaluate_citylearn.py": {
1157
+ "bytes": 2742,
1158
+ "sha256": "fed121f6a5e4ec864842d0c256b2bbf75d93af2ac52613269fe7daa5d6b9df00"
1159
+ },
1160
+ "tests/test_history_diagnostics.py": {
1161
+ "bytes": 9349,
1162
+ "sha256": "47cc6e270f9759fb76d3db4e4323b07022beb335615e7ceed087c920b63951ce"
1163
+ },
1164
+ "tests/test_robustness_release.py": {
1165
+ "bytes": 1919,
1166
+ "sha256": "54bd1e9e943e311b6f21484810618b65e847c93e902c71438aa5fec553546c05"
1167
+ },
1168
+ "tests/test_thermal_challenges.py": {
1169
+ "bytes": 5851,
1170
+ "sha256": "06e4930f7253596eb9af4495418b18e1ac92ec6bce674a21a5bdd737f2cc64b7"
1171
+ },
1172
+ "tests/test_thermal_evaluation.py": {
1173
+ "bytes": 6071,
1174
+ "sha256": "d26e8cc8ee9fda48335dc2417e08e92f93b0e56de95d18817e48b164a45055cc"
1175
+ },
1176
+ "tfm/__init__.py": {
1177
+ "bytes": 75,
1178
+ "sha256": "c4f4c4c5f9d9f9d6d8b54e62cbfbf8167b8a5423bf204d809d5135fbefc059fa"
1179
+ },
1180
+ "tfm/api/__init__.py": {
1181
+ "bytes": 0,
1182
+ "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
1183
+ },
1184
+ "tfm/api/profile.py": {
1185
+ "bytes": 3252,
1186
+ "sha256": "9de088be206053b13ab515d12e308d14a93c5fb2ff1db5773c1f6f23777abc32"
1187
+ },
1188
+ "tfm/api/server.py": {
1189
+ "bytes": 18168,
1190
+ "sha256": "428bd963c2833c04b47f19c26573f2f28296585d394d05544ac6de860a47e33d"
1191
+ },
1192
+ "tfm/data/__init__.py": {
1193
+ "bytes": 1495,
1194
+ "sha256": "2bd24e2f5d48edc4678ad827f3a3046de256b8d5731d90f19c6fabee269daa66"
1195
+ },
1196
+ "tfm/data/episodes.py": {
1197
+ "bytes": 5274,
1198
+ "sha256": "a512ba251e45017ee75d9278385bb8f3d26f5ac43bf9d994c912d8e15cead803"
1199
+ },
1200
+ "tfm/data/external.py": {
1201
+ "bytes": 8604,
1202
+ "sha256": "6a00cd71156b423b02a180f5d6db31d83953320ec4380a8ced6f005427df51e3"
1203
+ },
1204
+ "tfm/data/generate.py": {
1205
+ "bytes": 11465,
1206
+ "sha256": "a3cffad4440a1aa9cd41eac48dcf43fbdc19b4d169200e33bf7dcf0e760d33e3"
1207
+ },
1208
+ "tfm/data/normalization.py": {
1209
+ "bytes": 2831,
1210
+ "sha256": "8f51104b43c2855944c550f9f8699654605196c71214e9d7746d90500a05a3a3"
1211
+ },
1212
+ "tfm/data/schema.py": {
1213
+ "bytes": 2744,
1214
+ "sha256": "c1e09c0b273e0cb31044e5d7575b4e6387ff2a5d7fd2d550c6b23bc72ccc2078"
1215
+ },
1216
+ "tfm/data/windows.py": {
1217
+ "bytes": 5071,
1218
+ "sha256": "b23a0ff480d3c49b325abd266298243eae7be9c279ee91516fc4d01c5a145f12"
1219
+ },
1220
+ "tfm/evaluation/__init__.py": {
1221
+ "bytes": 427,
1222
+ "sha256": "fd447a5e6bf4b857d9b0670f8ac8c70111bf3dfe5755eefbe68c4d86b13a36d5"
1223
+ },
1224
+ "tfm/evaluation/baselines.py": {
1225
+ "bytes": 6888,
1226
+ "sha256": "2dd709f31cbef652b252bec7e4ad63425933f7141657e73b9803f4f98eeda843"
1227
+ },
1228
+ "tfm/evaluation/benchmark.py": {
1229
+ "bytes": 37929,
1230
+ "sha256": "2f817104b27c343bab44940840b93a9aae4fb9f42d5851c6496920c1eb5d2155"
1231
+ },
1232
+ "tfm/evaluation/calibration.py": {
1233
+ "bytes": 5811,
1234
+ "sha256": "cf5bdbccccd6f13628ae871b8fd7e7b1900306e1ebdf00003147d083cf3cc5ca"
1235
+ },
1236
+ "tfm/evaluation/catboost_artifact.py": {
1237
+ "bytes": 16461,
1238
+ "sha256": "5b70bb928cc29b4f09cbaacaed716208e1fc5b507f8cd4811b52c770e9268d0d"
1239
+ },
1240
+ "tfm/evaluation/comparison.py": {
1241
+ "bytes": 15257,
1242
+ "sha256": "c53cf03e4a689b2d85c00a52dca6303216c92f61ec91a4b68f66f36a3cc588de"
1243
+ },
1244
+ "tfm/evaluation/context.py": {
1245
+ "bytes": 10633,
1246
+ "sha256": "f7f981c0b9d635681c7c4e8ee606e279b0bf6eacbe3f3a8a2d6a4fd1279e97e9"
1247
+ },
1248
+ "tfm/evaluation/counterfactual.py": {
1249
+ "bytes": 4765,
1250
+ "sha256": "6d8c00284449997e8c338b227ebd2f4d72885c555a6d4bef0fb290742be95259"
1251
+ },
1252
+ "tfm/evaluation/experiments.py": {
1253
+ "bytes": 4616,
1254
+ "sha256": "cb4ad0870c90b7ce54ad082cd5cb8289d4036ff81418732fa5280eee25b5a03d"
1255
+ },
1256
+ "tfm/evaluation/metrics.py": {
1257
+ "bytes": 3557,
1258
+ "sha256": "04016e0647788caa5d5cd1b9a036018761d7867b29f10cc064bbb4ae2351dcd7"
1259
+ },
1260
+ "tfm/evaluation/runner.py": {
1261
+ "bytes": 2048,
1262
+ "sha256": "3296ca3f3fd036a2173edc71c23d6430f57ca8f141a20c8e5a7cc6a20cffd7e0"
1263
+ },
1264
+ "tfm/evaluation/train_catboost.py": {
1265
+ "bytes": 5180,
1266
+ "sha256": "6605555f6139827b57c5485ccd0fbb0cf60367da21c2a12009ef0a48b1cc176f"
1267
+ },
1268
+ "tfm/export/__init__.py": {
1269
+ "bytes": 0,
1270
+ "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
1271
+ },
1272
+ "tfm/export/onnx.py": {
1273
+ "bytes": 7776,
1274
+ "sha256": "3cec8744848d40551ae0c770aa0cc1d4177189a1a1ac20f2bd5c3bd04f2ce40d"
1275
+ },
1276
+ "tfm/models/__init__.py": {
1277
+ "bytes": 157,
1278
+ "sha256": "c20432c425b2305a275f22968a539619a57037c7c8af67baceca979a4ee0e35c"
1279
+ },
1280
+ "tfm/models/core.py": {
1281
+ "bytes": 10257,
1282
+ "sha256": "53c6c32671006759d3a1aeab93cc2349371203414a0a1febf618e9c2b779a0c5"
1283
+ },
1284
+ "tfm/simulators/__init__.py": {
1285
+ "bytes": 156,
1286
+ "sha256": "0c2179025d61bad9948adb2e4ae3130a844ddf92e1094ba3b93836d6d75b1e8b"
1287
+ },
1288
+ "tfm/simulators/physics.py": {
1289
+ "bytes": 11047,
1290
+ "sha256": "03893cd813d86f12d0e335cbdbd4ea588d2c8ac97566254725fb32e774c413cc"
1291
+ },
1292
+ "tfm/training/__init__.py": {
1293
+ "bytes": 200,
1294
+ "sha256": "5fdc517285d815cc8972aac6a60c974f82354ce29f6413d3e0c856c30edcf5bd"
1295
+ },
1296
+ "tfm/training/train.py": {
1297
+ "bytes": 32180,
1298
+ "sha256": "d9e58852db730ccecff009c24654e794b3be40a3d8d6985c4757774ab6e0898d"
1299
+ },
1300
+ "uv.lock": {
1301
+ "bytes": 461099,
1302
+ "sha256": "622b4c7d12d2d20f0061f132b6f50e505de99852a1285ad280cba9b2aa32509f"
1303
+ }
1304
+ },
1305
+ "full_study_complete": false,
1306
+ "normalization_sha256": "33b890560ea8c2a881a2a4e125f19e2beb7a69029ec406d4b7e6a13b9880f3e4",
1307
+ "original_dataset_revision": "1baf2150adb20ff95ae8b81b7f4a985d81edbfe4",
1308
+ "original_manifest_sha256": "bfa4012aa501c8645851c8bf03095b8eacaa63da8f7ae4b2a559ac01d3cd41b7",
1309
+ "scaling_deferred": true,
1310
+ "status": "completed_diagnostics_retrained_controls_pending",
1311
+ "subset": "robustness-seed42-20261011",
1312
+ "version": "tfm-diagnostics-release-v1"
1313
+ }
studies/robustness-seed42-20261011/scripts/behavior_support.py ADDED
@@ -0,0 +1,609 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Retrospective behavior checks using observations already inside H24.
2
+
3
+ Fixed policy: anchor 15, terminal observation 23, 120-minute backtest, 99th
4
+ training-residual quantile, minimum 30 scores, normalized driver std < 0.05.
5
+ No future target, hidden parameter or held-out error participates in flag fitting.
6
+ An unflagged retrospective fit never certifies a new system or future forecast.
7
+ """
8
+
9
+ from __future__ import annotations
10
+
11
+ import argparse
12
+ import hashlib
13
+ import json
14
+ from pathlib import Path
15
+
16
+ import numpy as np
17
+ import torch
18
+
19
+ VERSION = "tfm-retrospective-behavior-support-v1"
20
+ POLICY = {
21
+ "history_steps": 24,
22
+ "anchor_index": 15,
23
+ "end_index": 23,
24
+ "horizon_minutes": 120,
25
+ "quantile": 0.99,
26
+ "quantile_method": "higher",
27
+ "min_count": 30,
28
+ "low_driver_std": 0.05,
29
+ }
30
+ INPUT_KEYS = (
31
+ "history",
32
+ "history_mask",
33
+ "history_dt",
34
+ "current",
35
+ "current_mask",
36
+ "future_actions",
37
+ "future_exogenous",
38
+ "future_dt",
39
+ "field_mask",
40
+ "field_ids",
41
+ "unit_ids",
42
+ "role_ids",
43
+ "query_fields",
44
+ "query_horizons",
45
+ )
46
+ OBSERVATION_KEYS = (
47
+ "history",
48
+ "history_mask",
49
+ "history_dt",
50
+ "current",
51
+ "current_mask",
52
+ "field_mask",
53
+ "field_ids",
54
+ "unit_ids",
55
+ "role_ids",
56
+ )
57
+
58
+
59
+ def _tensor(value):
60
+ return value.detach().cpu() if torch.is_tensor(value) else torch.as_tensor(value)
61
+
62
+
63
+ def _observations(batch):
64
+ # Do not inspect or even convert any other mapping value.
65
+ if not set(OBSERVATION_KEYS) <= batch.keys():
66
+ raise ValueError("Retrospective assessment needs history, time, current masks and schema identities")
67
+ data = {key: _tensor(batch[key]) for key in OBSERVATION_KEYS}
68
+ if data["current"].ndim == 1:
69
+ data = {key: value[None] for key, value in data.items()}
70
+ n, fields = data["current"].shape
71
+ if data["history"].shape != (n, 24, fields) or data["history_mask"].shape != (n, 24, fields):
72
+ raise ValueError("Behavior backtest requires exactly [sample, 24, field] history")
73
+ if data["history_dt"].shape != (n, 24):
74
+ raise ValueError("History time shape must be [sample, 24]")
75
+ for key in OBSERVATION_KEYS[4:]:
76
+ if data[key].shape != (n, fields):
77
+ raise ValueError(f"{key} must match current shape")
78
+ if torch.any(~torch.isin(data["role_ids"], torch.tensor([0, 1, 2]))):
79
+ raise ValueError("Role IDs must use state=0, action=1, exogenous=2")
80
+ return data
81
+
82
+
83
+ def build_history_backtest(batch):
84
+ """Construct past-only network inputs; observed terminal delta stays outside.
85
+
86
+ The H24 prefix contains original rows 0..15, left padded by 8 masked rows.
87
+ The new anchor is original index 15. Its action/exogenous values are masked
88
+ from prefix/current and become the first planned transition, with rows 15..22
89
+ forming the entire plan. Original request future inputs are never inspected.
90
+ """
91
+ data = _observations(batch)
92
+ n, fields = data["current"].shape
93
+ active = data["field_mask"].bool()
94
+ roles = data["role_ids"]
95
+ state = active & (roles == 0)
96
+ driver = active & (roles != 0)
97
+ history = data["history"].float()
98
+ observed = data["history_mask"].bool() & torch.isfinite(history) & active[:, None]
99
+ current_ok = data["current_mask"].bool() & torch.isfinite(data["current"]) & state
100
+ query_mask = state & observed[:, 15] & current_ok
101
+ safe_history = torch.where(observed, history, 0)
102
+ prefix = safe_history[:, :16].clone()
103
+ prefix_mask = observed[:, :16].clone()
104
+ prefix_mask[:, -1] &= state
105
+ prefix[:, -1] = torch.where(prefix_mask[:, -1], prefix[:, -1], 0)
106
+ padding = torch.zeros((n, 8, fields), dtype=prefix.dtype)
107
+ history_backtest = torch.cat((padding, prefix), dim=1)
108
+ mask_backtest = torch.cat((padding.bool(), prefix_mask), dim=1)
109
+ relative = data["history_dt"].float() - data["history_dt"][:, 15:16].float()
110
+ # Padding time slots precede the observed prefix at the canonical cadence.
111
+ padded_times = relative[:, :1] - torch.arange(8, 0, -1).float()[None] * 15
112
+ history_dt = torch.cat((padded_times, relative[:, :16]), dim=1)
113
+ past_plan = safe_history[:, 15:23]
114
+ inputs = {
115
+ "history": history_backtest,
116
+ "history_mask": mask_backtest.float(),
117
+ "history_dt": history_dt,
118
+ "current": prefix[:, -1],
119
+ "current_mask": prefix_mask[:, -1].float(),
120
+ "future_actions": torch.where((active & (roles == 1))[:, None], past_plan, 0),
121
+ "future_exogenous": torch.where((active & (roles == 2))[:, None], past_plan, 0),
122
+ "future_dt": relative[:, 16:24],
123
+ "query_fields": data["field_ids"].long(),
124
+ "query_horizons": torch.full((n, fields), 120.0),
125
+ **{key: data[key].clone() for key in ("field_mask", "field_ids", "unit_ids", "role_ids")},
126
+ }
127
+ delta = data["current"].float() - safe_history[:, 15]
128
+ statuses, reasons, group_keys = [], [], []
129
+ eligible = torch.ones(n, dtype=torch.bool)
130
+ for i in range(n):
131
+ codes = []
132
+ valid_times = torch.isfinite(relative[i]).all() and torch.allclose(
133
+ relative[i],
134
+ torch.arange(-15, 9).float() * 15,
135
+ atol=1e-4,
136
+ rtol=0,
137
+ )
138
+ if not valid_times:
139
+ codes.append("invalid_or_noncanonical_time_context")
140
+ eligible[i] = False
141
+ if not observed[i, :15, state[i]].any():
142
+ codes.append("missing_past_state_history")
143
+ eligible[i] = False
144
+ if not query_mask[i].any():
145
+ codes.append("no_observed_backtest_state_pair")
146
+ eligible[i] = False
147
+ if driver[i].any() and not observed[i, 15:23, driver[i]].all():
148
+ codes.append("missing_known_past_driver_plan")
149
+ eligible[i] = False
150
+ driver_values = [history[i, :23, f][observed[i, :23, f]] for f in torch.where(driver[i])[0]]
151
+ stds = [float(values.std(unbiased=False)) for values in driver_values if len(values) >= 2]
152
+ if not driver_values:
153
+ codes.append("no_observed_driver_fields")
154
+ elif len(stds) != len(driver_values):
155
+ codes.append("insufficient_observed_driver_history")
156
+ elif max(stds) < POLICY["low_driver_std"]:
157
+ codes.append("low_excitation_heuristic")
158
+ complete = bool(observed[i, :16, state[i]].all() and observed[i, :15, driver[i]].all())
159
+ if not complete:
160
+ codes.append("incomplete_past_context")
161
+ schema = [(int(data["field_ids"][i, f]), int(roles[i, f])) for f in torch.where(active[i])[0]]
162
+ # Do not pool across schemas or complete/partial history availability.
163
+ group_keys.append(
164
+ json.dumps(sorted(schema), separators=(",", ":")) + ("|complete" if complete else "|partial")
165
+ )
166
+ statuses.append("unavailable" if not eligible[i] else "limited" if codes else "available")
167
+ reasons.append(codes)
168
+ return {
169
+ "inputs": inputs,
170
+ "observed_delta": delta.numpy(),
171
+ "query_mask": query_mask.numpy(),
172
+ "eligible": eligible.numpy(),
173
+ "availability_status": statuses,
174
+ "reasoning_codes": reasons,
175
+ "group_keys": group_keys,
176
+ "fields": data["field_ids"].numpy(),
177
+ }
178
+
179
+
180
+ def retrospective_behavior_scores(model, batch, *, device="cpu"):
181
+ """Score residual/sigma against an already-observed current sensor value."""
182
+ prepared = build_history_backtest(batch)
183
+ shape = prepared["query_mask"].shape
184
+ mean, std = np.full(shape, np.nan), np.full(shape, np.nan)
185
+ selected = np.flatnonzero(prepared["eligible"])
186
+ if len(selected):
187
+ inputs = {key: prepared["inputs"][key][selected].to(device) for key in INPUT_KEYS}
188
+ with torch.inference_mode():
189
+ predicted, log_std = model(inputs)
190
+ if tuple(predicted.shape) != (len(selected), shape[1]) or log_std.shape != predicted.shape:
191
+ raise ValueError("Backtest model outputs must match [sample, field query]")
192
+ mean[selected] = predicted.detach().cpu().numpy()
193
+ std[selected] = log_std.detach().cpu().exp().numpy()
194
+ valid = (
195
+ prepared["query_mask"]
196
+ & prepared["eligible"][:, None]
197
+ & np.isfinite(mean)
198
+ & np.isfinite(std)
199
+ & (std > 0)
200
+ )
201
+ score = np.full(shape, np.nan)
202
+ score[valid] = np.abs(prepared["observed_delta"] - mean)[valid] / std[valid]
203
+ for i in range(len(valid)):
204
+ if prepared["eligible"][i] and not np.array_equal(valid[i], prepared["query_mask"][i]):
205
+ prepared["reasoning_codes"][i].append("nonfinite_model_backtest")
206
+ prepared["availability_status"][i] = "limited" if valid[i].any() else "unavailable"
207
+ return {
208
+ key: prepared[key] for key in ("fields", "group_keys", "availability_status", "reasoning_codes")
209
+ } | {
210
+ "version": VERSION,
211
+ "policy": dict(POLICY),
212
+ "scores": score,
213
+ "valid": valid,
214
+ "predicted_past_delta": mean,
215
+ "past_std": std,
216
+ "observed_past_delta": prepared["observed_delta"],
217
+ }
218
+
219
+
220
+ def fit_behavior_support(
221
+ results,
222
+ *,
223
+ training_system_ids,
224
+ forbidden_system_ids=(),
225
+ checkpoint_sha256,
226
+ normalization_sha256,
227
+ transform_id,
228
+ source_sha256,
229
+ training_anchor_provenance,
230
+ ):
231
+ """Fit fixed 99th-percentile training scores; no held-out threshold selection."""
232
+ ids = sorted(set(training_system_ids))
233
+ if not ids or set(ids) & set(forbidden_system_ids):
234
+ raise ValueError("Behavior thresholds require disjoint training-only systems")
235
+ if not all(
236
+ (checkpoint_sha256, normalization_sha256, transform_id, source_sha256, training_anchor_provenance)
237
+ ):
238
+ raise ValueError(
239
+ "Behavior threshold provenance must include model, normalization, source, transform and anchors"
240
+ )
241
+ results = [results] if isinstance(results, dict) else results
242
+ collected = {}
243
+ for result in results:
244
+ if result["version"] != VERSION or result["policy"] != POLICY:
245
+ raise ValueError("Behavior policy mismatch")
246
+ for i, j in zip(*np.where(result["valid"])):
247
+ key = result["group_keys"][i] + "|" + str(int(result["fields"][i, j]))
248
+ collected.setdefault(key, []).append(float(result["scores"][i, j]))
249
+ entries = {}
250
+ for key, values in collected.items():
251
+ if not np.isfinite(values).all():
252
+ raise ValueError("Training behavior scores must be finite")
253
+ entries[key] = {
254
+ "count": len(values),
255
+ "threshold": float(np.quantile(values, POLICY["quantile"], method="higher"))
256
+ if len(values) >= POLICY["min_count"]
257
+ else None,
258
+ }
259
+ artifact = {
260
+ "version": VERSION,
261
+ "policy": dict(POLICY),
262
+ "training_system_ids": ids,
263
+ "checkpoint_sha256": sorted(checkpoint_sha256),
264
+ "normalization_sha256": normalization_sha256,
265
+ "transform_id": transform_id,
266
+ "source_sha256": source_sha256,
267
+ "training_anchor_provenance": training_anchor_provenance,
268
+ "entries": entries,
269
+ "interpretation": "Observation-only retrospective residual heuristic; no confidence probability or future coverage guarantee.",
270
+ "limitations": [
271
+ "Thresholds use model training systems and may be optimistic.",
272
+ "Overlapping sampled windows are correlated; no independent-window false-alarm guarantee.",
273
+ "No flag cannot prove a new system safe or identify hidden room size.",
274
+ "A good fit under constant drivers cannot resolve hidden action response.",
275
+ "Prefix history is shortened by 8 masked rows; this is a fixed retrospective assay, not ordinary deployment context.",
276
+ ],
277
+ }
278
+ artifact["sha256"] = hashlib.sha256(json.dumps(artifact, sort_keys=True).encode()).hexdigest()
279
+ return artifact
280
+
281
+
282
+ def assess_behavior_support(
283
+ result,
284
+ artifact,
285
+ *,
286
+ checkpoint_sha256,
287
+ normalization_sha256,
288
+ transform_id,
289
+ evaluation_system_ids=(),
290
+ ):
291
+ if (
292
+ artifact.get("version") != VERSION
293
+ or artifact.get("policy") != POLICY
294
+ or result.get("policy") != POLICY
295
+ ):
296
+ raise ValueError("Behavior support policy mismatch")
297
+ for key, actual in (
298
+ ("checkpoint_sha256", sorted(checkpoint_sha256)),
299
+ ("normalization_sha256", normalization_sha256),
300
+ ("transform_id", transform_id),
301
+ ):
302
+ if artifact[key] != actual:
303
+ raise ValueError(f"Behavior {key} provenance mismatch")
304
+ if set(evaluation_system_ids) & set(artifact["training_system_ids"]):
305
+ raise ValueError("Behavior reference/evaluation systems overlap training thresholds")
306
+ assessments = []
307
+ for i, group in enumerate(result["group_keys"]):
308
+ codes = list(result["reasoning_codes"][i])
309
+ fields, flagged, checked = [], [], 0
310
+ for j in np.flatnonzero(result["valid"][i]):
311
+ field = int(result["fields"][i, j])
312
+ entry = artifact["entries"].get(group + "|" + str(field))
313
+ score = float(result["scores"][i, j])
314
+ threshold = entry["threshold"] if entry else None
315
+ if threshold is not None:
316
+ checked += 1
317
+ if score > threshold:
318
+ flagged.append(field)
319
+ fields.append(
320
+ {
321
+ "field_id": field,
322
+ "score": score,
323
+ "threshold": threshold,
324
+ "training_score_count": entry["count"] if entry else 0,
325
+ "status": "behavior_mismatch"
326
+ if field in flagged
327
+ else "within_training_retrospective_threshold"
328
+ if threshold is not None
329
+ else "unavailable_matching_schema_or_samples",
330
+ }
331
+ )
332
+ if flagged:
333
+ codes.append("observed_behavior_mismatch")
334
+ if not checked:
335
+ codes.append("matching_behavior_threshold_unavailable")
336
+ assessments.append(
337
+ {
338
+ "status": "unavailable" if not fields else "limited" if not checked or codes else "assessed",
339
+ "behavior_flag": bool(flagged),
340
+ "flagged_field_ids": flagged,
341
+ "fields": fields,
342
+ "reasoning_codes": codes,
343
+ "group_key": group,
344
+ "thresholds_available": checked,
345
+ "threshold_sha256": artifact["sha256"],
346
+ "confidence_probability": None,
347
+ "future_targets_consumed": False,
348
+ "hidden_physical_parameters_consumed": False,
349
+ }
350
+ )
351
+ return assessments
352
+
353
+
354
+ def summarize_behavior(assessments):
355
+ count = len(assessments)
356
+ available = sum(bool(row["thresholds_available"]) for row in assessments)
357
+ flags = sum(row["behavior_flag"] for row in assessments)
358
+ by_field = {}
359
+ for row in assessments:
360
+ for entry in row["fields"]:
361
+ field = str(entry["field_id"])
362
+ counts = by_field.setdefault(field, {"available_windows": 0, "flagged_windows": 0})
363
+ if entry["threshold"] is not None:
364
+ counts["available_windows"] += 1
365
+ counts["flagged_windows"] += int(entry["status"] == "behavior_mismatch")
366
+ for counts in by_field.values():
367
+ counts["empirical_flag_rate"] = (
368
+ counts["flagged_windows"] / counts["available_windows"] if counts["available_windows"] else None
369
+ )
370
+ return {
371
+ "windows": count,
372
+ "threshold_available_windows": available,
373
+ "behavior_flagged_windows": flags,
374
+ "behavior_flag_rate_all_windows": flags / count if count else None,
375
+ "behavior_flag_rate_available_windows": flags / available if available else None,
376
+ "limited_windows": sum(row["status"] == "limited" for row in assessments),
377
+ "unavailable_windows": sum(row["status"] == "unavailable" for row in assessments),
378
+ "by_field": by_field,
379
+ "interpretation": "Empirical retrospective flag rate, not a confidence probability.",
380
+ }
381
+
382
+
383
+ def _sha(path):
384
+ return hashlib.sha256(Path(path).read_bytes()).hexdigest()
385
+
386
+
387
+ def _save(path, value):
388
+ path = Path(path)
389
+ path.parent.mkdir(parents=True, exist_ok=True)
390
+ temporary = path.with_suffix(path.suffix + ".tmp")
391
+ temporary.write_text(json.dumps(value, indent=2, allow_nan=False) + "\n")
392
+ temporary.replace(path)
393
+
394
+
395
+ def run_diagnostic(args):
396
+ """Bounded source/thermal/CityLearn assay; future errors are analysis only."""
397
+ from torch.utils.data import DataLoader
398
+
399
+ from scripts.citylearn_validation import COHORTS as CITY_COHORTS
400
+ from scripts.citylearn_validation import load_citylearn_windows
401
+ from scripts.evaluate_citylearn import TRANSFORM, sparse_battery_episode
402
+ from scripts.thermal_challenges import COHORTS as THERMAL_COHORTS
403
+ from scripts.thermal_challenges import load_challenge_windows
404
+ from tfm.data import Normalization, WindowDataset
405
+ from tfm.evaluation.benchmark import collect_predictions, sampling_provenance, summarize
406
+ from tfm.training.train import load_checkpoint, load_manifest_episodes
407
+
408
+ torch.set_num_threads(1)
409
+ checkpoint_path = Path(args.checkpoint)
410
+ norm_path = checkpoint_path.with_name("normalization.json")
411
+ source_path = Path(args.data) / "manifest.json"
412
+ model, checkpoint = load_checkpoint(checkpoint_path, "cpu")
413
+ norm = Normalization.load(norm_path)
414
+ if checkpoint["model_config"]["history_steps"] != 24 or checkpoint["seed"] != 42:
415
+ raise ValueError("Fixed behavior assay requires completed H24 seed-42 checkpoint")
416
+ if (
417
+ _sha(norm_path) != checkpoint["normalization_sha256"]
418
+ or _sha(source_path) != checkpoint["manifest_sha256"]
419
+ ):
420
+ raise ValueError("Frozen normalization/source manifest mismatch")
421
+ if set(norm.training_system_ids) != set(checkpoint["training_system_ids"]):
422
+ raise ValueError("Frozen normalization training provenance mismatch")
423
+ manifest, splits = load_manifest_episodes(args.data)
424
+ used_ids = {
425
+ e.system_id
426
+ for split in ("train", "test")
427
+ for e in splits[split]
428
+ if e.family_id in ("thermal", "battery")
429
+ }
430
+ hashes = {
431
+ entry["path"]: _sha(Path(args.data) / entry["path"])
432
+ for entry in manifest["episodes"]
433
+ if entry["system_id"] in used_ids
434
+ }
435
+ if any(checkpoint["episode_sha256"].get(key) != value for key, value in hashes.items()):
436
+ raise ValueError("Original source episode fingerprint mismatch")
437
+ helpers = [
438
+ Path("scripts/evaluate_citylearn.py"),
439
+ Path("scripts/citylearn_validation.py"),
440
+ Path("scripts/thermal_challenges.py"),
441
+ ]
442
+ files = [
443
+ checkpoint_path,
444
+ norm_path,
445
+ source_path,
446
+ Path(args.thermal_manifest),
447
+ Path(args.citylearn_manifest),
448
+ Path(__file__),
449
+ *helpers,
450
+ ]
451
+ binding = {str(path): _sha(path) for path in files}
452
+ sources = {path: _sha(path) for path in checkpoint["source_sha256"]}
453
+ if sources != checkpoint["source_sha256"]:
454
+ raise ValueError("Frozen model/data source differs from checkpoint")
455
+ sources.update({str(path): _sha(path) for path in [Path(__file__), *helpers]})
456
+ report = {
457
+ "version": VERSION,
458
+ "status": "running",
459
+ "policy": dict(POLICY),
460
+ "binding_sha256": binding,
461
+ "source_sha256": sources,
462
+ "used_source_episode_sha256": hashes,
463
+ "seed": args.seed,
464
+ "samples_per_population": args.samples,
465
+ "profiles": {},
466
+ "weights_updated": False,
467
+ "thresholds_fit_on_training_systems_only": True,
468
+ "future_errors_used_for_flag_fitting": False,
469
+ "real_world_validation": False,
470
+ "confidence_probability": None,
471
+ }
472
+ output = Path(args.output)
473
+ common = {"history_steps": 24, "future_steps": 8, "samples": args.samples}
474
+
475
+ def score(dataset):
476
+ # Mapping may include targets: whitelist conversion prevents consumption.
477
+ return [
478
+ retrospective_behavior_scores(model, batch)
479
+ for batch in DataLoader(dataset, batch_size=args.batch_size, shuffle=False)
480
+ ]
481
+
482
+ for family, transform in (("thermal", "thermal-full-canonical-observations-v1"), ("battery", TRANSFORM)):
483
+ partitions = {name: [ep for ep in eps if ep.family_id == family] for name, eps in splits.items()}
484
+ if family == "battery":
485
+ partitions = {
486
+ name: [sparse_battery_episode(ep) for ep in eps] for name, eps in partitions.items()
487
+ }
488
+ train = WindowDataset(partitions["train"], norm, seed=args.seed, **common)
489
+ original = WindowDataset(partitions["test"], norm, seed=args.seed + 1, **common)
490
+ train_ids = {ep.system_id for ep in partitions["train"]}
491
+ prohibited = set().union(
492
+ *(
493
+ set(checkpoint[key])
494
+ for key in ("validation_selection_system_ids", "calibration_system_ids", "test_system_ids")
495
+ )
496
+ )
497
+ profile = fit_behavior_support(
498
+ score(train),
499
+ training_system_ids=train_ids,
500
+ forbidden_system_ids=prohibited,
501
+ checkpoint_sha256=[binding[str(checkpoint_path)]],
502
+ normalization_sha256=binding[str(norm_path)],
503
+ transform_id=transform,
504
+ source_sha256=sources,
505
+ training_anchor_provenance=sampling_provenance(train, source_path),
506
+ )
507
+ _save(output / family / "thresholds.json", profile)
508
+ populations = [("original_heldout", original, source_path)]
509
+ if family == "thermal":
510
+ populations += [
511
+ (
512
+ cohort,
513
+ load_challenge_windows(
514
+ args.thermal_manifest, norm, cohort, samples=args.samples, seed=args.seed + 1
515
+ ),
516
+ Path(args.thermal_manifest),
517
+ )
518
+ for cohort in THERMAL_COHORTS
519
+ ]
520
+ else:
521
+ populations += [
522
+ (
523
+ cohort,
524
+ load_citylearn_windows(
525
+ args.citylearn_manifest, norm, cohort, 24, args.samples, args.seed + 1
526
+ ),
527
+ Path(args.citylearn_manifest),
528
+ )
529
+ for cohort in CITY_COHORTS
530
+ ]
531
+ rows = {}
532
+ for name, dataset, population_manifest in populations:
533
+ ids = {ep.system_id for ep in dataset.episodes}
534
+ if ids & (
535
+ set(checkpoint["training_system_ids"])
536
+ | set(checkpoint["validation_selection_system_ids"])
537
+ | set(checkpoint["calibration_system_ids"])
538
+ ):
539
+ raise ValueError("Behavior evaluation system leakage")
540
+ assessments = []
541
+ for result in score(dataset):
542
+ assessments += assess_behavior_support(
543
+ result,
544
+ profile,
545
+ checkpoint_sha256=[binding[str(checkpoint_path)]],
546
+ normalization_sha256=binding[str(norm_path)],
547
+ transform_id=transform,
548
+ evaluation_system_ids=ids,
549
+ )
550
+ # Ordinary future forecast labels are used only after flags are fixed,
551
+ # solely to report retrospective flag/error association descriptively.
552
+ mean, log_std, arrays = collect_predictions(model, dataset, "cpu", args.batch_size)
553
+ forecast_rows = summarize(mean, log_std, arrays, [family] * len(dataset), norm)
554
+ selected_error = np.abs(mean - arrays["target"])
555
+ future_mask = arrays["target_mask"].astype(bool)
556
+ flag = np.asarray([row["behavior_flag"] for row in assessments])[:, None]
557
+ error_groups = {}
558
+ for label, chosen in (("flagged", flag), ("unflagged", ~flag)):
559
+ selected = future_mask & chosen
560
+ error_groups[label] = {
561
+ "query_count": int(selected.sum()),
562
+ "normalized_delta_mae": float(selected_error[selected].mean())
563
+ if selected.any()
564
+ else None,
565
+ }
566
+ rows[name] = {
567
+ **summarize_behavior(assessments),
568
+ "anchor_provenance": sampling_provenance(dataset, population_manifest),
569
+ "future_error_analysis_only": {
570
+ "selective_error": error_groups,
571
+ "metrics_by_field_horizon": forecast_rows,
572
+ },
573
+ }
574
+ _save(output / family / (name + "-assessments.json"), {"assessments": assessments})
575
+ print(
576
+ json.dumps({"family": family, "population": name, **summarize_behavior(assessments)}),
577
+ flush=True,
578
+ )
579
+ report["profiles"][family] = {
580
+ "transform_id": transform,
581
+ "threshold_sha256": profile["sha256"],
582
+ "populations": rows,
583
+ }
584
+ _save(output / "behavior-diagnostic.json", report)
585
+ if binding != {str(path): _sha(path) for path in files}:
586
+ raise RuntimeError("Diagnostic source/data/checkpoint binding changed during evaluation")
587
+ report.update(status="completed", frozen_bindings_unchanged=True)
588
+ _save(output / "behavior-diagnostic.json", report)
589
+ return report
590
+
591
+
592
+ def main():
593
+ parser = argparse.ArgumentParser(description=__doc__)
594
+ parser.add_argument("--data", default="artifacts/poc/data")
595
+ parser.add_argument("--checkpoint", default="artifacts/full/transformer/seed-42/checkpoint.pt")
596
+ parser.add_argument("--thermal-manifest", default="artifacts/robustness/thermal-challenges/manifest.json")
597
+ parser.add_argument("--citylearn-manifest", default="artifacts/robustness/citylearn/manifest.json")
598
+ parser.add_argument("--output", default="artifacts/robustness/behavior")
599
+ parser.add_argument("--samples", type=int, default=512)
600
+ parser.add_argument("--batch-size", type=int, default=64)
601
+ parser.add_argument("--seed", type=int, default=20261012)
602
+ args = parser.parse_args()
603
+ if min(args.samples, args.batch_size) < 1:
604
+ parser.error("Sample/batch counts must be positive")
605
+ run_diagnostic(args)
606
+
607
+
608
+ if __name__ == "__main__":
609
+ main()
studies/robustness-seed42-20261011/scripts/build_robustness_release.py ADDED
@@ -0,0 +1,218 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Package completed seed-42 diagnostics, separately from the running full study."""
2
+
3
+ from __future__ import annotations
4
+
5
+ import argparse
6
+ import hashlib
7
+ import json
8
+ import shutil
9
+ from pathlib import Path
10
+
11
+ SUBSET = "robustness-seed42-20261011"
12
+ COHORT_DIRS = (
13
+ "history",
14
+ "thermal-challenges",
15
+ "thermal-evaluation",
16
+ "challenge-interventions",
17
+ "citylearn",
18
+ "behavior",
19
+ )
20
+ HELPERS = (
21
+ "history_diagnostics",
22
+ "thermal_challenges",
23
+ "evaluate_thermal_challenges",
24
+ "challenge_interventions",
25
+ "citylearn_validation",
26
+ "evaluate_citylearn",
27
+ "context_uncertainty",
28
+ "behavior_support",
29
+ "build_robustness_release",
30
+ )
31
+ TESTS = (
32
+ "behavior_support",
33
+ "challenge_interventions",
34
+ "citylearn_validation",
35
+ "context_uncertainty",
36
+ "evaluate_citylearn",
37
+ "history_diagnostics",
38
+ "thermal_challenges",
39
+ "thermal_evaluation",
40
+ "robustness_release",
41
+ )
42
+
43
+
44
+ def sha256(path):
45
+ return hashlib.sha256(Path(path).read_bytes()).hexdigest()
46
+
47
+
48
+ def check_hashes(root, mapping):
49
+ root = Path(root).resolve()
50
+ for name, expected in mapping.items():
51
+ raw = root / name
52
+ path = raw.resolve()
53
+ if raw.is_symlink() or not path.is_relative_to(root) or not path.is_file():
54
+ raise ValueError(f"Unsafe or missing provenance file: {name}")
55
+ if sha256(path) != expected:
56
+ raise ValueError(f"Changed provenance file: {name}")
57
+
58
+
59
+ def copy_files(source, destination):
60
+ if source.is_symlink():
61
+ raise ValueError(f"Symlink cannot be released: {source}")
62
+ paths = [source] if source.is_file() else sorted(source.rglob("*"))
63
+ for path in paths:
64
+ if path.is_symlink():
65
+ raise ValueError(f"Symlink cannot be released: {path}")
66
+ if not path.is_file() or "__pycache__" in path.parts or path.suffix == ".pyc":
67
+ continue
68
+ if path.name in ("last.pt", "runner.lock") or path.suffix in (".log", ".tmp"):
69
+ raise ValueError(f"Unfinished or transient artifact cannot be released: {path}")
70
+ target = destination if source.is_file() else destination / path.relative_to(source)
71
+ target.parent.mkdir(parents=True, exist_ok=True)
72
+ shutil.copyfile(path, target)
73
+
74
+
75
+ def build(root, output, citylearn_license):
76
+ root, output = Path(root).resolve(), Path(output).resolve()
77
+ if output.exists() and any(output.iterdir()):
78
+ raise ValueError("Existing releases cannot be overwritten")
79
+ evidence = root / "artifacts/robustness"
80
+
81
+ def load(p):
82
+ return json.loads(p.read_text())
83
+
84
+ completed = (
85
+ "history/history-diagnostics.json",
86
+ "thermal-evaluation/thermal-evaluation.json",
87
+ "challenge-interventions/interventions.json",
88
+ "citylearn/evaluation/citylearn-evaluation.json",
89
+ "behavior/behavior-diagnostic.json",
90
+ )
91
+ reports = {name: load(evidence / name) for name in completed}
92
+ for name, report in reports.items():
93
+ if report.get("status") not in ("complete", "completed"):
94
+ raise ValueError(f"Diagnostic is unfinished: {name}")
95
+ weights = root / "artifacts/full/transformer/seed-42"
96
+ run = load(weights / "run.json")
97
+ if run.get("status") != "complete":
98
+ raise ValueError("Seed-42 training must be complete")
99
+ cp_hash, norm_hash = sha256(weights / "checkpoint.pt"), sha256(weights / "normalization.json")
100
+ original = root / "artifacts/poc/data"
101
+ check_hashes(original, run["episode_sha256"])
102
+ if sha256(original / "manifest.json") != run["manifest_sha256"]:
103
+ raise ValueError("Original dataset manifest changed")
104
+ if norm_hash != run["normalization_sha256"]:
105
+ raise ValueError("Training normalization changed")
106
+ check_hashes(root, run["source_sha256"])
107
+ history = reports[completed[0]]
108
+ check_hashes(evidence / "history", history["outputs_sha256"])
109
+ hp = load(evidence / "history/provenance.json")
110
+ check_hashes(root, hp["source_sha256"])
111
+ if history["provenance_sha256"] != sha256(evidence / "history/provenance.json"):
112
+ raise ValueError("History provenance changed")
113
+ for report in reports.values():
114
+ if "checkpoint_sha256" in report:
115
+ actual = report["checkpoint_sha256"]
116
+ if actual not in (cp_hash, [cp_hash]):
117
+ raise ValueError("Diagnostic checkpoint mismatch")
118
+ thermal = reports[completed[1]]
119
+ check_hashes(root, thermal["helper_sha256"])
120
+ check_hashes(root, {"scripts/evaluate_thermal_challenges.py": thermal["script_sha256"]})
121
+ for key in ("citylearn/evaluation/citylearn-evaluation.json", "behavior/behavior-diagnostic.json"):
122
+ r = reports[key]
123
+ check_hashes(root, r.get("binding_file_sha256", r.get("binding_sha256", {})))
124
+ check_hashes(root, r.get("source_sha256", r.get("frozen_tfm_sources_sha256", {})))
125
+ if "predictions_sha256" in r:
126
+ check_hashes(evidence / "citylearn/evaluation", r["predictions_sha256"])
127
+ check_hashes(
128
+ evidence / "challenge-interventions",
129
+ load(evidence / "challenge-interventions/verification.json")["outputs_sha256"],
130
+ )
131
+ for cohort in ("thermal-challenges", "citylearn"):
132
+ manifest = load(evidence / cohort / "manifest.json")
133
+ for episode in manifest["episodes"]:
134
+ check_hashes(evidence / cohort, {episode["path"]: episode["sha256"]})
135
+ if "native_audit_path" in episode:
136
+ check_hashes(
137
+ evidence / cohort, {episode["native_audit_path"]: episode["native_audit_sha256"]}
138
+ )
139
+ for name, digest in load(evidence / "preliminary-verification.json")["files_sha256"].items():
140
+ check_hashes(root, {name: digest})
141
+ output.mkdir(parents=True, exist_ok=True)
142
+ for cohort in COHORT_DIRS:
143
+ copy_files(evidence / cohort, output / "artifacts/robustness" / cohort)
144
+ for name in ("preliminary-results.md", "preliminary-verification.json"):
145
+ copy_files(evidence / name, output / "artifacts/robustness" / name)
146
+ copy_files(root / "tfm", output / "tfm")
147
+ for name in HELPERS:
148
+ copy_files(root / f"scripts/{name}.py", output / f"scripts/{name}.py")
149
+ for name in TESTS:
150
+ copy_files(root / f"tests/test_{name}.py", output / f"tests/test_{name}.py")
151
+ for config in sorted((root / "configs").glob("*.json")):
152
+ copy_files(config, output / "configs" / config.name)
153
+ for name in ("context-uncertainty", "robustness-study", "contracts"):
154
+ copy_files(root / f"docs/{name}.md", output / f"docs/{name}.md")
155
+ for name in ("pyproject.toml", "uv.lock", "LICENSE"):
156
+ copy_files(root / name, output / name)
157
+ for name in ("checkpoint.pt", "normalization.json", "run.json"):
158
+ copy_files(weights / name, output / "artifacts/full/transformer/seed-42" / name)
159
+ copy_files(original / "manifest.json", output / "artifacts/poc/data/manifest.json")
160
+ copy_files(Path(citylearn_license), output / "licenses/CityLearn-3.0.2-LICENSE")
161
+ (output / "README.md").write_text(
162
+ "# Completed robustness diagnostics — frozen full-data seed 42\n\n"
163
+ "This is a preliminary research evidence bundle, not a replacement for the v0.1.0 model release. "
164
+ "Only seed 42's full-data training is included. The full three-seed/MLP/GRU/specialist study "
165
+ "and five retrained robustness controls are still pending. Scaling is deferred.\n\n"
166
+ "[Findings](artifacts/robustness/preliminary-results.md) include negative intervention results "
167
+ "and severe interval undercoverage on independent thermal and actual CityLearn battery-device physics. "
168
+ "CityLearn validation is synthetic device validation, not building or real-world validation. "
169
+ "Richer evidence flags remain experimental helpers, not deployed API behavior.\n\n"
170
+ "This folder preserves its own paths and checksums. Download this study prefix, install the pinned "
171
+ "uv.lock environment, and run helpers from this folder. The original 1,200-system dataset is "
172
+ "the immutable archive at [v0.1.0](https://huggingface.co/maxffarrell/transition-foundation-model/"
173
+ "resolve/v0.1.0/datasets/synthetic-1200-v0.1.2.tar.gz); extract it here before original-distribution "
174
+ "diagnostics. Its manifest SHA256 must match release.json. CityLearn-generated episodes and "
175
+ "source-package hashes are included; the third-party runtime is installed separately. "
176
+ "CityLearn's MIT license is retained in licenses/. Historical absolute paths in provenance "
177
+ "identify the generating workstation; executable helper defaults use relative paths.\n\n"
178
+ "Run `uv sync --locked --extra dev --extra baselines --extra export` and "
179
+ "`uv run --no-sync python -m pytest -q`. Native CityLearn generation tests require separately "
180
+ "installed CityLearn 3.0.2 with compatible NumPy<2; data reading and inference do not.\n\n"
181
+ "The original smoke CatBoost counterfactual numbers predate corrected inactive-field padding; "
182
+ "use subsequent corrected reports for action-effect comparisons.\n"
183
+ )
184
+ inventory = {
185
+ str(p.relative_to(output)): {"bytes": p.stat().st_size, "sha256": sha256(p)}
186
+ for p in sorted(output.rglob("*"))
187
+ if p.is_file()
188
+ }
189
+ release = {
190
+ "version": "tfm-diagnostics-release-v1",
191
+ "subset": SUBSET,
192
+ "status": "completed_diagnostics_retrained_controls_pending",
193
+ "full_study_complete": False,
194
+ "scaling_deferred": True,
195
+ "checkpoint_sha256": cp_hash,
196
+ "normalization_sha256": norm_hash,
197
+ "original_manifest_sha256": run["manifest_sha256"],
198
+ "original_dataset_revision": "1baf2150adb20ff95ae8b81b7f4a985d81edbfe4",
199
+ "files": inventory,
200
+ }
201
+ (output / "release.json").write_text(json.dumps(release, indent=2, sort_keys=True) + "\n")
202
+ sums = {name: item["sha256"] for name, item in inventory.items()}
203
+ sums["release.json"] = sha256(output / "release.json")
204
+ (output / "SHA256SUMS").write_text("".join(f"{value} {name}\n" for name, value in sorted(sums.items())))
205
+ return {
206
+ "output": str(output),
207
+ "files": len(sums) + 1,
208
+ "bytes": sum(p.stat().st_size for p in output.rglob("*") if p.is_file()),
209
+ }
210
+
211
+
212
+ if __name__ == "__main__":
213
+ parser = argparse.ArgumentParser(description=__doc__)
214
+ parser.add_argument("--root", default=".")
215
+ parser.add_argument("--output", required=True)
216
+ parser.add_argument("--citylearn-license", required=True)
217
+ args = parser.parse_args()
218
+ print(json.dumps(build(args.root, args.output, args.citylearn_license)))
studies/robustness-seed42-20261011/scripts/challenge_interventions.py ADDED
@@ -0,0 +1,404 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Paired action effects on independently implemented synthetic thermal systems.
2
+
3
+ Reconstruct the hidden simulator snapshot ONLY for ground-truth branching. The
4
+ network and local-history baseline receive canonical observed window tensors,
5
+ never hidden parameters, seed, cohort identity, targets or internal state.
6
+ """
7
+
8
+ from __future__ import annotations
9
+
10
+ import argparse
11
+ import hashlib
12
+ import json
13
+ import os
14
+ import sys
15
+ from pathlib import Path
16
+
17
+ import numpy as np
18
+ import torch
19
+
20
+ sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
21
+
22
+ from scripts.thermal_challenges import ( # noqa: E402
23
+ COHORTS,
24
+ VERSION,
25
+ ThermalParameters,
26
+ integrate_interval,
27
+ load_challenge_windows,
28
+ sha256,
29
+ )
30
+ from tfm.data import FIELD_INDEX, ROLE_IDS, Normalization # noqa: E402
31
+ from tfm.evaluation.baselines import HistoryLinearBaseline, PersistenceBaseline # noqa: E402
32
+ from tfm.evaluation.counterfactual import action_effect_metrics # noqa: E402
33
+ from tfm.training.train import load_checkpoint # noqa: E402
34
+
35
+ INPUT_KEYS = (
36
+ "history",
37
+ "history_mask",
38
+ "history_dt",
39
+ "current",
40
+ "current_mask",
41
+ "future_actions",
42
+ "future_exogenous",
43
+ "future_dt",
44
+ "field_mask",
45
+ "field_ids",
46
+ "unit_ids",
47
+ "role_ids",
48
+ "query_fields",
49
+ "query_horizons",
50
+ )
51
+ STATE_FIELDS = ("indoor_temp", "mass_temp", "energy_kwh")
52
+ STATE_INDICES = [FIELD_INDEX[name] for name in STATE_FIELDS]
53
+ HORIZONS = (15, 30, 60, 120)
54
+ STEPS = (1, 2, 4, 8)
55
+ EXOGENOUS_INDICES = [FIELD_INDEX[name] for name in ("outdoor_temp", "solar_gain", "occupancy_gain")]
56
+ OBJECTIVE = "Sum of squared indoor_temp deviation from 21 degC at 15,30,60,120 minutes; units degC^2."
57
+ POLICY = (
58
+ "A is the recorded future HVAC command plan. B adds 6 electrical kW to all eight commands, "
59
+ "clipped to public [-15,15] kW. If that leaves the entire plan unchanged, B subtracts 6 kW instead. "
60
+ "Both branches start from the identical replayed snapshot with identical external disturbances. "
61
+ "Hidden actuator saturation can legitimately make different requested plans physically equivalent."
62
+ )
63
+
64
+
65
+ def initial_internal_state(episode):
66
+ """Reproduce exactly the independent generator's three initial-state draws."""
67
+ random = np.random.default_rng(episode.seed)
68
+ air = random.uniform(18, 24)
69
+ return np.array([air, air + random.uniform(-2, 2), air + random.uniform(-1, 1), 0.0, 0.0])
70
+
71
+
72
+ def internal_to_observed(state):
73
+ return np.asarray(state)[..., [0, 1, 4]]
74
+
75
+
76
+ def replay_snapshot(episode, anchor):
77
+ if episode.hidden_parameters.get("challenge_version") != VERSION:
78
+ raise ValueError("Independent generator version differs from episode")
79
+ parameters = ThermalParameters(**episode.hidden_parameters["parameters"])
80
+ state = initial_internal_state(episode)
81
+ for step in range(anchor):
82
+ state = integrate_interval(
83
+ parameters,
84
+ state,
85
+ episode.actions[step, FIELD_INDEX["hvac_power"]],
86
+ episode.exogenous[step, EXOGENOUS_INDICES],
87
+ step / 4,
88
+ )
89
+ if not np.allclose(
90
+ internal_to_observed(state), episode.values[anchor, STATE_INDICES], rtol=1e-10, atol=1e-10
91
+ ):
92
+ raise ValueError("Independent snapshot replay differs from stored observed trajectory")
93
+ return parameters, state
94
+
95
+
96
+ def paired_plans(episode, anchor):
97
+ original = episode.actions[anchor : anchor + 8, FIELD_INDEX["hvac_power"]].copy()
98
+ if original.shape != (8,):
99
+ raise ValueError("Action plans require eight future steps")
100
+ alternative = np.clip(original + 6.0, -15.0, 15.0)
101
+ if np.array_equal(original, alternative):
102
+ alternative = np.clip(original - 6.0, -15.0, 15.0)
103
+ if np.array_equal(original, alternative):
104
+ raise ValueError("Requested action intervention must change the observed plan")
105
+ return np.stack([original, alternative])
106
+
107
+
108
+ def branch_replay(parameters, snapshot, commands, disturbances, anchor):
109
+ """Ground-truth-only path: never passed to any predictor."""
110
+ trajectories = []
111
+ for candidate in commands:
112
+ state, trajectory = np.asarray(snapshot).copy(), []
113
+ for step, (command, exogenous) in enumerate(zip(candidate, disturbances)):
114
+ state = integrate_interval(parameters, state, command, exogenous, (anchor + step) / 4)
115
+ trajectory.append(internal_to_observed(state))
116
+ trajectories.append(trajectory)
117
+ return np.asarray(trajectories)
118
+
119
+
120
+ def observed_input_batch(sample, episode, commands, normalization):
121
+ """Whitelist canonical window tensors by stripping labels and replacing controls."""
122
+ batch = {
123
+ key: value.unsqueeze(0).repeat((len(commands),) + (1,) * value.ndim)
124
+ for key in INPUT_KEYS
125
+ for value in (sample[key],)
126
+ }
127
+ actions = np.zeros((len(commands), 8, len(normalization.mean)), dtype=np.float32)
128
+ active = (ROLE_IDS == 1) & episode.field_mask
129
+ field = FIELD_INDEX["hvac_power"]
130
+ if not active[field]:
131
+ raise ValueError("Thermal intervention requires an active public HVAC control slot")
132
+ actions[:, :, field] = (commands - normalization.mean[field]) / normalization.scale[field]
133
+ actions[:, :, ~active] = 0
134
+ batch["future_actions"] = torch.tensor(actions)
135
+ return batch
136
+
137
+
138
+ def predict_states(predictor, batch, normalization):
139
+ if not all(batch["current_mask"][:, STATE_INDICES].bool().flatten()):
140
+ raise ValueError("Action ranking on absolute temperature requires observed current thermal state")
141
+ with torch.inference_mode():
142
+ mean, _ = predictor(batch)
143
+ mean = mean.numpy()
144
+ result = np.zeros((len(mean), len(HORIZONS), len(STATE_FIELDS)))
145
+ for candidate in range(len(mean)):
146
+ for h, horizon in enumerate(HORIZONS):
147
+ for f, field in enumerate(STATE_INDICES):
148
+ selected = (batch["query_fields"][candidate].numpy() == field) & (
149
+ batch["query_horizons"][candidate].numpy() == horizon
150
+ )
151
+ if selected.sum() != 1:
152
+ raise ValueError("Missing/ambiguous public field and horizon query")
153
+ current = (
154
+ float(batch["current"][candidate, field]) * normalization.scale[field]
155
+ + normalization.mean[field]
156
+ )
157
+ result[candidate, h, f] = current + mean[candidate, selected][0] * normalization.scale[field]
158
+ return result
159
+
160
+
161
+ def objective(states):
162
+ return np.sum((np.asarray(states)[:, :, 0] - 21.0) ** 2, axis=1)
163
+
164
+
165
+ def aggregate(records):
166
+ result = []
167
+ for cohort in sorted({record["cohort"] for record in records}):
168
+ for model in sorted({record["model"] for record in records}):
169
+ selected = [
170
+ record for record in records if record["cohort"] == cohort and record["model"] == model
171
+ ]
172
+ if not selected:
173
+ continue
174
+ effects = []
175
+ for horizon in HORIZONS:
176
+ for field in STATE_FIELDS:
177
+ rows = [
178
+ row
179
+ for record in selected
180
+ for row in record["effects"]
181
+ if row["field"] == field and row["horizon_minutes"] == horizon
182
+ ]
183
+ effects.append(
184
+ {
185
+ "field": field,
186
+ "horizon_minutes": horizon,
187
+ "action_effect_mae": float(np.mean([row["action_effect_mae"] for row in rows])),
188
+ "action_effect_rmse": float(
189
+ np.sqrt(np.mean([row["action_effect_rmse"] ** 2 for row in rows]))
190
+ ),
191
+ "sample_count": len(rows),
192
+ }
193
+ )
194
+ pairs = sum(record["ranking_comparable_pairs"] for record in selected)
195
+ result.append(
196
+ {
197
+ "cohort": cohort,
198
+ "model": model,
199
+ "sample_count": len(selected),
200
+ "effects": effects,
201
+ "ranking_comparable_pairs": pairs,
202
+ "ranking_accuracy": sum(
203
+ (record["ranking_accuracy"] or 0) * record["ranking_comparable_pairs"]
204
+ for record in selected
205
+ )
206
+ / pairs
207
+ if pairs
208
+ else None,
209
+ "mean_regret_degC_squared": float(np.mean([record["regret"] for record in selected])),
210
+ }
211
+ )
212
+ return result
213
+
214
+
215
+ def run(args):
216
+ torch.set_num_threads(1)
217
+ output, manifest_path = Path(args.output), Path(args.manifest)
218
+ output.mkdir(parents=True, exist_ok=True)
219
+ manifest = json.loads(manifest_path.read_text())
220
+ generator_path = Path(__file__).with_name("thermal_challenges.py")
221
+ if manifest["generator_sha256"] != sha256(generator_path):
222
+ raise ValueError("Challenge generator source differs from dataset receipt")
223
+ if manifest["version"] != VERSION or manifest.get("external_real_world_validation"):
224
+ raise ValueError("Only the recorded independent synthetic challenge generator is supported")
225
+ model, checkpoint = load_checkpoint(args.checkpoint, "cpu")
226
+ model.eval()
227
+ normalization_path = Path(args.checkpoint).with_name("normalization.json")
228
+ norm = Normalization.load(normalization_path)
229
+ if checkpoint.get("normalization_sha256") != sha256(normalization_path):
230
+ raise ValueError("Checkpoint normalization fingerprint differs")
231
+ if set(norm.training_system_ids) != set(checkpoint["training_system_ids"]):
232
+ raise ValueError("Checkpoint and normalization training provenance differs")
233
+ models = {
234
+ "foundation_model": model,
235
+ "persistence": PersistenceBaseline(),
236
+ "history_linear_system_id": HistoryLinearBaseline(ROLE_IDS),
237
+ }
238
+ provenance = {
239
+ "checkpoint": str(Path(args.checkpoint).resolve()),
240
+ "checkpoint_sha256": sha256(args.checkpoint),
241
+ "normalization_sha256": sha256(normalization_path),
242
+ "manifest_sha256": sha256(manifest_path),
243
+ "generator_sha256": sha256(generator_path),
244
+ "args": vars(args),
245
+ "policy": POLICY,
246
+ "objective": OBJECTIVE,
247
+ "external_real_world_validation": False,
248
+ "source_sha256": {
249
+ str(path): sha256(path)
250
+ for path in [
251
+ Path(__file__),
252
+ generator_path,
253
+ Path("tfm/data/windows.py"),
254
+ Path("tfm/models/core.py"),
255
+ Path("tfm/evaluation/baselines.py"),
256
+ Path("tfm/evaluation/counterfactual.py"),
257
+ ]
258
+ },
259
+ "runtime": {
260
+ "torch": torch.__version__,
261
+ "numpy": np.__version__,
262
+ "device": "cpu",
263
+ "threads": 1,
264
+ "OMP_NUM_THREADS": os.environ.get("OMP_NUM_THREADS"),
265
+ "MKL_NUM_THREADS": os.environ.get("MKL_NUM_THREADS"),
266
+ },
267
+ "anchors": [],
268
+ "episode_sha256": {e["path"]: e["sha256"] for e in manifest["episodes"] if e["split"] == "test"},
269
+ }
270
+ records, truth_effects, maximum_error = [], [], 0.0
271
+ for cohort_index, cohort in enumerate(COHORTS):
272
+ dataset = load_challenge_windows(
273
+ manifest_path,
274
+ norm,
275
+ cohort=cohort,
276
+ history_steps=checkpoint["model_config"]["history_steps"],
277
+ samples=args.samples_per_cohort,
278
+ seed=args.seed + cohort_index,
279
+ )
280
+ for index, (episode_index, anchor) in enumerate(dataset.indices):
281
+ episode, sample = dataset.episodes[episode_index], dataset[index]
282
+ if set(checkpoint.get("training_system_ids", [])) & {str(episode.system_id)}:
283
+ raise ValueError("Challenge system occurred in checkpoint training")
284
+ parameters, snapshot = replay_snapshot(episode, anchor)
285
+ commands = paired_plans(episode, anchor)
286
+ disturbances = episode.exogenous[anchor : anchor + 8, EXOGENOUS_INDICES]
287
+ trajectories = branch_replay(parameters, snapshot, commands, disturbances, anchor)
288
+ replay_error = float(
289
+ np.max(np.abs(trajectories[0] - episode.values[anchor + 1 : anchor + 9, STATE_INDICES]))
290
+ )
291
+ maximum_error = max(maximum_error, replay_error)
292
+ if not np.allclose(
293
+ trajectories[0],
294
+ episode.values[anchor + 1 : anchor + 9, STATE_INDICES],
295
+ rtol=1e-10,
296
+ atol=1e-10,
297
+ ):
298
+ raise ValueError("Observed future plan replay differs from stored challenge trajectory")
299
+ truth = trajectories[:, np.asarray(STEPS) - 1]
300
+ batch = observed_input_batch(sample, episode, commands, norm)
301
+ anchor_id = {
302
+ "cohort": cohort,
303
+ "system_id": str(episode.system_id),
304
+ "anchor_step": int(anchor),
305
+ "sample_index": index,
306
+ "commands_kw": commands.tolist(),
307
+ "shared_exogenous_sha256": hashlib.sha256(disturbances.tobytes()).hexdigest(),
308
+ "snapshot_replay_verified": True,
309
+ "observed_plan_replay_max_error": replay_error,
310
+ }
311
+ provenance["anchors"].append(anchor_id)
312
+ truth_effects.append(
313
+ {
314
+ **anchor_id,
315
+ "true_effects": (truth[1] - truth[0]).tolist(),
316
+ "true_objective_degC_squared": objective(truth).tolist(),
317
+ }
318
+ )
319
+ for name, predictor in models.items():
320
+ predictions = predict_states(predictor, batch, norm)
321
+ record = action_effect_metrics(
322
+ predictions,
323
+ truth,
324
+ objective(predictions),
325
+ objective(truth),
326
+ field_names=STATE_FIELDS,
327
+ horizons=HORIZONS,
328
+ )
329
+ records.append({**anchor_id, "model": name, **record})
330
+ print(
331
+ json.dumps(
332
+ {"cohort": cohort, "snapshots": len(dataset), "observed_plan_max_replay_error": maximum_error}
333
+ ),
334
+ flush=True,
335
+ )
336
+ provenance["anchor_sha256"] = hashlib.sha256(
337
+ json.dumps(provenance["anchors"], sort_keys=True).encode()
338
+ ).hexdigest()
339
+ save(output / "provenance.json", provenance)
340
+ report = {
341
+ "status": "completed",
342
+ "external_real_world_validation": False,
343
+ "objective": OBJECTIVE,
344
+ "policy": POLICY,
345
+ "provenance_sha256": sha256(output / "provenance.json"),
346
+ "observed_plan_max_replay_error": maximum_error,
347
+ "summary": aggregate(records),
348
+ "true_effect_summary": [
349
+ {
350
+ "cohort": cohort,
351
+ "field": field,
352
+ "horizon_minutes": horizon,
353
+ "mean_absolute_true_effect": float(
354
+ np.mean(
355
+ [
356
+ abs(record["true_effects"][h][f])
357
+ for record in truth_effects
358
+ if record["cohort"] == cohort
359
+ ]
360
+ )
361
+ ),
362
+ "nonzero_effect_snapshots": sum(
363
+ abs(record["true_effects"][h][f]) > 1e-10
364
+ for record in truth_effects
365
+ if record["cohort"] == cohort
366
+ ),
367
+ "snapshot_count": sum(record["cohort"] == cohort for record in truth_effects),
368
+ }
369
+ for cohort in COHORTS
370
+ for f, field in enumerate(STATE_FIELDS)
371
+ for h, horizon in enumerate(HORIZONS)
372
+ ],
373
+ "truth_effects": truth_effects,
374
+ "records": records,
375
+ "limitation": "Independently implemented synthetic equations, not external real-world validation. "
376
+ "Only two candidate plans and 16 snapshots per cohort by default; saturation can remove the actual action effect. "
377
+ "Ranking/regret applies solely to the declared indoor-temperature objective; no energy/safety/controller claim.",
378
+ }
379
+ save(output / "interventions.json", report)
380
+ save(
381
+ output / "summary.json",
382
+ {key: value for key, value in report.items() if key not in ("records", "truth_effects")},
383
+ )
384
+ return report
385
+
386
+
387
+ def save(path, value):
388
+ temporary = path.with_suffix(path.suffix + ".tmp")
389
+ temporary.write_text(json.dumps(value, indent=2, allow_nan=False) + "\n")
390
+ temporary.replace(path)
391
+
392
+
393
+ def parser():
394
+ result = argparse.ArgumentParser(description=__doc__)
395
+ result.add_argument("--manifest", default="artifacts/robustness/thermal-challenges/manifest.json")
396
+ result.add_argument("--checkpoint", default="artifacts/full/transformer/seed-42/checkpoint.pt")
397
+ result.add_argument("--output", default="artifacts/robustness/challenge-interventions")
398
+ result.add_argument("--samples-per-cohort", type=int, default=16)
399
+ result.add_argument("--seed", type=int, default=20261012)
400
+ return result
401
+
402
+
403
+ if __name__ == "__main__":
404
+ run(parser().parse_args())
studies/robustness-seed42-20261011/scripts/citylearn_validation.py ADDED
@@ -0,0 +1,261 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """SOC-only third-party CityLearn device validation using unmodified Battery code.
2
+
3
+ Generation requires the isolated CityLearn runtime on PYTHONPATH. Evaluation
4
+ does not import CityLearn and uses the unchanged canonical WindowDataset.
5
+ """
6
+
7
+ import argparse
8
+ import hashlib
9
+ import importlib.metadata
10
+ import inspect
11
+ import json
12
+ import platform
13
+ from pathlib import Path
14
+
15
+ import numpy as np
16
+
17
+ from tfm.data.episodes import Episode, load_episode, save_episode
18
+ from tfm.data.schema import FIELD_INDEX, NUM_FIELDS
19
+ from tfm.data.windows import WindowDataset
20
+
21
+ ROOT = Path(__file__).resolve().parents[1]
22
+ VERSION = "citylearn-device-validation-1.0"
23
+ COHORTS = ("citylearn_battery_common_scale", "citylearn_battery_scale_shift")
24
+ SCHEMA = ["state_of_charge", "battery_power"]
25
+ SECONDS = 900
26
+ HOURS = SECONDS / 3600
27
+ DOCS = "https://www.citylearn.net/api/citylearn.energy_model.html"
28
+
29
+
30
+ def sha256(path):
31
+ return hashlib.sha256(Path(path).read_bytes()).hexdigest()
32
+
33
+
34
+ def load_citylearn():
35
+ """No monkey patches, dependency installation, or rewritten equations."""
36
+ if int(np.__version__.split(".")[0]) >= 2:
37
+ raise RuntimeError("Generation needs CityLearn-compatible NumPy<2; use the isolated runtime PYTHONPATH")
38
+ try:
39
+ from citylearn.base import EpisodeTracker
40
+ from citylearn.energy_model import Battery
41
+ except ModuleNotFoundError as error:
42
+ raise RuntimeError("Generation needs the isolated CityLearn runtime on PYTHONPATH") from error
43
+ version = importlib.metadata.version("citylearn")
44
+ if version != "3.0.2":
45
+ raise RuntimeError(f"Expected pinned CityLearn 3.0.2, got {version}")
46
+ return Battery, EpisodeTracker
47
+
48
+
49
+ def parameters(cohort, seed):
50
+ if cohort not in COHORTS:
51
+ raise ValueError(f"Unknown cohort {cohort}")
52
+ rng = np.random.default_rng(seed)
53
+ if cohort == "citylearn_battery_common_scale":
54
+ capacity = rng.uniform(7, 55)
55
+ nominal_power = rng.uniform(2, 14)
56
+ else:
57
+ capacity = rng.uniform(1, 4) if seed % 2 else rng.uniform(75, 160)
58
+ nominal_power = rng.uniform(0.4, 2) if seed % 2 else rng.uniform(21, 38)
59
+ depth = rng.uniform(0.7, 1)
60
+ return {"capacity": capacity, "nominal_power": nominal_power,
61
+ "depth_of_discharge": depth, "initial_soc": rng.uniform(max(1 - depth, 0.15), 0.85),
62
+ "efficiency": rng.uniform(0.85, 0.98), "loss_coefficient": rng.uniform(0.0001, 0.004),
63
+ "capacity_loss_coefficient": rng.uniform(1e-5, 1e-4),
64
+ "random_seed": int(seed), "seconds_per_time_step": SECONDS, "time_step_ratio": 1.0}
65
+
66
+
67
+ def requested_commands(steps, seed):
68
+ """Precommitted electrical kW schedule; never computed from future state."""
69
+ rng = np.random.default_rng(seed)
70
+ block = np.repeat(rng.choice([-20., -10., -3., 0., 0., 3., 10., 20.],
71
+ int(np.ceil(steps / 5))), 5)[:steps]
72
+ phase = rng.uniform(0, 2 * np.pi)
73
+ t = np.arange(steps)
74
+ return np.clip(block + 0.7 * np.sin(t * 0.08 + t**2 * 0.002 + phase), -20, 20)
75
+
76
+
77
+ def make_battery(p, steps):
78
+ Battery, EpisodeTracker = load_citylearn()
79
+ tracker = EpisodeTracker(0, steps)
80
+ tracker.next_episode(steps + 1, False, False, p["random_seed"])
81
+ battery = Battery(**p, episode_tracker=tracker)
82
+ battery.reset()
83
+ return battery
84
+
85
+
86
+ def advance_battery(battery, requested_kw):
87
+ """Record initial SOC before advancing, then run native loss and charging.
88
+
89
+ next_time_step() applies one physical step's standby losses and copies SOC;
90
+ charge() then applies energy for that same step. ratio=1 ensures kWh are
91
+ physical-step kWh. No fake temperature, voltage, or ambient observations.
92
+ """
93
+ if not np.isfinite(requested_kw) or not -20 <= requested_kw <= 20:
94
+ raise ValueError("Requested electrical power outside canonical bounds")
95
+ if battery.time_step_ratio != 1 or battery.seconds_per_time_step != SECONDS:
96
+ raise ValueError("Expected ratio=1 and 900-second physical steps")
97
+ battery.next_time_step()
98
+ battery.charge(float(requested_kw) * HOURS)
99
+ t = battery.time_step
100
+ return {"soc": float(battery.soc[t]), "requested_energy_kwh": float(requested_kw) * HOURS,
101
+ "applied_energy_kwh": float(battery.energy_balance[t]),
102
+ "efficiency": float(battery.efficiency), "degraded_capacity": float(battery.degraded_capacity)}
103
+
104
+
105
+ def generate_episode(cohort, system_seed, episode_seed, steps=192):
106
+ p = parameters(cohort, system_seed)
107
+ battery = make_battery(p, steps)
108
+ commands = requested_commands(steps, episode_seed)
109
+ soc = [float(battery.soc[0])]
110
+ audit = []
111
+ for command in commands:
112
+ native = advance_battery(battery, command)
113
+ soc.append(native["soc"])
114
+ audit.append(native)
115
+ soc = np.asarray(soc)
116
+ if not np.isfinite(soc).all() or np.any((soc < 0) | (soc > 1)):
117
+ raise ValueError("Invalid native CityLearn SOC")
118
+ values = np.zeros((steps + 1, NUM_FIELDS))
119
+ masks = np.zeros_like(values, dtype=bool)
120
+ actions = np.zeros((steps, NUM_FIELDS))
121
+ values[:, FIELD_INDEX["state_of_charge"]] = soc
122
+ masks[:, FIELD_INDEX["state_of_charge"]] = True
123
+ actions[:, FIELD_INDEX["battery_power"]] = commands
124
+ values[:-1, FIELD_INDEX["battery_power"]] = commands
125
+ masks[:-1, FIELD_INDEX["battery_power"]] = True
126
+ episode = Episode(
127
+ f"citylearn-device-{cohort}-{system_seed}", "battery", list(SCHEMA),
128
+ np.arange(steps + 1) * SECONDS / 60, values, masks, actions, np.zeros_like(actions),
129
+ {"source": "CityLearn.energy_model.Battery", "source_version": "3.0.2", "parameters": p,
130
+ "omitted_unavailable_observations": ["battery_temp", "terminal_voltage", "ambient_temp"],
131
+ "action_semantics": "requested electrical kW, converted to kWh per 900-second step before native Battery.charge; native power/SOC limits apply"},
132
+ int(episode_seed), "test",
133
+ )
134
+ audit_arrays = {key: np.array([row[key] for row in audit]) for key in audit[0]}
135
+ audit_arrays["requested_power_kw"] = commands
136
+ audit_arrays["applied_power_kw"] = audit_arrays["applied_energy_kwh"] / HOURS
137
+ return episode, audit_arrays
138
+
139
+
140
+ def load_citylearn_windows(manifest, normalization, cohort=None, history_steps=24, samples=None, seed=0):
141
+ """No CityLearn import needed to evaluate; only SOC targets are present."""
142
+ path = Path(manifest)
143
+ if path.is_dir():
144
+ path = path / "manifest.json"
145
+ data = json.loads(path.read_text())
146
+ if data.get("version") != VERSION or data.get("training_permitted") is not False:
147
+ raise ValueError("Not an immutable test-only CityLearn validation manifest")
148
+ entries = [e for e in data["episodes"] if cohort is None or e["cohort"] == cohort]
149
+ if not entries:
150
+ raise ValueError("No episodes match cohort")
151
+ if set(normalization.training_system_ids).intersection(e["system_id"] for e in entries):
152
+ raise ValueError("CityLearn test systems overlap normalization training provenance")
153
+ episodes = []
154
+ for entry in entries:
155
+ file = path.parent / entry["path"]
156
+ if sha256(file) != entry["sha256"]:
157
+ raise ValueError("CityLearn episode hash mismatch")
158
+ ep = load_episode(file)
159
+ if ep.schema != SCHEMA or ep.split != "test" or ep.system_id != entry["system_id"]:
160
+ raise ValueError("CityLearn SOC-only schema, split, or system ID mismatch")
161
+ episodes.append(ep)
162
+ return WindowDataset(episodes, normalization, history_steps=history_steps,
163
+ future_steps=8, samples=samples, seed=seed)
164
+
165
+
166
+ def generate_dataset(output, systems_per_cohort=24, steps=192, seed=183127):
167
+ if systems_per_cohort < 1 or steps < 32:
168
+ raise ValueError("Need positive systems and >=32 transitions")
169
+ Battery, _ = load_citylearn()
170
+ output = Path(output)
171
+ output.mkdir(parents=True, exist_ok=True)
172
+ package = Path(inspect.getfile(Battery)).parent
173
+ source_hashes = {str(p.relative_to(package)): sha256(p) for p in sorted(package.rglob("*.py"))}
174
+ frozen = sorted((ROOT / "tfm").rglob("*.py")) + sorted((ROOT / "configs").glob("*.json"))
175
+ frozen_hashes = {str(p.relative_to(ROOT)): sha256(p) for p in frozen}
176
+ entries = []
177
+ for ci, cohort in enumerate(COHORTS):
178
+ for i in range(systems_per_cohort):
179
+ system_seed = seed + 100000 * (ci + 1) + i
180
+ episode_seed = system_seed + 10000000
181
+ ep, audit = generate_episode(cohort, system_seed, episode_seed, steps)
182
+ relative = Path("episodes") / f"{ep.system_id}.npz"
183
+ native_relative = Path("native-audit") / f"{ep.system_id}.npz"
184
+ save_episode(ep, output / relative)
185
+ (output / "native-audit").mkdir(exist_ok=True)
186
+ np.savez_compressed(output / native_relative, **audit)
187
+ limited = np.abs(audit["applied_power_kw"] - audit["requested_power_kw"]) > 1e-5
188
+ entries.append({"system_id": ep.system_id, "family_id": "battery", "cohort": cohort,
189
+ "split": "test", "path": str(relative), "sha256": sha256(output / relative),
190
+ "native_audit_path": str(native_relative), "native_audit_sha256": sha256(output / native_relative),
191
+ "system_seed": system_seed, "episode_seed": episode_seed,
192
+ "transitions": steps, "window_anchors_h24_k8": steps - 30,
193
+ "native_limited_transitions": int(limited.sum()),
194
+ "parameters": ep.hidden_parameters["parameters"]})
195
+ if frozen_hashes != {str(p.relative_to(ROOT)): sha256(p) for p in frozen}:
196
+ raise RuntimeError("Frozen TFM source/config files changed during collection")
197
+ if source_hashes != {str(p.relative_to(package)): sha256(p) for p in sorted(package.rglob("*.py"))}:
198
+ raise RuntimeError("CityLearn sources changed during collection")
199
+ config = {"version": VERSION, "systems_per_cohort": systems_per_cohort, "steps": steps,
200
+ "seed": seed, "seconds_per_time_step": SECONDS, "time_step_ratio": 1.0}
201
+ (output / "config.json").write_text(json.dumps(config, indent=2) + "\n")
202
+ versions = {name: importlib.metadata.version(name) for name in
203
+ ("citylearn", "numpy", "pandas", "platformdirs", "requests", "simplejson", "pyyaml", "urllib3", "charset-normalizer", "pytz")}
204
+ manifest = {
205
+ "version": VERSION, "source": "CityLearn", "source_version": "3.0.2",
206
+ "source_docs_url": DOCS, "source_repository_url": "https://github.com/citylearn-project/CityLearn",
207
+ "kind": "third_party_synthetic_device_validation", "real_world_validation": False,
208
+ "building_or_control_environment_validation": False, "training_permitted": False,
209
+ "source_runtime_path": str(package.parent), "source_python_sha256": source_hashes,
210
+ "frozen_tfm_sources_sha256": frozen_hashes, "frozen_sources_unchanged_during_generation": True,
211
+ "generator_sha256": sha256(__file__), "config_file_sha256": sha256(output / "config.json"),
212
+ "versions": versions, "python": platform.python_version(), "config": config,
213
+ "schema": SCHEMA, "target_fields": ["state_of_charge"],
214
+ "unit_mapping": {"state_of_charge": "native Battery.soc fraction in [0,1]",
215
+ "battery_power": "precommitted requested electrical kW; charge receives kW * 0.25h kWh",
216
+ "timestamps": "minutes; native step=900 seconds"},
217
+ "transition_order": "Record SOC, native next_time_step (standby loss), native charge(requested_kWh), record next SOC",
218
+ "input_mapping_validation": "ratio=1 and seconds=900 enforced; zero/linear-energy/native-saturation tests; no realized future energy/power in input tensors",
219
+ "schema_omissions": ["battery_temp", "terminal_voltage", "ambient_temp"],
220
+ "evaluation_limits": "Only SOC transition targets; no battery temperature/voltage, no building dynamics, no HVAC-control environment, no real-world sensor data. Other CityLearn capabilities and optional dependencies were not exercised.",
221
+ "normalization_policy": "Frozen model training normalization only; no test/calibration fitting",
222
+ "episodes": entries,
223
+ }
224
+ (output / "manifest.json").write_text(json.dumps(manifest, indent=2, sort_keys=True) + "\n")
225
+ (output / "README.md").write_text(
226
+ "# CityLearn third-party battery device validation\n\n"
227
+ "These episodes execute the actual unmodified `citylearn.energy_model.Battery` from PyPI CityLearn 3.0.2. "
228
+ "This is third-party synthetic device physics validation, not real-world, building, or full CityLearn control-environment validation. "
229
+ "Only native SOC is available as a state. Temperature, voltage, and ambient temperature are omitted, not fabricated.\n\n"
230
+ "The command schedule is fixed before any transitions and is requested electrical kW within canonical +/-20kW. "
231
+ "Each command is converted to kWh using 900/3600 hours and passed to native Battery.charge with time_step_ratio=1. "
232
+ "Native SOC/power limits, efficiency curves, depth of discharge, degradation, and standby losses execute unchanged. "
233
+ "The native call order advances time (losses), then charges. Realized energy/power is stored separately in native-audit; "
234
+ "it never enters history/future action tensors. There is no reinterpretation of a CityLearn environment action fraction or thermostat setpoint.\n\n"
235
+ "Load with `scripts.citylearn_validation.load_citylearn_windows(manifest, saved_training_normalization, cohort=None)`. "
236
+ "It uses the unchanged WindowDataset and never imports CityLearn, refits normalization, or reads native-audit. "
237
+ "This tests SOC generalization while exposing only history SOC and planned electrical commands.\n\n"
238
+ "Generate: `PYTHONPATH=/private/tmp/tfm-citylearn-runtime OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 "
239
+ ".venv/bin/python -m scripts.citylearn_validation`. The temporary runtime contains CityLearn and only its device-path "
240
+ "import dependencies, including NumPy 1.26.4 to honor CityLearn's NumPy<2 declaration. No second torch, OpenStudio, "
241
+ "Gymnasium, DOE data, or torchvision was installed. The active .venv and lockfile are untouched. "
242
+ "Source hashes, installed versions, system seeds, parameter draws, canonical and audit hashes are in manifest.json.\n\n"
243
+ f"Official API documentation: {DOCS}\n"
244
+ )
245
+ return manifest
246
+
247
+
248
+ def main():
249
+ parser = argparse.ArgumentParser(description=__doc__)
250
+ parser.add_argument("--output", default="artifacts/robustness/citylearn")
251
+ parser.add_argument("--systems-per-cohort", type=int, default=24)
252
+ parser.add_argument("--steps", type=int, default=192)
253
+ parser.add_argument("--seed", type=int, default=183127)
254
+ args = parser.parse_args()
255
+ manifest = generate_dataset(args.output, args.systems_per_cohort, args.steps, args.seed)
256
+ print(json.dumps({"manifest": str(Path(args.output) / "manifest.json"),
257
+ "episodes": len(manifest["episodes"]), "cohorts": list(COHORTS)}))
258
+
259
+
260
+ if __name__ == "__main__":
261
+ main()
studies/robustness-seed42-20261011/scripts/context_uncertainty.py ADDED
@@ -0,0 +1,443 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Observation-only context flags and held-out residual intervals.
2
+
3
+ Inputs and intervals use the existing normalized state-delta convention. This
4
+ module does not train or change the model. It deliberately never reads targets,
5
+ simulator parameters, system IDs, or hidden room metadata while assessing context.
6
+ """
7
+
8
+ from __future__ import annotations
9
+
10
+ import hashlib
11
+ import json
12
+ from collections.abc import Mapping
13
+
14
+ import numpy as np
15
+
16
+ from tfm.evaluation.calibration import entry_key, finite_quantile
17
+
18
+ VERSION = "tfm-context-uncertainty-v1"
19
+ POLICY = {
20
+ "low_action_std": 0.05,
21
+ "weak_driver_std": 0.05,
22
+ "support_std_margin": 0.5,
23
+ "support_min_count": 10,
24
+ }
25
+ OBSERVATION_KEYS = (
26
+ "history",
27
+ "history_mask",
28
+ "current",
29
+ "current_mask",
30
+ "future_actions",
31
+ "future_exogenous",
32
+ "history_dt",
33
+ "future_dt",
34
+ "field_mask",
35
+ "role_ids",
36
+ )
37
+
38
+
39
+ def _array(value):
40
+ if hasattr(value, "detach"):
41
+ value = value.detach().cpu().numpy()
42
+ return np.asarray(value)
43
+
44
+
45
+ def _observations(batch):
46
+ # Explicit allowlist: do not even convert/inspect other mapping values.
47
+ data = {key: _array(batch[key]) for key in OBSERVATION_KEYS if key in batch}
48
+ required = {"history", "history_mask", "current", "current_mask", "field_mask", "role_ids"}
49
+ if not required <= data.keys():
50
+ raise ValueError(f"Missing observation arrays: {sorted(required - data.keys())}")
51
+ if data["current"].ndim == 1:
52
+ data = {key: value[None] for key, value in data.items()}
53
+ n, f = data["current"].shape
54
+ if data["history"].ndim != 3 or data["history"].shape[::2] != (n, f):
55
+ raise ValueError("history must have [sample, time, field] shape")
56
+ if data["history_mask"].shape != data["history"].shape:
57
+ raise ValueError("History values and masks must match")
58
+ for key in ("current_mask", "field_mask", "role_ids"):
59
+ if data[key].shape != (n, f):
60
+ raise ValueError(f"{key} must match current [sample, field] shape")
61
+ if np.any(~np.isin(data["role_ids"], [0, 1, 2])):
62
+ raise ValueError("Roles must use state=0, action=1, exogenous=2")
63
+ for key in ("future_actions", "future_exogenous"):
64
+ if key in data and (data[key].ndim != 3 or data[key].shape[::2] != (n, f)):
65
+ raise ValueError(f"{key} must have [sample, time, field] shape")
66
+ return data
67
+
68
+
69
+ def _row_features(data, i):
70
+ """Fieldwise features only from masked values and supplied time arrays."""
71
+ features = {}
72
+ active = data["field_mask"][i].astype(bool)
73
+ roles = data["role_ids"][i]
74
+ for field in np.flatnonzero(active):
75
+ role = int(roles[field])
76
+ h = data["history"][i, :, field]
77
+ observed = data["history_mask"][i, :, field].astype(bool) & np.isfinite(h)
78
+ if role != 0 and len(observed):
79
+ observed[-1] = False # Anchor controls describe a future transition.
80
+ values = h[observed]
81
+ prefix = f"{field}:"
82
+ if len(values):
83
+ features[prefix + "history_mean"] = float(values.mean())
84
+ features[prefix + "history_std"] = float(values.std())
85
+ if role == 0 and data["current_mask"][i, field] and np.isfinite(data["current"][i, field]):
86
+ features[prefix + "current"] = float(data["current"][i, field])
87
+ key = {1: "future_actions", 2: "future_exogenous"}.get(role)
88
+ if key in data:
89
+ future = data[key][i, :, field]
90
+ if len(future) and np.all(np.isfinite(future)):
91
+ features[prefix + "future_mean"] = float(future.mean())
92
+ features[prefix + "future_min"] = float(future.min())
93
+ features[prefix + "future_max"] = float(future.max())
94
+ for key in ("history_dt", "future_dt"):
95
+ if key in data:
96
+ times = data[key][i]
97
+ if len(times) and np.all(np.isfinite(times)):
98
+ features[key + ":span"] = float(np.ptp(times))
99
+ return features
100
+
101
+
102
+ def fit_context_support(
103
+ batches,
104
+ *,
105
+ training_system_ids,
106
+ forbidden_system_ids=(),
107
+ normalization_sha256,
108
+ transform_id="identity",
109
+ ):
110
+ """Fit fixed min/max feature envelopes from TRAIN windows only.
111
+
112
+ IDs declare input provenance; callers must supply actual training windows.
113
+ No test/calibration residual or test threshold tuning is used here.
114
+ """
115
+ ids = sorted(set(training_system_ids))
116
+ if not ids or set(ids) & set(forbidden_system_ids):
117
+ raise ValueError("Support systems must be training-only and disjoint from forbidden systems")
118
+ if not normalization_sha256 or not transform_id:
119
+ raise ValueError("Support requires normalization and transform provenance")
120
+ batches = [batches] if isinstance(batches, Mapping) else batches
121
+ collected = {}
122
+ for batch in batches:
123
+ data = _observations(batch)
124
+ for i in range(len(data["current"])):
125
+ for key, value in _row_features(data, i).items():
126
+ collected.setdefault(key, []).append(value)
127
+ entries = {}
128
+ for key, values in collected.items():
129
+ v = np.asarray(values)
130
+ if len(v) < POLICY["support_min_count"]:
131
+ continue
132
+ margin = max(float(v.std()) * POLICY["support_std_margin"], 1e-6)
133
+ entries[key] = {"min": float(v.min()) - margin, "max": float(v.max()) + margin, "count": len(v)}
134
+ artifact = {
135
+ "version": VERSION,
136
+ "training_system_ids": ids,
137
+ "normalization_sha256": normalization_sha256,
138
+ "transform_id": transform_id,
139
+ "policy": dict(POLICY),
140
+ "entries": entries,
141
+ "limitation": "Marginal feature-envelope heuristic; no guarantee of detecting hidden physical shift.",
142
+ }
143
+ artifact["sha256"] = hashlib.sha256(json.dumps(artifact, sort_keys=True).encode()).hexdigest()
144
+ return artifact
145
+
146
+
147
+ def assess_context(batch, support=None):
148
+ """Return per-window flags without rejecting useful sparse-context estimates.
149
+
150
+ Missing hidden room size cannot be detected from normal temperature alone.
151
+ Low variation/weak excitation are heuristics, not proofs of nonidentifiability.
152
+ A false shift flag likewise does not prove that a window is in distribution.
153
+ """
154
+ data = _observations(batch)
155
+ if support is not None and (support.get("version") != VERSION or support.get("policy") != POLICY):
156
+ raise ValueError("Unsupported support version or policy")
157
+ results = []
158
+ for i in range(len(data["current"])):
159
+ active = data["field_mask"][i].astype(bool)
160
+ roles = data["role_ids"][i]
161
+ states = active & (roles == 0)
162
+ h = data["history"][i]
163
+ hm = data["history_mask"][i].astype(bool) & np.isfinite(h)
164
+ current_ok = data["current_mask"][i].astype(bool) & np.isfinite(data["current"][i])
165
+ codes = []
166
+ hist_states = hm[:-1, states]
167
+ if not hist_states.any():
168
+ codes.append("missing_history")
169
+ driver_stds = {}
170
+ for role, name in ((1, "action"), (2, "exogenous")):
171
+ fields = np.flatnonzero(active & (roles == role))
172
+ observed = [h[:-1, field][hm[:-1, field]] for field in fields]
173
+ if len(fields) and not any(len(v) for v in observed):
174
+ codes.append(f"absent_historical_{name}")
175
+ if len(fields) and any(len(v) < 2 for v in observed):
176
+ codes.append(f"incomplete_historical_{name}")
177
+ stds = [float(v.std()) for v in observed if len(v) >= 2]
178
+ if stds:
179
+ driver_stds[role] = max(stds)
180
+ if role == 1 and stds and max(stds) < POLICY["low_action_std"]:
181
+ codes.append("low_observed_action_variation")
182
+ key = f"future_{'actions' if role == 1 else 'exogenous'}"
183
+ if len(fields) and (key not in data or not np.all(np.isfinite(data[key][i, :, fields]))):
184
+ codes.append(f"incomplete_planned_{name}")
185
+ expected_drivers = set(roles[active & (roles != 0)].tolist())
186
+ if (
187
+ expected_drivers
188
+ and expected_drivers <= driver_stds.keys()
189
+ and all(driver_stds[role] < POLICY["weak_driver_std"] for role in expected_drivers)
190
+ ):
191
+ codes.append("weak_excitation_heuristic")
192
+ if not states.any() or not np.all(current_ok[states]) or not np.all(hm[:, states]):
193
+ codes.append("incomplete_sensors")
194
+ eligible = bool(current_ok[states].any() or hm[:, states].any())
195
+ if not eligible:
196
+ codes.append("no_observed_state")
197
+ if "history_dt" not in data or "future_dt" not in data:
198
+ codes.append("time_context_unavailable")
199
+ elif any(
200
+ not np.all(np.isfinite(data[key][i])) or np.any(np.diff(data[key][i]) <= 0)
201
+ for key in ("history_dt", "future_dt")
202
+ ):
203
+ codes.append("invalid_time_context")
204
+ features = _row_features(data, i)
205
+ violations, checked = [], 0
206
+ if support is not None:
207
+ for key, value in features.items():
208
+ entry = support["entries"].get(key)
209
+ if entry:
210
+ checked += 1
211
+ if not entry["min"] <= value <= entry["max"]:
212
+ violations.append(key)
213
+ if violations:
214
+ codes.append("outside_training_support")
215
+ if not checked:
216
+ codes.append("training_support_unavailable")
217
+ # Fixed observation-only strata; never derive cutoffs from evaluation errors.
218
+ stratum = "+".join(codes) if codes else "observed_context"
219
+ results.append(
220
+ {
221
+ "version": VERSION,
222
+ "estimate_eligible": eligible,
223
+ "reasoning_codes": codes,
224
+ "stratum": stratum,
225
+ "uncertainty_status": "unsupported_state"
226
+ if not eligible
227
+ else "shift_no_coverage_claim"
228
+ if violations
229
+ else "limited_context"
230
+ if codes
231
+ else "observed_context",
232
+ "training_support_checked_features": checked,
233
+ "support_violations": violations,
234
+ "support_sha256": support["sha256"] if support is not None else None,
235
+ "hidden_physical_parameters_assessed": False,
236
+ }
237
+ )
238
+ return results
239
+
240
+
241
+ def _predictions(mean, std, target, fields, horizons, mask):
242
+ mean, std, fields, horizons, mask = (_array(x) for x in (mean, std, fields, horizons, mask))
243
+ mask = mask.astype(bool)
244
+ if mean.ndim != 2 or any(x.shape != mean.shape for x in (std, fields, horizons, mask)):
245
+ raise ValueError("Prediction arrays must match [sample, query] shapes")
246
+ if np.any(std[mask] <= 0) or np.any(~np.isfinite(std[mask])) or np.any(~np.isfinite(mean[mask])):
247
+ raise ValueError("Active predictions must be finite with positive std")
248
+ if np.any(~np.isfinite(horizons[mask])) or np.any(horizons[mask] <= 0):
249
+ raise ValueError("Active query horizons must be positive and finite")
250
+ if target is not None:
251
+ target = _array(target)
252
+ if target.shape != mean.shape or np.any(~np.isfinite(target[mask])):
253
+ raise ValueError("Active targets must be finite and match predictions")
254
+ return mean, std, target, fields, horizons, mask
255
+
256
+
257
+ def fit_context_calibration(
258
+ mean,
259
+ std,
260
+ target,
261
+ fields,
262
+ horizons,
263
+ mask,
264
+ assessments,
265
+ *,
266
+ calibration_system_ids,
267
+ training_system_ids,
268
+ validation_selection_system_ids=(),
269
+ test_system_ids=(),
270
+ checkpoint_sha256,
271
+ normalization_sha256,
272
+ transform_id,
273
+ coverage=0.9,
274
+ min_count=30,
275
+ ):
276
+ """Fit per-field/horizon/flag-stratum residual multipliers on separate systems.
277
+
278
+ The same masking transform MUST be applied before model prediction on both
279
+ calibration and evaluation windows, with its exact identity recorded here.
280
+ """
281
+ ids = sorted(set(calibration_system_ids))
282
+ prohibited = set(training_system_ids) | set(validation_selection_system_ids) | set(test_system_ids)
283
+ if not ids or set(ids) & prohibited:
284
+ raise ValueError("Calibration systems must be disjoint from training, selection and test")
285
+ if not checkpoint_sha256 or not normalization_sha256 or not transform_id:
286
+ raise ValueError("Calibration requires checkpoint, normalization and transform provenance")
287
+ if min_count < 1 or not 0 < coverage < 1:
288
+ raise ValueError("Calibration min_count must be positive and coverage in (0,1)")
289
+ mean, std, target, fields, horizons, mask = _predictions(mean, std, target, fields, horizons, mask)
290
+ if len(assessments) != len(mean):
291
+ raise ValueError("One context assessment is required per sample")
292
+ strata = np.asarray([row["stratum"] for row in assessments])
293
+ support_ids = {row.get("support_sha256") for row in assessments}
294
+ if len(support_ids) != 1 or any(row.get("version") != VERSION for row in assessments):
295
+ raise ValueError("Calibration assessments require one matching context support policy")
296
+ scores = np.abs(target - mean) / np.maximum(std, 1e-12)
297
+ entries, pooled = {}, {}
298
+
299
+ def entry(values):
300
+ multiplier = finite_quantile(values, coverage)
301
+ return {
302
+ "count": len(values),
303
+ "multiplier": float(multiplier) if len(values) >= min_count and np.isfinite(multiplier) else None,
304
+ }
305
+
306
+ for field, horizon in sorted(set(zip(fields[mask].tolist(), horizons[mask].tolist()))):
307
+ selected = mask & (fields == field) & (horizons == horizon)
308
+ key = entry_key(field, horizon)
309
+ pooled[key] = entry(scores[selected])
310
+ for stratum in sorted(set(strata[np.any(selected, axis=1)])):
311
+ entries[key + "|" + stratum] = entry(scores[selected & (strata[:, None] == stratum)])
312
+ return {
313
+ "version": VERSION,
314
+ "nominal_coverage": coverage,
315
+ "min_count": min_count,
316
+ "checkpoint_sha256": sorted(checkpoint_sha256),
317
+ "normalization_sha256": normalization_sha256,
318
+ "transform_id": transform_id,
319
+ "context_policy": dict(POLICY),
320
+ "calibration_system_ids": ids,
321
+ "support_sha256": next(iter(support_ids)),
322
+ "training_system_ids": sorted(set(training_system_ids)),
323
+ "validation_selection_system_ids": sorted(set(validation_selection_system_ids)),
324
+ "test_system_ids": sorted(set(test_system_ids)),
325
+ "entries": entries,
326
+ "pooled_entries": pooled,
327
+ "limitation": "Sampled windows are correlated. Finite-sample coverage requires exchangeability; "
328
+ "no system-level or shift guarantee. Pooled fallback has no stratum-conditional coverage claim.",
329
+ }
330
+
331
+
332
+ def apply_context_calibration(
333
+ mean,
334
+ std,
335
+ fields,
336
+ horizons,
337
+ mask,
338
+ assessments,
339
+ calibration,
340
+ *,
341
+ checkpoint_sha256,
342
+ normalization_sha256,
343
+ transform_id,
344
+ evaluation_system_ids=(),
345
+ ):
346
+ """Return normalized delta bounds and explicit per-query availability/status."""
347
+ if calibration.get("version") != VERSION or calibration.get("context_policy") != POLICY:
348
+ raise ValueError("Unsupported context calibration or policy")
349
+ for key, value in (
350
+ ("checkpoint_sha256", sorted(checkpoint_sha256)),
351
+ ("normalization_sha256", normalization_sha256),
352
+ ("transform_id", transform_id),
353
+ ):
354
+ if calibration.get(key) != value:
355
+ raise ValueError(f"Calibration {key} provenance mismatch")
356
+ ids = set(calibration["calibration_system_ids"])
357
+ prohibited = set(evaluation_system_ids)
358
+ for key in ("training_system_ids", "validation_selection_system_ids", "test_system_ids"):
359
+ prohibited.update(calibration[key])
360
+ if not ids or ids & prohibited:
361
+ raise ValueError("Calibration system leakage")
362
+ mean, std, _, fields, horizons, mask = _predictions(mean, std, None, fields, horizons, mask)
363
+ if len(assessments) != len(mean):
364
+ raise ValueError("One context assessment is required per sample")
365
+ if any(
366
+ row.get("version") != VERSION or row.get("support_sha256") != calibration["support_sha256"]
367
+ for row in assessments
368
+ ):
369
+ raise ValueError("Calibration assessment support provenance mismatch")
370
+ lower, upper = np.full(mean.shape, np.nan), np.full(mean.shape, np.nan)
371
+ status = np.full(mean.shape, "inactive", dtype=object)
372
+ counts = np.zeros(mean.shape, dtype=int)
373
+ for i, j in zip(*np.where(mask)):
374
+ row = assessments[i]
375
+ if not row["estimate_eligible"]:
376
+ status[i, j] = "no_observed_state"
377
+ continue
378
+ key = entry_key(fields[i, j], horizons[i, j])
379
+ entry = calibration["entries"].get(key + "|" + row["stratum"])
380
+ selected_status = "stratified"
381
+ if entry is None or entry["multiplier"] is None:
382
+ entry = calibration["pooled_entries"].get(key)
383
+ selected_status = "pooled_fallback_no_conditional_coverage"
384
+ if entry is None or entry["multiplier"] is None or entry["count"] < calibration["min_count"]:
385
+ status[i, j] = "insufficient_calibration_samples"
386
+ counts[i, j] = 0 if entry is None else entry["count"]
387
+ continue
388
+ multiplier = entry["multiplier"]
389
+ if not np.isfinite(multiplier) or multiplier < 0:
390
+ raise ValueError("Invalid calibration multiplier")
391
+ lower[i, j] = mean[i, j] - multiplier * std[i, j]
392
+ upper[i, j] = mean[i, j] + multiplier * std[i, j]
393
+ counts[i, j] = entry["count"]
394
+ if "outside_training_support" in row["reasoning_codes"]:
395
+ selected_status += "_shift_no_coverage_claim"
396
+ status[i, j] = selected_status
397
+ return {
398
+ "lower": lower,
399
+ "upper": upper,
400
+ "status": status,
401
+ "calibration_count": counts,
402
+ "available": mask & np.isfinite(lower) & np.isfinite(upper),
403
+ }
404
+
405
+
406
+ def context_interval_metrics(mean, target, mask, assessments, intervals):
407
+ """Coverage and width together; fixed-flag selective error, no test tuning."""
408
+ mean, target, mask = _array(mean), _array(target), _array(mask).astype(bool)
409
+ if target.shape != mean.shape or mask.shape != mean.shape or len(assessments) != len(mean):
410
+ raise ValueError("Metric arrays and assessments must align")
411
+ lower, upper = intervals["lower"], intervals["upper"]
412
+ available = mask & intervals["available"]
413
+ flags = sorted({flag for row in assessments for flag in row["reasoning_codes"]})
414
+
415
+ def summarize(selected):
416
+ valid = selected & available
417
+ count = int(selected.sum())
418
+ interval_count = int(valid.sum())
419
+ return {
420
+ "query_count": count,
421
+ "interval_count": interval_count,
422
+ "calibration_availability": interval_count / count if count else None,
423
+ "mae": float(np.abs(target - mean)[selected].mean()) if count else None,
424
+ "coverage": float(((target >= lower) & (target <= upper))[valid].mean())
425
+ if interval_count
426
+ else None,
427
+ "mean_width": float((upper - lower)[valid].mean()) if interval_count else None,
428
+ }
429
+
430
+ selective = {}
431
+ for flag in flags:
432
+ present = np.asarray([flag in row["reasoning_codes"] for row in assessments])[:, None]
433
+ selective[flag] = {"flagged": summarize(mask & present), "unflagged": summarize(mask & ~present)}
434
+ return {
435
+ "overall": summarize(mask),
436
+ "selective_error_by_flag": selective,
437
+ "status_counts": {
438
+ str(status): int(np.sum(intervals["status"][mask] == status))
439
+ for status in set(intervals["status"][mask].tolist())
440
+ },
441
+ "units": "normalized_delta",
442
+ "threshold_selection": "fixed_policy_not_test_tuned",
443
+ }
studies/robustness-seed42-20261011/scripts/evaluate_citylearn.py ADDED
@@ -0,0 +1,257 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Evaluate frozen seed-42 forecasts on actual third-party CityLearn SOC data.
2
+
3
+ Calibration is zero-shot with respect to CityLearn: only original reserved
4
+ battery calibration systems contribute residuals, with the exact SOC/power
5
+ schema used for CityLearn. Empirical shifted coverage is not a guarantee.
6
+ """
7
+
8
+ from __future__ import annotations
9
+
10
+ import argparse
11
+ import json
12
+ from collections import Counter
13
+ from pathlib import Path
14
+
15
+ import numpy as np
16
+ import torch
17
+ from torch.utils.data import DataLoader
18
+
19
+ from scripts.citylearn_validation import COHORTS, SCHEMA, load_citylearn_windows
20
+ from scripts.context_uncertainty import assess_context, fit_context_support
21
+ from tfm.data import Normalization, WindowDataset
22
+ from tfm.data.episodes import Episode
23
+ from tfm.data.schema import FIELD_INDEX, ROLE_IDS
24
+ from tfm.evaluation.baselines import HistoryLinearBaseline, PersistenceBaseline
25
+ from tfm.evaluation.benchmark import collect_predictions, sampling_provenance, sha256, summarize
26
+ from tfm.evaluation.calibration import entry_key, fit_calibration
27
+ from tfm.training.train import load_checkpoint, load_manifest_episodes
28
+
29
+ SOC = FIELD_INDEX["state_of_charge"]
30
+ POWER = FIELD_INDEX["battery_power"]
31
+ TRANSFORM = "citylearn-soc-and-requested-power-only-v1"
32
+
33
+
34
+ def save(path, value):
35
+ path = Path(path)
36
+ path.parent.mkdir(parents=True, exist_ok=True)
37
+ path.write_text(json.dumps(value, indent=2, allow_nan=False) + "\n")
38
+
39
+
40
+ def sparse_battery_episode(ep):
41
+ """Copy observations/plans into the exact external two-field schema.
42
+
43
+ No ambient, temperature or voltage survives in any canonical array. Source
44
+ SOC and pre-existing requested power are retained without reinterpretation.
45
+ Source split, IDs and targets remain their original reserved provenance.
46
+ """
47
+ if ep.family_id != "battery":
48
+ raise ValueError("Sparse SOC transform only accepts battery episodes")
49
+ values = np.zeros_like(ep.values)
50
+ mask = np.zeros_like(ep.observation_mask)
51
+ actions = np.zeros_like(ep.actions)
52
+ values[:, [SOC, POWER]] = ep.values[:, [SOC, POWER]]
53
+ mask[:, [SOC, POWER]] = ep.observation_mask[:, [SOC, POWER]]
54
+ actions[:, POWER] = ep.actions[:, POWER]
55
+ return Episode(ep.system_id, ep.family_id, list(SCHEMA), ep.timestamps.copy(), values, mask,
56
+ actions, np.zeros_like(ep.exogenous), dict(ep.hidden_parameters), ep.seed, ep.split)
57
+
58
+
59
+ def input_batches(dataset, batch_size):
60
+ for batch in DataLoader(dataset, batch_size=batch_size, shuffle=False):
61
+ yield {key: value for key, value in batch.items() if key not in ("target", "target_mask")}
62
+
63
+
64
+ def physical_soc(mean, arrays, norm):
65
+ anchor = arrays["current"][:, SOC] * norm.scale[SOC] + norm.mean[SOC]
66
+ return anchor[:, None] + mean * norm.scale[SOC]
67
+
68
+
69
+ def prediction_summary(mean, arrays, norm):
70
+ mask = arrays["target_mask"].astype(bool) & (arrays["query_fields"] == SOC)
71
+ prediction = physical_soc(mean, arrays, norm)
72
+ error = np.abs(mean - arrays["target"]) * norm.scale[SOC]
73
+ violation = (prediction < 0) | (prediction > 1)
74
+ rows = []
75
+ for horizon in sorted(set(arrays["query_horizons"][mask].tolist())):
76
+ selected = mask & (arrays["query_horizons"] == horizon)
77
+ rows.append({"horizon_minutes": float(horizon), "count": int(selected.sum()),
78
+ "soc_mae_fraction": float(error[selected].mean()),
79
+ "raw_soc_bounds_violation_rate": float(violation[selected].mean())})
80
+ return {"soc_mae_fraction": float(error[mask].mean()),
81
+ "soc_mae_percentage_points": float(error[mask].mean() * 100),
82
+ "raw_soc_bounds_violation_rate": float(violation[mask].mean()),
83
+ "by_horizon": rows, "clipping_applied": False}
84
+
85
+
86
+ def calibrated_half_width(log_std, arrays, calibration):
87
+ half = np.zeros_like(log_std)
88
+ for row in range(len(half)):
89
+ for query in np.flatnonzero(arrays["target_mask"][row]):
90
+ key = entry_key(arrays["query_fields"][row, query], arrays["query_horizons"][row, query])
91
+ entry = calibration["entries"].get(key)
92
+ if not entry or entry["multiplier"] is None:
93
+ raise ValueError(f"Sparse source calibration unavailable for active query {key}")
94
+ half[row, query] = entry["multiplier"] * np.exp(log_std[row, query])
95
+ return half
96
+
97
+
98
+ def evaluate(args):
99
+ torch.set_num_threads(1)
100
+ torch.set_num_interop_threads(1)
101
+ checkpoint_path, data_path, output = Path(args.checkpoint), Path(args.data), Path(args.output)
102
+ norm_path = checkpoint_path.with_name("normalization.json")
103
+ source_path = data_path / "manifest.json"
104
+ model, checkpoint = load_checkpoint(checkpoint_path, "cpu")
105
+ model.eval()
106
+ norm = Normalization.load(norm_path)
107
+ if sha256(norm_path) != checkpoint["normalization_sha256"]:
108
+ raise ValueError("Checkpoint normalization fingerprint mismatch")
109
+ if set(norm.training_system_ids) != set(checkpoint["training_system_ids"]):
110
+ raise ValueError("Normalization training provenance mismatch")
111
+ source_manifest, splits = load_manifest_episodes(data_path)
112
+ if sha256(source_path) != checkpoint["manifest_sha256"]:
113
+ raise ValueError("Original manifest fingerprint mismatch")
114
+ external_manifest = json.loads(Path(args.manifest).read_text())
115
+ external_ids = {e["system_id"] for e in external_manifest["episodes"]}
116
+ original_partitions = {key: set(checkpoint[key]) for key in
117
+ ("training_system_ids", "validation_selection_system_ids", "calibration_system_ids", "test_system_ids")}
118
+ if any(external_ids & ids for ids in original_partitions.values()):
119
+ raise ValueError("CityLearn system IDs overlap source data")
120
+ if any(a & b for i, a in enumerate(original_partitions.values())
121
+ for b in list(original_partitions.values())[i + 1:]):
122
+ raise ValueError("Original checkpoint partition overlap")
123
+ cal_episodes = [ep for ep in splits["validation"]
124
+ if ep.family_id == "battery" and ep.system_id in original_partitions["calibration_system_ids"]]
125
+ train_episodes = [ep for ep in splits["train"] if ep.family_id == "battery"]
126
+ if not cal_episodes or not train_episodes:
127
+ raise ValueError("Original battery training/reserved calibration unavailable")
128
+ cal_ids = {ep.system_id for ep in cal_episodes}
129
+ train_ids = {ep.system_id for ep in train_episodes}
130
+ used_source_ids = cal_ids | train_ids
131
+ used_entries = [entry for entry in source_manifest["episodes"] if entry["system_id"] in used_source_ids]
132
+ original_episode_hashes = {entry["path"]: sha256(data_path / entry["path"]) for entry in used_entries}
133
+ if any(checkpoint["episode_sha256"].get(key) != value for key, value in original_episode_hashes.items()):
134
+ raise ValueError("Original training/calibration battery episode hash changed")
135
+ history = checkpoint["model_config"]["history_steps"]
136
+ common = dict(history_steps=history, future_steps=8, samples=args.samples)
137
+ calibration_data = WindowDataset([sparse_battery_episode(ep) for ep in cal_episodes], norm,
138
+ seed=args.seed + 1, **common)
139
+ support_data = WindowDataset([sparse_battery_episode(ep) for ep in train_episodes], norm,
140
+ seed=args.seed + 2, **common)
141
+ binding_files = [checkpoint_path, norm_path, source_path, Path(args.manifest)]
142
+ binding_hashes = {str(p): sha256(p) for p in binding_files}
143
+ source_hashes = {path: sha256(path) for path in external_manifest["frozen_tfm_sources_sha256"]}
144
+ if source_hashes != external_manifest["frozen_tfm_sources_sha256"]:
145
+ raise ValueError("TFM/config source changed since CityLearn data collection")
146
+ generator_path = Path("scripts/citylearn_validation.py")
147
+ if sha256(generator_path) != external_manifest["generator_sha256"]:
148
+ raise ValueError("External generator no longer matches collected episodes")
149
+ report = {
150
+ "version": "tfm-citylearn-device-evaluation-v1", "status": "running",
151
+ "source_scope": "actual third-party CityLearn 3.0.2 synthetic battery device; SOC only",
152
+ "real_world_validation": False, "building_control_validation": False,
153
+ "weights_updated": False, "normalization_refit": False, "external_target_calibration": False,
154
+ "coverage_under_shift_validated": False,
155
+ "interval_status": "zero-shot wrt CityLearn; original reserved SOC/power-schema calibration; shifted coverage descriptive only",
156
+ "binding_file_sha256": binding_hashes, "model_state_sha256": checkpoint["model_state_sha256"],
157
+ "evaluation_sha256": sha256(__file__), "source_generator_sha256": external_manifest["generator_sha256"],
158
+ "source_citylearn_python_sha256": external_manifest["source_python_sha256"],
159
+ "original_used_episode_sha256": original_episode_hashes,
160
+ "citylearn_episode_sha256": {e["path"]: e["sha256"] for e in external_manifest["episodes"]},
161
+ "frozen_tfm_sources_sha256": source_hashes,
162
+ "helper_sha256": {path: sha256(path) for path in ("scripts/context_uncertainty.py", "tfm/evaluation/benchmark.py", "tfm/evaluation/calibration.py", "tfm/evaluation/baselines.py")},
163
+ "seed": args.seed, "checkpoint_seed": checkpoint["seed"], "samples_per_cohort": args.samples,
164
+ "cpu_threads": 1, "transform_id": TRANSFORM,
165
+ "target_fields": ["state_of_charge"], "input_schema": SCHEMA,
166
+ "calibration_system_ids": sorted(cal_ids), "support_training_system_ids": sorted(train_ids),
167
+ "calibration_anchor_provenance": sampling_provenance(calibration_data, source_path),
168
+ "support_anchor_provenance": sampling_provenance(support_data, source_path),
169
+ "limitations": [
170
+ "Only SOC is measured; no thermal/voltage/ambient state or building/environment validation.",
171
+ "Requested precommitted electrical kW is used as action, never realized future power.",
172
+ "Sparse calibration removes original battery temperature, voltage and ambient in all arrays; no test targets are used to fit intervals.",
173
+ "Windows are sampled with replacement and overlap; no independent system-level confidence bound or coverage guarantee.",
174
+ "Training feature-envelope flags are heuristic and cannot reliably reveal hidden scale/physics shift.",
175
+ "Predictions and intervals remain raw; no SOC clipping is applied.",
176
+ ],
177
+ "cohorts": {},
178
+ }
179
+ forbidden = set().union(*(ids for key, ids in original_partitions.items() if key != "training_system_ids")) | external_ids
180
+ support = fit_context_support(input_batches(support_data, args.batch_size), training_system_ids=train_ids,
181
+ forbidden_system_ids=forbidden, normalization_sha256=sha256(norm_path), transform_id=TRANSFORM)
182
+ save(output / "support.json", support)
183
+ candidates = {"transformer": model, "persistence": PersistenceBaseline(),
184
+ "history_linear": HistoryLinearBaseline(ROLE_IDS)}
185
+ calibrations = {}
186
+ for name, candidate in candidates.items():
187
+ mean, log_std, arrays = collect_predictions(candidate, calibration_data, "cpu", args.batch_size)
188
+ calibration = fit_calibration(
189
+ mean, np.exp(log_std), arrays["target"], arrays["query_fields"], arrays["query_horizons"], arrays["target_mask"],
190
+ calibration_system_ids=cal_ids, training_system_ids=checkpoint["training_system_ids"],
191
+ validation_selection_system_ids=checkpoint["validation_selection_system_ids"],
192
+ test_system_ids=original_partitions["test_system_ids"] | external_ids,
193
+ checkpoint_sha256=[sha256(checkpoint_path)], normalization_sha256=sha256(norm_path), coverage=0.9,
194
+ )
195
+ calibration["input_transform_id"] = TRANSFORM
196
+ calibration["model_name"] = name
197
+ calibrations[name] = calibration
198
+ save(output / f"{name}-sparse-calibration.json", calibration)
199
+ for cohort in COHORTS:
200
+ dataset = load_citylearn_windows(args.manifest, norm, cohort, history, args.samples, args.seed)
201
+ cohort_report = {"anchor_provenance": sampling_provenance(dataset, args.manifest), "models": {}}
202
+ for name, candidate in candidates.items():
203
+ mean, log_std, arrays = collect_predictions(candidate, dataset, "cpu", args.batch_size)
204
+ raw_metrics = summarize(mean, log_std, arrays, ["battery"] * len(dataset), norm)
205
+ cal_metrics = summarize(mean, log_std, arrays, ["battery"] * len(dataset), norm, calibrations[name])
206
+ entry = prediction_summary(mean, arrays, norm)
207
+ entry.update(uncalibrated_gaussian90=raw_metrics, zero_shot_sparse_calibrated90=cal_metrics)
208
+ half_width = calibrated_half_width(log_std, arrays, calibrations[name])
209
+ prediction = physical_soc(mean, arrays, norm)
210
+ np.savez_compressed(output / f"{cohort}-{name}-predictions.npz", mean_delta_normalized=mean,
211
+ log_std_normalized=log_std, target_delta_normalized=arrays["target"],
212
+ target_mask=arrays["target_mask"], query_fields=arrays["query_fields"],
213
+ horizons_minutes=arrays["query_horizons"], soc_prediction_raw=prediction,
214
+ soc_target=physical_soc(arrays["target"], arrays, norm),
215
+ calibrated90_lower_soc=prediction - half_width * norm.scale[SOC],
216
+ calibrated90_upper_soc=prediction + half_width * norm.scale[SOC])
217
+ if name == "transformer":
218
+ assessments = assess_context(arrays, support)
219
+ entry["context_flags"] = {
220
+ "flagged_windows": sum(bool(a["reasoning_codes"]) for a in assessments),
221
+ "reason_counts": dict(Counter(code for a in assessments for code in a["reasoning_codes"])),
222
+ "interpretation": "training-derived marginal envelope and fixed context heuristics; shift coverage unvalidated",
223
+ }
224
+ save(output / f"{cohort}-context.json", {"assessments": assessments})
225
+ cohort_report["models"][name] = entry
226
+ report["cohorts"][cohort] = cohort_report
227
+ save(output / "citylearn-evaluation.json", report)
228
+ print(json.dumps({"cohort": cohort, "soc_mae_fraction": {
229
+ name: row["soc_mae_fraction"] for name, row in cohort_report["models"].items()}}), flush=True)
230
+ if binding_hashes != {str(p): sha256(p) for p in binding_files}:
231
+ raise RuntimeError("Frozen data/normalization/checkpoint binding changed during evaluation")
232
+ if source_hashes != {path: sha256(path) for path in source_hashes}:
233
+ raise RuntimeError("Frozen TFM/config source changed during evaluation")
234
+ report["status"] = "completed"
235
+ report["frozen_bindings_unchanged_after_evaluation"] = True
236
+ report["predictions_sha256"] = {p.name: sha256(p) for p in output.glob("*-predictions.npz")}
237
+ save(output / "citylearn-evaluation.json", report)
238
+ return report
239
+
240
+
241
+ def main():
242
+ parser = argparse.ArgumentParser(description=__doc__)
243
+ parser.add_argument("--data", default="artifacts/poc/data")
244
+ parser.add_argument("--checkpoint", default="artifacts/full/transformer/seed-42/checkpoint.pt")
245
+ parser.add_argument("--manifest", default="artifacts/robustness/citylearn/manifest.json")
246
+ parser.add_argument("--output", default="artifacts/robustness/citylearn/evaluation")
247
+ parser.add_argument("--samples", type=int, default=512)
248
+ parser.add_argument("--batch-size", type=int, default=64)
249
+ parser.add_argument("--seed", type=int, default=20261011)
250
+ args = parser.parse_args()
251
+ if min(args.samples, args.batch_size) <= 0:
252
+ parser.error("Positive sample/batch counts required")
253
+ evaluate(args)
254
+
255
+
256
+ if __name__ == "__main__":
257
+ main()
studies/robustness-seed42-20261011/scripts/evaluate_thermal_challenges.py ADDED
@@ -0,0 +1,298 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Evaluate independently implemented thermal holdouts without fitting on them.
2
+
3
+ Context support uses training observations; all residual calibration uses the
4
+ original checkpoint's reserved calibration systems. Shift coverage is empirical,
5
+ not a calibrated guarantee on challenge populations.
6
+ """
7
+
8
+ from __future__ import annotations
9
+
10
+ import argparse
11
+ import json
12
+ import sys
13
+ from collections import Counter
14
+ from pathlib import Path
15
+
16
+ import numpy as np
17
+ import torch
18
+ from torch.utils.data import DataLoader
19
+
20
+ sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
21
+
22
+ from scripts.context_uncertainty import ( # noqa: E402
23
+ apply_context_calibration,
24
+ assess_context,
25
+ context_interval_metrics,
26
+ fit_context_calibration,
27
+ fit_context_support,
28
+ )
29
+ from scripts.history_diagnostics import HistoryView # noqa: E402
30
+ from scripts.thermal_challenges import COHORTS, load_challenge_windows # noqa: E402
31
+ from tfm.data import Normalization, WindowDataset # noqa: E402
32
+ from tfm.data.schema import ROLE_IDS # noqa: E402
33
+ from tfm.evaluation.baselines import HistoryLinearBaseline, PersistenceBaseline # noqa: E402
34
+ from tfm.evaluation.benchmark import collect_predictions, sampling_provenance, sha256, summarize # noqa: E402
35
+ from tfm.training.train import load_checkpoint, load_manifest_episodes # noqa: E402
36
+
37
+
38
+ def save(path, value):
39
+ path = Path(path)
40
+ path.parent.mkdir(parents=True, exist_ok=True)
41
+ tmp = path.with_suffix(path.suffix + ".tmp")
42
+ tmp.write_text(json.dumps(value, indent=2, allow_nan=False) + "\n")
43
+ tmp.replace(path)
44
+
45
+
46
+ def observations(view, batch_size):
47
+ for batch in DataLoader(view, batch_size=batch_size, shuffle=False):
48
+ yield {k: v for k, v in batch.items() if k not in ("target", "target_mask")}
49
+
50
+
51
+ def air_summary(rows):
52
+ rows = [r for r in rows if r["field"] == "indoor_temp"]
53
+ return {
54
+ "mae_degC": float(np.mean([r["mae"] for r in rows])),
55
+ "by_horizon": [{"minutes": r["horizon_minutes"], "mae_degC": r["mae"]} for r in rows],
56
+ }
57
+
58
+
59
+ def evaluate(args):
60
+ torch.set_num_threads(1)
61
+ checkpoint_path = Path(args.checkpoint)
62
+ norm_path = checkpoint_path.with_name("normalization.json")
63
+ model, checkpoint = load_checkpoint(checkpoint_path, "cpu")
64
+ norm = Normalization.load(norm_path)
65
+ if sha256(norm_path) != checkpoint["normalization_sha256"]:
66
+ raise ValueError("Checkpoint normalization fingerprint mismatch")
67
+ if set(norm.training_system_ids) != set(checkpoint["training_system_ids"]):
68
+ raise ValueError("Normalization training provenance differs")
69
+ source_manifest, splits = load_manifest_episodes(args.data)
70
+ source_path = Path(args.data) / "manifest.json"
71
+ if sha256(source_path) != checkpoint["manifest_sha256"]:
72
+ raise ValueError("Original data manifest fingerprint mismatch")
73
+ episode_hashes = checkpoint["episode_sha256"]
74
+ for relative_path, fingerprint in episode_hashes.items():
75
+ if sha256(source_path.parent / relative_path) != fingerprint:
76
+ raise ValueError(f"Original episode bytes differ from checkpoint: {relative_path}")
77
+ cal_ids = set(checkpoint["calibration_system_ids"])
78
+ cal_episodes = [e for e in splits["validation"] if e.system_id in cal_ids]
79
+ if {e.system_id for e in cal_episodes} != cal_ids:
80
+ raise ValueError("Reserved calibration systems unavailable")
81
+ history = checkpoint["model_config"]["history_steps"]
82
+ common = dict(history_steps=history, future_steps=8, samples=args.samples)
83
+ train_data = WindowDataset(splits["train"], norm, seed=args.seed + 2, **common)
84
+ cal_data = WindowDataset(cal_episodes, norm, seed=args.seed + 1, **common)
85
+ test_data = WindowDataset(splits["test"], norm, seed=args.seed, **common)
86
+ fingerprints = [sha256(checkpoint_path)]
87
+ output = Path(args.output)
88
+ challenge_manifest = json.loads(Path(args.manifest).read_text())
89
+ challenge_ids = {e["system_id"] for e in challenge_manifest["episodes"]}
90
+ forbidden = (
91
+ set(checkpoint["validation_selection_system_ids"]) | cal_ids | set(checkpoint["test_system_ids"])
92
+ )
93
+ if challenge_ids & (forbidden | set(checkpoint["training_system_ids"])):
94
+ raise ValueError("Challenge system IDs overlap source partitions")
95
+ report = {
96
+ "version": "tfm-independent-thermal-evaluation-v1",
97
+ "status": "running",
98
+ "checkpoint_sha256": fingerprints,
99
+ "normalization_sha256": sha256(norm_path),
100
+ "source_manifest_sha256": sha256(source_path),
101
+ "source_data_bundle_sha256": checkpoint["data_bundle_sha256"],
102
+ "original_episode_hashes_verified": len(episode_hashes),
103
+ "air_temperature_scale_degC": float(norm.scale[0]),
104
+ "challenge_manifest_sha256": sha256(args.manifest),
105
+ "script_sha256": sha256(__file__),
106
+ "helper_sha256": {
107
+ str(p): sha256(p)
108
+ for p in [
109
+ Path("scripts/context_uncertainty.py"),
110
+ Path("scripts/history_diagnostics.py"),
111
+ Path("scripts/thermal_challenges.py"),
112
+ ]
113
+ },
114
+ "seed": args.seed,
115
+ "samples_per_cohort_condition": args.samples,
116
+ "calibration_ids": sorted(cal_ids),
117
+ "calibration_anchor_provenance": sampling_provenance(cal_data, source_path),
118
+ "support_anchor_provenance": sampling_provenance(train_data, source_path),
119
+ "source_generator_version": source_manifest["generator_version"],
120
+ "interval_status": "original_distribution_calibration_unvalidated_under_challenge_shift",
121
+ "external_validation": "independent_in_house_synthetic_equations_not_real_world_or_third_party",
122
+ "conditions": {},
123
+ "limitations": [
124
+ "Temperature-only condition scores only currently observed indoor temperature; full-field averages have a different population.",
125
+ "Context flags cannot discover every hidden capacity change from ordinary temperature readings.",
126
+ "Calibration uses correlated windows; challenge coverage is descriptive, not a guarantee.",
127
+ "No predictions, normalization, flag thresholds or intervals are fitted using challenge targets.",
128
+ "The fitted linear baseline requires complete active-state transition rows; under temperature-only masks it falls back to persistence.",
129
+ ],
130
+ }
131
+ models = {
132
+ "transformer": model,
133
+ "persistence": PersistenceBaseline(),
134
+ "history_linear": HistoryLinearBaseline(ROLE_IDS),
135
+ }
136
+ for condition in ("full_history", "current_only", "thermal_temperature_only"):
137
+ transform_id = "history-diagnostics-v1:" + condition
138
+ train_view, cal_view = HistoryView(train_data, condition), HistoryView(cal_data, condition)
139
+ support = fit_context_support(
140
+ observations(train_view, args.batch_size),
141
+ training_system_ids=checkpoint["training_system_ids"],
142
+ forbidden_system_ids=forbidden | challenge_ids,
143
+ normalization_sha256=sha256(norm_path),
144
+ transform_id=transform_id,
145
+ )
146
+ save(output / condition / "support.json", support)
147
+ cal_mean, cal_log, cal_arrays = collect_predictions(model, cal_view, "cpu", args.batch_size)
148
+ cal_assessment = assess_context(cal_arrays, support)
149
+ calibration = fit_context_calibration(
150
+ cal_mean,
151
+ np.exp(cal_log),
152
+ cal_arrays["target"],
153
+ cal_arrays["query_fields"],
154
+ cal_arrays["query_horizons"],
155
+ cal_arrays["target_mask"],
156
+ cal_assessment,
157
+ calibration_system_ids=cal_ids,
158
+ training_system_ids=checkpoint["training_system_ids"],
159
+ validation_selection_system_ids=checkpoint["validation_selection_system_ids"],
160
+ test_system_ids=set(checkpoint["test_system_ids"]) | challenge_ids,
161
+ checkpoint_sha256=fingerprints,
162
+ normalization_sha256=sha256(norm_path),
163
+ transform_id=transform_id,
164
+ )
165
+ save(output / condition / "calibration.json", calibration)
166
+ reference_view = HistoryView(test_data, condition)
167
+ ref_mean, ref_log, ref_arrays = collect_predictions(model, reference_view, "cpu", args.batch_size)
168
+ ref_assessment = assess_context(ref_arrays, support)
169
+ ref_intervals = apply_context_calibration(
170
+ ref_mean,
171
+ np.exp(ref_log),
172
+ ref_arrays["query_fields"],
173
+ ref_arrays["query_horizons"],
174
+ ref_arrays["target_mask"],
175
+ ref_assessment,
176
+ calibration,
177
+ checkpoint_sha256=fingerprints,
178
+ normalization_sha256=sha256(norm_path),
179
+ transform_id=transform_id,
180
+ evaluation_system_ids=checkpoint["test_system_ids"],
181
+ )
182
+ ref_air = ref_arrays["target_mask"].astype(bool) & (ref_arrays["query_fields"] == 0)
183
+ reference = {
184
+ "anchor_provenance": sampling_provenance(test_data, source_path),
185
+ "air_context_interval_metrics": context_interval_metrics(
186
+ ref_mean, ref_arrays["target"], ref_air, ref_assessment, ref_intervals
187
+ ),
188
+ "outside_training_support_rate": sum(
189
+ "outside_training_support" in a["reasoning_codes"] for a in ref_assessment
190
+ )
191
+ / len(ref_assessment),
192
+ "any_evidence_flag_rate": sum(bool(a["reasoning_codes"]) for a in ref_assessment)
193
+ / len(ref_assessment),
194
+ }
195
+ thermal_assessments = [
196
+ a
197
+ for a, (ep, _) in zip(ref_assessment, test_data.indices)
198
+ if test_data.episodes[ep].family_id == "thermal"
199
+ ]
200
+ reference["thermal_only"] = {
201
+ "windows": len(thermal_assessments),
202
+ "outside_training_support_rate": sum(
203
+ "outside_training_support" in a["reasoning_codes"] for a in thermal_assessments
204
+ )
205
+ / len(thermal_assessments),
206
+ "any_evidence_flag_rate": sum(bool(a["reasoning_codes"]) for a in thermal_assessments)
207
+ / len(thermal_assessments),
208
+ }
209
+ results = {}
210
+ for cohort in COHORTS:
211
+ data = load_challenge_windows(
212
+ args.manifest, norm, cohort, samples=args.samples, seed=args.seed, history_steps=history
213
+ )
214
+ view = HistoryView(data, condition)
215
+ families = ["thermal"] * len(view)
216
+ result = {"anchor_provenance": sampling_provenance(data, args.manifest), "models": {}}
217
+ for name, candidate in models.items():
218
+ mean, log_std, arrays = collect_predictions(candidate, view, "cpu", args.batch_size)
219
+ rows = summarize(mean, log_std, arrays, families, norm)
220
+ entry = {"air_temperature": air_summary(rows), "metrics": rows}
221
+ if name == "transformer":
222
+ assessment = assess_context(arrays, support)
223
+ intervals = apply_context_calibration(
224
+ mean,
225
+ np.exp(log_std),
226
+ arrays["query_fields"],
227
+ arrays["query_horizons"],
228
+ arrays["target_mask"],
229
+ assessment,
230
+ calibration,
231
+ checkpoint_sha256=fingerprints,
232
+ normalization_sha256=sha256(norm_path),
233
+ transform_id=transform_id,
234
+ evaluation_system_ids={e.system_id for e in data.episodes},
235
+ )
236
+ air_mask = arrays["target_mask"].astype(bool) & (arrays["query_fields"] == 0)
237
+ interval_metrics = context_interval_metrics(
238
+ mean, arrays["target"], air_mask, assessment, intervals
239
+ )
240
+ entry.update(
241
+ context_interval_metrics=interval_metrics,
242
+ reason_counts=dict(
243
+ Counter(code for a in assessment for code in a["reasoning_codes"])
244
+ ),
245
+ flagged_windows=sum(bool(a["reasoning_codes"]) for a in assessment),
246
+ interval_status=report["interval_status"],
247
+ )
248
+ np.savez_compressed(
249
+ output / condition / f"{cohort}-predictions.npz",
250
+ mean=mean,
251
+ log_std=log_std,
252
+ target=arrays["target"],
253
+ mask=arrays["target_mask"],
254
+ fields=arrays["query_fields"],
255
+ horizons=arrays["query_horizons"],
256
+ lower=intervals["lower"],
257
+ upper=intervals["upper"],
258
+ )
259
+ save(output / condition / f"{cohort}-context.json", {"assessments": assessment})
260
+ result["models"][name] = entry
261
+ results[cohort] = result
262
+ print(
263
+ json.dumps(
264
+ {
265
+ "condition": condition,
266
+ "cohort": cohort,
267
+ "air_mae_degC": {
268
+ name: row["air_temperature"]["mae_degC"] for name, row in result["models"].items()
269
+ },
270
+ }
271
+ ),
272
+ flush=True,
273
+ )
274
+ report["conditions"][condition] = results
275
+ report.setdefault("original_distribution_reference", {})[condition] = reference
276
+ save(output / "thermal-evaluation.json", report)
277
+ report["status"] = "completed"
278
+ save(output / "thermal-evaluation.json", report)
279
+ return report
280
+
281
+
282
+ def main():
283
+ p = argparse.ArgumentParser(description=__doc__)
284
+ p.add_argument("--data", default="artifacts/poc/data")
285
+ p.add_argument("--checkpoint", default="artifacts/full/transformer/seed-42/checkpoint.pt")
286
+ p.add_argument("--manifest", default="artifacts/robustness/thermal-challenges/manifest.json")
287
+ p.add_argument("--output", default="artifacts/robustness/thermal-evaluation")
288
+ p.add_argument("--samples", type=int, default=512)
289
+ p.add_argument("--batch-size", type=int, default=64)
290
+ p.add_argument("--seed", type=int, default=20261011)
291
+ args = p.parse_args()
292
+ if min(args.samples, args.batch_size) < 1:
293
+ p.error("Sample and batch counts must be positive")
294
+ evaluate(args)
295
+
296
+
297
+ if __name__ == "__main__":
298
+ main()
studies/robustness-seed42-20261011/scripts/history_diagnostics.py ADDED
@@ -0,0 +1,394 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Same-checkpoint history perturbations on identical held-out anchors.
2
+
3
+ Inference-time removal/shuffle/swap is distribution shift, not a fair comparison
4
+ with a retrained no-history model. Calibration uses separately reserved systems
5
+ under each exact perturbation; it cannot establish coverage on new dynamics.
6
+ """
7
+
8
+ from __future__ import annotations
9
+
10
+ import argparse
11
+ import json
12
+ import os
13
+ import sys
14
+ from pathlib import Path
15
+
16
+ import numpy as np
17
+ import torch
18
+ from torch.utils.data import Dataset
19
+
20
+ sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
21
+
22
+ from tfm.data import FIELD_INDEX, Normalization, WindowDataset # noqa: E402
23
+ from tfm.evaluation.benchmark import ( # noqa: E402
24
+ collect_predictions,
25
+ evaluate_counterfactual,
26
+ sampling_provenance,
27
+ sha256,
28
+ summarize,
29
+ )
30
+ from tfm.evaluation.calibration import fit_calibration # noqa: E402
31
+ from tfm.training.train import load_checkpoint, load_manifest_episodes # noqa: E402
32
+
33
+ CONDITIONS = (
34
+ "full_history",
35
+ "current_only",
36
+ "history_branch_disabled",
37
+ "past_state_only",
38
+ "truncate_4",
39
+ "truncate_8",
40
+ "temporal_shuffle",
41
+ "matched_family_swap",
42
+ "thermal_temperature_only",
43
+ )
44
+ LIMITATION = (
45
+ "Same trained weights; inference-time history perturbations are distribution shift, "
46
+ "not fair retrained no-history comparisons. Masked rows retain time/missingness tokens. "
47
+ "history_branch_disabled zeros the encoder history contribution using its existing use_history switch. "
48
+ "Shuffle permutes past observations and masks jointly, keeps time slots/current/future plan. "
49
+ "Swap uses a different held-out system of the same family, matched only on observed current state "
50
+ "and donor availability covering every observed recipient current field; it disrupts history/current consistency and is not an isolated "
51
+ "causal identification test. Calibration is perturbation-specific at sampled-window level, "
52
+ "overlapping windows are correlated, with no guarantee for novel dynamics."
53
+ )
54
+
55
+
56
+ def perturb(sample, condition, *, seed=0, donor=None, family=None):
57
+ """Only history values/masks change. Labels never determine an intervention."""
58
+ if condition not in CONDITIONS:
59
+ raise ValueError(f"Unknown condition: {condition}")
60
+ result = {key: value.clone() for key, value in sample.items()}
61
+ history, mask = result["history"], result["history_mask"]
62
+ if condition in ("current_only", "history_branch_disabled"):
63
+ history[:-1] = 0
64
+ mask[:-1] = 0
65
+ elif condition == "past_state_only":
66
+ selected = result["role_ids"] != 0
67
+ history[:-1, selected] = 0
68
+ mask[:-1, selected] = 0
69
+ elif condition.startswith("truncate_"):
70
+ length = int(condition.rsplit("_", 1)[1])
71
+ history[:-length] = 0
72
+ mask[:-length] = 0
73
+ elif condition == "temporal_shuffle":
74
+ order = np.random.default_rng(seed).permutation(len(history) - 1)
75
+ history[:-1] = history[:-1][order].clone()
76
+ mask[:-1] = mask[:-1][order].clone()
77
+ elif condition == "thermal_temperature_only" and family == "thermal":
78
+ selected = (result["role_ids"] == 0) & (result["field_ids"] != FIELD_INDEX["indoor_temp"])
79
+ history[:, selected] = 0
80
+ mask[:, selected] = 0
81
+ result["current"][selected] = 0
82
+ result["current_mask"][selected] = 0
83
+ result["target_mask"] &= result["current_mask"].gather(0, result["query_fields"]).bool()
84
+ elif condition == "matched_family_swap":
85
+ if donor is None:
86
+ raise ValueError("Matched swap requires an independently selected donor")
87
+ history[:-1] = donor["history"][:-1]
88
+ mask[:-1] = donor["history_mask"][:-1]
89
+ return result
90
+
91
+
92
+ def match_donors(dataset):
93
+ """Nearest normalized observed current, same family, donor observes recipient current fields, other system.
94
+
95
+ No targets, future observations, future plans or hidden simulator parameters
96
+ participate in matching. All candidates are within this one reserved split.
97
+ """
98
+ current, masks, families, systems = [], [], [], []
99
+ for index, (episode_index, _) in enumerate(dataset.indices):
100
+ sample = dataset[index]
101
+ episode = dataset.episodes[episode_index]
102
+ current.append(sample["current"].numpy())
103
+ masks.append(sample["current_mask"].numpy().astype(bool))
104
+ families.append(episode.family_id)
105
+ systems.append(episode.system_id)
106
+ current, masks = np.asarray(current), np.asarray(masks)
107
+ families, systems = np.asarray(families), np.asarray(systems)
108
+ donors, distances = [], []
109
+ for index in range(len(dataset)):
110
+ eligible = (families == families[index]) & (systems != systems[index])
111
+ eligible &= np.all(masks | ~masks[index], axis=1)
112
+ candidates = np.flatnonzero(eligible)
113
+ if not masks[index].any() or not len(candidates):
114
+ raise ValueError(f"No same-family/recipient-field-covering other-system donor at sample {index}")
115
+ distance = np.mean((current[candidates][:, masks[index]] - current[index, masks[index]]) ** 2, axis=1)
116
+ selected = int(np.argmin(distance))
117
+ donors.append(int(candidates[selected]))
118
+ distances.append(float(np.sqrt(distance[selected])))
119
+ return donors, distances
120
+
121
+
122
+ class HistoryView(Dataset):
123
+ def __init__(self, dataset, condition, *, seed=0, donors=None):
124
+ self.base, self.condition, self.seed, self.donors = dataset, condition, seed, donors
125
+ for key in ("episodes", "indices", "normalization", "history_steps", "future_steps"):
126
+ setattr(self, key, getattr(dataset, key))
127
+
128
+ def __len__(self):
129
+ return len(self.base)
130
+
131
+ def __getitem__(self, index):
132
+ donor = self.base[self.donors[index]] if self.condition == "matched_family_swap" else None
133
+ return perturb(
134
+ self.base[index],
135
+ self.condition,
136
+ seed=self.seed + index,
137
+ donor=donor,
138
+ family=self.episodes[self.indices[index][0]].family_id,
139
+ )
140
+
141
+
142
+ class ModelView:
143
+ """Use the existing encoder switch without changing saved weights/config."""
144
+
145
+ def __init__(self, model, disable_history=False):
146
+ self.model, self.disable_history = model, disable_history
147
+
148
+ def __call__(self, batch):
149
+ if "target" in batch or "target_mask" in batch:
150
+ raise ValueError("Targets must remain outside the network inputs")
151
+ original = self.model.encoder.use_history
152
+ try:
153
+ if self.disable_history:
154
+ self.model.encoder.use_history = False
155
+ return self.model(batch)
156
+ finally:
157
+ self.model.encoder.use_history = original
158
+
159
+
160
+ def anchor_records(dataset):
161
+ return [
162
+ {
163
+ "system_id": str(dataset.episodes[e].system_id),
164
+ "family": str(dataset.episodes[e].family_id),
165
+ "episode_index": int(e),
166
+ "anchor_step": int(t),
167
+ }
168
+ for e, t in dataset.indices
169
+ ]
170
+
171
+
172
+ def validate_partitions(splits, checkpoint, normalization):
173
+ ids = {name: {str(e.system_id) for e in episodes} for name, episodes in splits.items()}
174
+ recorded = {
175
+ name: set(checkpoint.get(name, []))
176
+ for name in (
177
+ "training_system_ids",
178
+ "validation_selection_system_ids",
179
+ "calibration_system_ids",
180
+ "test_system_ids",
181
+ )
182
+ }
183
+ if not recorded["calibration_system_ids"]:
184
+ raise ValueError("Checkpoint must record separately reserved calibration systems")
185
+ if ids["test"] != recorded["test_system_ids"]:
186
+ raise ValueError("Test systems differ from checkpoint provenance")
187
+ if ids["train"] != recorded["training_system_ids"]:
188
+ raise ValueError("Training systems differ from checkpoint provenance")
189
+ if set(normalization.training_system_ids) != ids["train"]:
190
+ raise ValueError("Normalization training systems differ from checkpoint")
191
+ if not recorded["calibration_system_ids"] <= ids["validation"]:
192
+ raise ValueError("Reserved calibration systems must belong to validation")
193
+ groups = list(recorded.values())
194
+ if any(left & right for i, left in enumerate(groups) for right in groups[i + 1 :]):
195
+ raise ValueError("Checkpoint partition overlap would leak targets")
196
+ return recorded
197
+
198
+
199
+ def run(args):
200
+ torch.set_num_threads(1)
201
+ checkpoint_path, data_path, output = Path(args.checkpoint), Path(args.data), Path(args.output)
202
+ output.mkdir(parents=True, exist_ok=True)
203
+ manifest, splits = load_manifest_episodes(data_path)
204
+ model, checkpoint = load_checkpoint(checkpoint_path, "cpu")
205
+ model.eval()
206
+ normalization_path = checkpoint_path.with_name("normalization.json")
207
+ norm = Normalization.load(normalization_path)
208
+ checkpoint_hash, normalization_hash = sha256(checkpoint_path), sha256(normalization_path)
209
+ manifest_path = data_path / "manifest.json"
210
+ manifest_hash = sha256(manifest_path)
211
+ if checkpoint.get("normalization_sha256") != normalization_hash:
212
+ raise ValueError("Checkpoint normalization fingerprint differs")
213
+ if checkpoint.get("manifest_sha256") != manifest_hash:
214
+ raise ValueError("Checkpoint manifest fingerprint differs")
215
+ partitions = validate_partitions(splits, checkpoint, norm)
216
+ history = checkpoint["model_config"]["history_steps"]
217
+ datasets = {
218
+ "test": WindowDataset(
219
+ splits["test"], norm, history_steps=history, samples=args.samples, seed=args.seed
220
+ ),
221
+ "calibration": WindowDataset(
222
+ [e for e in splits["validation"] if str(e.system_id) in partitions["calibration_system_ids"]],
223
+ norm,
224
+ history_steps=history,
225
+ samples=args.calibration_samples,
226
+ seed=args.seed + 1,
227
+ ),
228
+ }
229
+ provenance = {
230
+ "checkpoint": str(checkpoint_path.resolve()),
231
+ "checkpoint_sha256": checkpoint_hash,
232
+ "normalization": str(normalization_path.resolve()),
233
+ "normalization_sha256": normalization_hash,
234
+ "manifest": str(manifest_path.resolve()),
235
+ "manifest_sha256": manifest_hash,
236
+ "episode_sha256": {
237
+ entry["path"]: sha256(data_path / entry["path"]) for entry in manifest["episodes"]
238
+ },
239
+ "source_sha256": {
240
+ str(path): sha256(path)
241
+ for path in [
242
+ Path(__file__),
243
+ Path("tfm/data/windows.py"),
244
+ Path("tfm/data/schema.py"),
245
+ Path("tfm/models/core.py"),
246
+ Path("tfm/evaluation/benchmark.py"),
247
+ Path("tfm/evaluation/calibration.py"),
248
+ Path("tfm/evaluation/metrics.py"),
249
+ ]
250
+ },
251
+ "partitions": {name: sorted(ids) for name, ids in partitions.items()},
252
+ "runtime": {
253
+ "torch": torch.__version__,
254
+ "numpy": np.__version__,
255
+ "device": "cpu",
256
+ "threads": 1,
257
+ "OMP_NUM_THREADS": os.environ.get("OMP_NUM_THREADS"),
258
+ "MKL_NUM_THREADS": os.environ.get("MKL_NUM_THREADS"),
259
+ },
260
+ "args": vars(args),
261
+ "limitation": LIMITATION,
262
+ }
263
+ donor_maps = {}
264
+ for split, dataset in datasets.items():
265
+ provenance[split] = {
266
+ **sampling_provenance(dataset, manifest_path),
267
+ "anchors": anchor_records(dataset),
268
+ }
269
+ if "matched_family_swap" in args.conditions:
270
+ donors, distances = match_donors(dataset)
271
+ donor_maps[split] = donors
272
+ provenance[split]["swap"] = {
273
+ "donor_indices": donors,
274
+ "current_standardized_rms_distance": distances,
275
+ "matching_uses_targets": False,
276
+ }
277
+ save_json(output / "provenance.json", provenance)
278
+ report = {
279
+ "status": "running",
280
+ "limitation": LIMITATION,
281
+ "conditions": {},
282
+ "checkpoint_sha256": checkpoint_hash,
283
+ "provenance_sha256": sha256(output / "provenance.json"),
284
+ }
285
+ for condition in args.conditions:
286
+ predictions, views = {}, {}
287
+ wrapped = ModelView(model, condition == "history_branch_disabled")
288
+ for split, dataset in datasets.items():
289
+ view = HistoryView(dataset, condition, seed=args.seed, donors=donor_maps.get(split))
290
+ views[split] = view
291
+ mean, log_std, arrays = collect_predictions(wrapped, view, "cpu", batch_size=args.batch_size)
292
+ predictions[split] = (mean, log_std, arrays)
293
+ # Labels are saved solely for downstream scoring/calibration, never model inputs.
294
+ np.savez_compressed(
295
+ output / f"{condition}-{split}.npz",
296
+ mean=mean,
297
+ log_std=log_std,
298
+ target=arrays["target"],
299
+ target_mask=arrays["target_mask"],
300
+ query_fields=arrays["query_fields"],
301
+ query_horizons=arrays["query_horizons"],
302
+ current_mask=arrays["current_mask"],
303
+ history_mask=arrays["history_mask"],
304
+ family=np.asarray([datasets[split].episodes[e].family_id for e, _ in dataset.indices]),
305
+ system_id=np.asarray([datasets[split].episodes[e].system_id for e, _ in dataset.indices]),
306
+ )
307
+ cm, cl, ca = predictions["calibration"]
308
+ calibration = fit_calibration(
309
+ cm,
310
+ np.exp(cl),
311
+ ca["target"],
312
+ ca["query_fields"],
313
+ ca["query_horizons"],
314
+ ca["target_mask"],
315
+ **{name: sorted(ids) for name, ids in partitions.items()},
316
+ checkpoint_sha256=[checkpoint_hash],
317
+ normalization_version=1,
318
+ normalization_sha256=normalization_hash,
319
+ )
320
+ save_json(output / f"calibration-{condition}.json", calibration)
321
+ mean, log_std, arrays = predictions["test"]
322
+ families = [datasets["test"].episodes[e].family_id for e, _ in datasets["test"].indices]
323
+ metrics = summarize(mean, log_std, arrays, families, norm, calibration=calibration)
324
+ result = {
325
+ "metrics": metrics,
326
+ "field_horizon_balanced_standardized_mae": float(
327
+ np.mean([row["standardized_mae"] for row in metrics])
328
+ ),
329
+ "field_horizon_balanced_standardized_nll": float(
330
+ np.mean([row["standardized_gaussian_nll"] for row in metrics])
331
+ ),
332
+ "calibration_scope": "this perturbation on independently reserved calibration systems",
333
+ "score_population": "thermal indoor_temp only; other families unchanged"
334
+ if condition == "thermal_temperature_only"
335
+ else "identical observed queries in all history comparisons",
336
+ }
337
+ if args.counterfactual_samples and condition == "thermal_temperature_only":
338
+ result["counterfactual"] = {
339
+ "status": "unrun",
340
+ "reason": "Existing all-state counterfactual helper requires observed current for delta decoding; "
341
+ "temperature-only forecast metrics are scored separately.",
342
+ }
343
+ elif args.counterfactual_samples:
344
+ result["counterfactual"] = evaluate_counterfactual(
345
+ wrapped,
346
+ views["test"],
347
+ norm,
348
+ "cpu",
349
+ samples=args.counterfactual_samples,
350
+ )
351
+ report["conditions"][condition] = result
352
+ save_json(output / "history-diagnostics.json", report)
353
+ print(
354
+ json.dumps(
355
+ {
356
+ "condition": condition,
357
+ "standardized_mae": result["field_horizon_balanced_standardized_mae"],
358
+ "standardized_nll": result["field_horizon_balanced_standardized_nll"],
359
+ }
360
+ ),
361
+ flush=True,
362
+ )
363
+ report["status"] = "completed"
364
+ report["outputs_sha256"] = {
365
+ path.name: sha256(path)
366
+ for path in sorted(output.iterdir())
367
+ if path.suffix in (".json", ".npz") and path.name != "history-diagnostics.json"
368
+ }
369
+ save_json(output / "history-diagnostics.json", report)
370
+ return report
371
+
372
+
373
+ def save_json(path, value):
374
+ temporary = path.with_suffix(path.suffix + ".tmp")
375
+ temporary.write_text(json.dumps(value, indent=2, allow_nan=False) + "\n")
376
+ temporary.replace(path)
377
+
378
+
379
+ def parser():
380
+ result = argparse.ArgumentParser(description=__doc__)
381
+ result.add_argument("--data", default="artifacts/poc/data")
382
+ result.add_argument("--checkpoint", default="artifacts/full/transformer/seed-42/checkpoint.pt")
383
+ result.add_argument("--output", default="artifacts/robustness/history")
384
+ result.add_argument("--samples", type=int, default=1024)
385
+ result.add_argument("--calibration-samples", type=int, default=512)
386
+ result.add_argument("--seed", type=int, default=42)
387
+ result.add_argument("--batch-size", type=int, default=32)
388
+ result.add_argument("--counterfactual-samples", type=int, default=12)
389
+ result.add_argument("--conditions", nargs="+", choices=CONDITIONS, default=list(CONDITIONS))
390
+ return result
391
+
392
+
393
+ if __name__ == "__main__":
394
+ run(parser().parse_args())
studies/robustness-seed42-20261011/scripts/thermal_challenges.py ADDED
@@ -0,0 +1,322 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Independent synthetic thermal holdouts. Never used for model training.
2
+
3
+ Run with OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 .venv/bin/python -m
4
+ scripts.thermal_challenges. This implements its own equations and SciPy
5
+ integrator; it never imports or steps the original simulator.
6
+ """
7
+
8
+ import argparse
9
+ import hashlib
10
+ import json
11
+ from dataclasses import asdict, dataclass
12
+ from pathlib import Path
13
+
14
+ import numpy as np
15
+ import scipy
16
+ from scipy.integrate import solve_ivp
17
+
18
+ from tfm.data.episodes import Episode, load_episode, save_episode
19
+ from tfm.data.schema import FIELD_INDEX, FIELD_SPECS, NUM_FIELDS, SCHEMAS
20
+ from tfm.data.windows import WindowDataset
21
+
22
+ VERSION = "independent-thermal-1.0"
23
+ COHORTS = ("parameter_shift_2node", "small_room_3node", "auditorium_3node")
24
+ ROOT = Path(__file__).resolve().parents[1]
25
+
26
+
27
+ @dataclass(frozen=True)
28
+ class ThermalParameters:
29
+ cohort: str
30
+ air_capacity: float # kWh / K
31
+ wall_capacity: float
32
+ furniture_capacity: float
33
+ envelope_conductance: float # kW / K, outside -> air
34
+ air_wall_conductance: float
35
+ air_furniture_conductance: float
36
+ wall_outside_conductance: float
37
+ hvac_efficiency: float # thermal kW / electrical kW
38
+ max_heating_power: float # electrical kW
39
+ max_cooling_power: float
40
+ actuator_lag_hours: float
41
+ ventilation_conductance: float
42
+ ventilation_period_hours: float
43
+ ventilation_phase_hours: float
44
+ ventilation_duty: float
45
+ solar_wall_fraction: float
46
+ occupancy_furniture_fraction: float
47
+
48
+
49
+ def sample_parameters(cohort, seed):
50
+ """Disjoint seeds define systems; all draws precede episode integration."""
51
+ if cohort not in COHORTS:
52
+ raise ValueError(f"Unknown cohort: {cohort}")
53
+ r = np.random.default_rng(seed)
54
+ if cohort == "parameter_shift_2node":
55
+ # Same deterministic two-node equations, different integrator and bounds.
56
+ return ThermalParameters(
57
+ cohort, r.uniform(0.12, 0.3), r.uniform(30, 55), 1.0,
58
+ r.uniform(0.025, 0.07), r.uniform(0.08, 0.18), 0.0, 0.0,
59
+ r.uniform(0.85, 1.05), r.uniform(1.1, 2.5), r.uniform(0.8, 2.2),
60
+ 0.0, 0.0, 6.0, 0.0, 0.4, 0.0, 0.0,
61
+ )
62
+ if cohort == "small_room_3node":
63
+ limits = (0.025, 0.10, 0.3, 1.5, 0.15, 0.8, 0.012, 0.045,
64
+ 0.025, 0.10, 0.015, 0.08, 0.015, 0.06, 0.3, 1.5, 0.2, 1.3, 0.01, 0.08)
65
+ else:
66
+ limits = (6.0, 18.0, 65.0, 180.0, 25.0, 80.0, 0.5, 1.2,
67
+ 1.0, 4.0, 0.5, 2.5, 0.2, 0.8, 18.0, 45.0, 18.0, 40.0, 0.5, 2.0)
68
+ draws = [r.uniform(limits[i], limits[i + 1]) for i in range(0, len(limits), 2)]
69
+ return ThermalParameters(
70
+ cohort, *draws[:7], r.uniform(0.8, 1.1), *draws[7:9],
71
+ r.uniform(0.03, 0.22), draws[9], r.uniform(3, 9), r.uniform(0, 3),
72
+ r.uniform(0.2, 0.65), r.uniform(0.1, 0.6), r.uniform(0.1, 0.45),
73
+ )
74
+
75
+
76
+ def ventilation_is_open(p, time_hours):
77
+ return ((time_hours + p.ventilation_phase_hours) % p.ventilation_period_hours
78
+ < p.ventilation_period_hours * p.ventilation_duty)
79
+
80
+
81
+ def thermal_rhs(p, state, command_kw, exogenous, ventilation_open=False):
82
+ """State is air/wall/furniture degC, electrical actuator kW, meter kWh.
83
+
84
+ Time is hours. Positive command heats; negative command cools. Internal
85
+ exchange cancels exactly in total stored heat; abs(electrical power)
86
+ increments the meter. mass_temp is the wall node, never an aggregate.
87
+ """
88
+ air, wall, furniture, actuator, _ = state
89
+ outside, solar, occupancy = exogenous
90
+ target = float(np.clip(command_kw, -p.max_cooling_power, p.max_heating_power))
91
+ delivered = actuator if p.actuator_lag_hours else target
92
+ vent = p.ventilation_conductance if ventilation_open else 0.0
93
+ outside_air = (p.envelope_conductance + vent) * (outside - air)
94
+ air_wall = p.air_wall_conductance * (wall - air)
95
+ air_furniture = p.air_furniture_conductance * (furniture - air)
96
+ outside_wall = p.wall_outside_conductance * (outside - wall)
97
+ solar_wall = p.solar_wall_fraction * solar
98
+ furniture_gain = p.occupancy_furniture_fraction * occupancy
99
+ air_rate = (outside_air + air_wall + air_furniture + p.hvac_efficiency * delivered
100
+ + solar - solar_wall + occupancy - furniture_gain) / p.air_capacity
101
+ wall_rate = (-air_wall + outside_wall + solar_wall) / p.wall_capacity
102
+ furniture_rate = (-air_furniture + furniture_gain) / p.furniture_capacity
103
+ actuator_rate = (target - actuator) / p.actuator_lag_hours if p.actuator_lag_hours else 0.0
104
+ return np.array([air_rate, wall_rate, furniture_rate, actuator_rate, abs(delivered)])
105
+
106
+
107
+ def integrate_interval(p, state, command_kw, exogenous, start_hours, duration_hours=0.25):
108
+ """Split at every infiltration event; adaptive solve_ivp within each segment."""
109
+ if duration_hours <= 0 or not np.isfinite(duration_hours):
110
+ raise ValueError("duration_hours must be finite and positive")
111
+ end = start_hours + duration_hours
112
+ boundaries = [start_hours, end]
113
+ if p.ventilation_conductance:
114
+ period, phase = p.ventilation_period_hours, p.ventilation_phase_hours
115
+ for cycle in range(int(np.floor((start_hours + phase) / period)) - 1,
116
+ int(np.ceil((end + phase) / period)) + 1):
117
+ for offset in (0.0, period * p.ventilation_duty):
118
+ event = cycle * period + offset - phase
119
+ if start_hours < event < end:
120
+ boundaries.append(event)
121
+ state = np.asarray(state, dtype=float).copy()
122
+ if not p.actuator_lag_hours:
123
+ state[3] = np.clip(command_kw, -p.max_cooling_power, p.max_heating_power)
124
+ for a, b in zip(sorted(boundaries)[:-1], sorted(boundaries)[1:]):
125
+ vent_open = ventilation_is_open(p, (a + b) / 2)
126
+ solution = solve_ivp(
127
+ lambda _t, y: thermal_rhs(p, y, command_kw, exogenous, vent_open),
128
+ (a, b), state, method="DOP853", rtol=1e-9, atol=1e-11,
129
+ max_step=min(0.125, b - a),
130
+ )
131
+ if not solution.success:
132
+ raise RuntimeError(solution.message)
133
+ state = solution.y[:, -1]
134
+ return state
135
+
136
+
137
+ def challenge_regime(cohort, steps, seed):
138
+ """Independent pulse/chirp/occupancy/weather patterns, not training policies."""
139
+ r = np.random.default_rng(seed)
140
+ t = np.arange(steps) / 4
141
+ phase = r.uniform(0, 2 * np.pi)
142
+ outdoor = 9 + 10 * np.sin(2 * np.pi * t / 29 + phase) + 3 * np.sin(t / 2.1)
143
+ outdoor += np.where((t > 19) & (t < 25), -7, 0)
144
+ solar = np.maximum(0, np.sin(2 * np.pi * (t - 6) / 24)) * r.uniform(0.15, 0.6)
145
+ occupancy = ((t % 24 > 10) & (t % 24 < 13)).astype(float) * r.uniform(0.06, 0.25)
146
+ occupancy += ((t % 24 > 18) & (t % 24 < 21)).astype(float) * 0.15
147
+ if cohort == "auditorium_3node":
148
+ solar *= 6
149
+ occupancy *= 8
150
+ # Saturating excitation alternates with zero/near-zero commands, plus chirp.
151
+ block_commands = r.choice([-15.0, -6.0, 0.0, 0.0, 6.0, 15.0], int(np.ceil(steps / 7)))
152
+ actions = np.repeat(block_commands, 7)[:steps]
153
+ actions = np.clip(actions + 1.2 * np.sin(0.25 * t + 0.04 * t**2 + phase), -15, 15)
154
+ return actions, np.column_stack([outdoor, solar, occupancy])
155
+
156
+
157
+ def generate_episode(cohort, system_seed, episode_seed, steps=192, split="test"):
158
+ if split not in ("test", "challenge_calibration"):
159
+ raise ValueError("Challenges cannot be train or tune episodes")
160
+ p = sample_parameters(cohort, system_seed)
161
+ r = np.random.default_rng(episode_seed)
162
+ initial = r.uniform(18, 24)
163
+ state = np.array([initial, initial + r.uniform(-2, 2), initial + r.uniform(-1, 1), 0, 0])
164
+ commands, disturbances = challenge_regime(cohort, steps, episode_seed)
165
+ trajectory = [state.copy()]
166
+ for i, (command, exog) in enumerate(zip(commands, disturbances)):
167
+ state = integrate_interval(p, state, command, exog, i / 4)
168
+ trajectory.append(state.copy())
169
+ trajectory = np.asarray(trajectory)
170
+ values = np.zeros((steps + 1, NUM_FIELDS))
171
+ observed = np.zeros_like(values, dtype=bool)
172
+ actions = np.zeros((steps, NUM_FIELDS))
173
+ exogenous = np.zeros_like(actions)
174
+ schema = list(SCHEMAS["thermal"])
175
+ for name, column in (("indoor_temp", 0), ("mass_temp", 1), ("energy_kwh", 4)):
176
+ values[:, FIELD_INDEX[name]] = trajectory[:, column]
177
+ observed[:, FIELD_INDEX[name]] = True
178
+ actions[:, FIELD_INDEX["hvac_power"]] = commands
179
+ for name, column in (("outdoor_temp", 0), ("solar_gain", 1), ("occupancy_gain", 2)):
180
+ exogenous[:, FIELD_INDEX[name]] = disturbances[:, column]
181
+ for name in schema:
182
+ idx = FIELD_INDEX[name]
183
+ if FIELD_SPECS[idx].role != "state":
184
+ source = actions if FIELD_SPECS[idx].role == "action" else exogenous
185
+ values[:-1, idx] = source[:, idx]
186
+ observed[:-1, idx] = True
187
+ spec = FIELD_SPECS[idx]
188
+ if np.any((values[observed[:, idx], idx] < spec.minimum)
189
+ | (values[observed[:, idx], idx] > spec.maximum)):
190
+ raise ValueError(f"Generated observed field outside public bounds: {name}")
191
+ system_id = f"challenge-{cohort}-{system_seed}"
192
+ episode = Episode(
193
+ system_id, "thermal", schema, np.arange(steps + 1) * 15.0, values,
194
+ observed, actions, exogenous,
195
+ {"challenge_version": VERSION, "parameters": asdict(p), "system_seed": system_seed,
196
+ "unobserved_state": ["furniture_temp", "actuator_power"],
197
+ "unobserved_disturbance": ["ventilation_switch"],
198
+ "mass_temp_definition": "wall temperature degC",
199
+ "energy_definition": "integral of absolute actual electrical actuator power kW over hours"},
200
+ episode_seed, split,
201
+ )
202
+ return episode
203
+
204
+
205
+ def sha256(path):
206
+ return hashlib.sha256(Path(path).read_bytes()).hexdigest()
207
+
208
+
209
+ def load_challenge_windows(manifest, normalization, cohort=None, split="test",
210
+ history_steps=24, samples=None, seed=0):
211
+ """Use frozen training normalization; no fitting or calibration occurs here."""
212
+ if split not in ("test", "challenge_calibration"):
213
+ raise ValueError("Only challenge test/calibration splits are available")
214
+ path = Path(manifest)
215
+ if path.is_dir():
216
+ path = path / "manifest.json"
217
+ data = json.loads(path.read_text())
218
+ entries = [e for e in data["episodes"] if e["split"] == split
219
+ and (cohort is None or e["cohort"] == cohort)]
220
+ if not entries:
221
+ raise ValueError("No challenge episodes match the requested cohort and split")
222
+ training_ids = set(normalization.training_system_ids)
223
+ if training_ids.intersection(e["system_id"] for e in entries):
224
+ raise ValueError("Challenge systems overlap normalization training provenance")
225
+ episodes = []
226
+ for e in entries:
227
+ file = path.parent / e["path"]
228
+ if sha256(file) != e["sha256"]:
229
+ raise ValueError(f"Challenge episode hash mismatch: {file}")
230
+ episode = load_episode(file)
231
+ if episode.system_id != e["system_id"] or episode.split != split:
232
+ raise ValueError("Challenge episode identity/split does not match manifest")
233
+ episodes.append(episode)
234
+ return WindowDataset(episodes, normalization, history_steps=history_steps,
235
+ future_steps=8, samples=samples, seed=seed)
236
+
237
+
238
+ def generate_dataset(output, systems_per_cohort=24, calibration_per_cohort=0, steps=192, seed=741230):
239
+ if systems_per_cohort < 1 or calibration_per_cohort < 0 or steps < 32:
240
+ raise ValueError("Need positive systems, nonnegative calibration count, and >=32 steps")
241
+ output = Path(output)
242
+ output.mkdir(parents=True, exist_ok=True)
243
+ source_paths = sorted((ROOT / "tfm").rglob("*.py")) + sorted((ROOT / "configs").glob("*.json"))
244
+ source_hashes = {str(p.relative_to(ROOT)): sha256(p) for p in source_paths}
245
+ config = {"version": VERSION, "systems_per_cohort": systems_per_cohort,
246
+ "calibration_per_cohort": calibration_per_cohort, "steps": steps, "seed": seed,
247
+ "interval_minutes": 15, "cohorts": list(COHORTS)}
248
+ config_bytes = json.dumps(config, sort_keys=True, separators=(",", ":")).encode()
249
+ (output / "config.json").write_text(json.dumps(config, indent=2) + "\n")
250
+ entries = []
251
+ for ci, cohort in enumerate(COHORTS):
252
+ for i in range(systems_per_cohort + calibration_per_cohort):
253
+ split = "test" if i < systems_per_cohort else "challenge_calibration"
254
+ system_seed = seed + 100000 * (ci + 1) + i
255
+ episode_seed = system_seed + 10000000
256
+ ep = generate_episode(cohort, system_seed, episode_seed, steps, split)
257
+ rel = Path("episodes") / f"{ep.system_id}.npz"
258
+ save_episode(ep, output / rel)
259
+ entries.append({"system_id": ep.system_id, "family_id": "thermal", "cohort": cohort,
260
+ "split": split, "path": str(rel), "sha256": sha256(output / rel),
261
+ "system_seed": system_seed, "episode_seed": episode_seed,
262
+ "transitions": steps, "window_anchors_h24_k8": steps - 30,
263
+ "parameters": ep.hidden_parameters["parameters"]})
264
+ after = {str(p.relative_to(ROOT)): sha256(p) for p in source_paths}
265
+ if after != source_hashes:
266
+ raise RuntimeError("Frozen core/config sources changed during generation")
267
+ manifest = {
268
+ "version": VERSION, "kind": "independently_implemented_synthetic_holdout",
269
+ "external_real_world_validation": False, "training_permitted": False,
270
+ "generator_sha256": sha256(__file__), "config_sha256": hashlib.sha256(config_bytes).hexdigest(),
271
+ "frozen_sources_sha256": source_hashes, "frozen_sources_unchanged_during_generation": True,
272
+ "versions": {"numpy": np.__version__, "scipy": scipy.__version__},
273
+ "integrator": {"library": "scipy.integrate.solve_ivp", "method": "DOP853",
274
+ "rtol": 1e-9, "atol": 1e-11, "max_step_hours": 0.125},
275
+ "cohort_meaning": {
276
+ "parameter_shift_2node": "Same deterministic two-node equations; new integrator and parameter bounds; no original process noise",
277
+ "small_room_3node": "Changed three-node equations, wall/outside link, furniture gains, switched ventilation, actuator lag, small capacities and saturation",
278
+ "auditorium_3node": "Same changed three-node topology, very large effective capacities; heater/cooler limits >15kW and therefore unreachable by canonical commands",
279
+ },
280
+ "units": {"capacity": "kWh/K", "conductance": "kW/K", "power": "kW", "time": "hours in ODE; minutes in episode"},
281
+ "schema_mapping": {"indoor_temp": "air node degC", "mass_temp": "wall node degC",
282
+ "energy_kwh": "actual electrical actuator absolute power integrated over hours",
283
+ "hvac_power": "requested electrical power kW before hidden saturation/lag"},
284
+ "schema_omissions": ["furniture temperature", "actual actuator power", "ventilation switch"],
285
+ "split_policy": "Test-only unseen systems; optional challenge_calibration systems are disjoint and are not train/tune/test. Never refit model or normalization here.",
286
+ "config": config, "episodes": entries,
287
+ }
288
+ (output / "manifest.json").write_text(json.dumps(manifest, indent=2, sort_keys=True) + "\n")
289
+ (output / "README.md").write_text(
290
+ "# Independent synthetic thermal challenges\n\n"
291
+ "This is an independently implemented synthetic simulator, not third-party or real-world validation. "
292
+ "The two-node parameter-shift cohort changes bounds and integration, not the deterministic equations. "
293
+ "The two three-node cohorts change equations and contain hidden furniture, switched infiltration, and actuator lag. "
294
+ "Capacities are effective kWh/K; these are stress-test scales, not calibrated claims about actual buildings. "
295
+ "Auditorium limits exceed canonical +/-15 kW commands, so those limits never bind; small-room limits do bind.\n\n"
296
+ "No challenge systems enter training or normalization. Optional challenge_calibration systems are separate from test "
297
+ "and must not be passed off as untouched test data after calibration. Use load_challenge_windows with the saved "
298
+ "training Normalization object to obtain canonical WindowDataset tensors; hidden parameters never appear in its inputs. "
299
+ "mass_temp is the wall node and energy_kwh is actual electrical consumption including lag, not thermal stored energy.\n\n"
300
+ "Reproduce with: OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 .venv/bin/python -m scripts.thermal_challenges "
301
+ "--output artifacts/robustness/thermal-challenges. Config, seeds, episode hashes, generator hash, library versions, "
302
+ "and frozen source hashes are in manifest.json.\n"
303
+ )
304
+ return manifest
305
+
306
+
307
+ def main():
308
+ parser = argparse.ArgumentParser(description=__doc__)
309
+ parser.add_argument("--output", default="artifacts/robustness/thermal-challenges")
310
+ parser.add_argument("--systems-per-cohort", type=int, default=24)
311
+ parser.add_argument("--calibration-per-cohort", type=int, default=0)
312
+ parser.add_argument("--steps", type=int, default=192)
313
+ parser.add_argument("--seed", type=int, default=741230)
314
+ args = parser.parse_args()
315
+ manifest = generate_dataset(args.output, args.systems_per_cohort, args.calibration_per_cohort,
316
+ args.steps, args.seed)
317
+ print(json.dumps({"manifest": str(Path(args.output) / "manifest.json"),
318
+ "episodes": len(manifest["episodes"]), "cohorts": list(COHORTS)}))
319
+
320
+
321
+ if __name__ == "__main__":
322
+ main()
studies/robustness-seed42-20261011/tests/test_behavior_support.py ADDED
@@ -0,0 +1,183 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import copy
2
+
3
+ import numpy as np
4
+ import pytest
5
+ import torch
6
+
7
+ from scripts.behavior_support import (
8
+ INPUT_KEYS,
9
+ assess_behavior_support,
10
+ build_history_backtest,
11
+ fit_behavior_support,
12
+ retrospective_behavior_scores,
13
+ summarize_behavior,
14
+ )
15
+
16
+
17
+ def batch(n=40, response=1, constant=False):
18
+ t = np.arange(24)
19
+ action = np.ones(24) if constant else 0.5 + np.sin(t)
20
+ exogenous = np.zeros(24) if constant else 0.1 * np.cos(t)
21
+ state = np.r_[0, np.cumsum(action[:-1] + exogenous[:-1])] * response
22
+ history = np.tile(np.column_stack([state, action, exogenous]), (n, 1, 1))
23
+ mask = np.ones_like(history)
24
+ mask[:, -1, 1:] = 0
25
+ return {
26
+ "history": history,
27
+ "history_mask": mask,
28
+ "history_dt": np.tile(np.arange(-345, 1, 15), (n, 1)),
29
+ "current": history[:, -1].copy(),
30
+ "current_mask": mask[:, -1].copy(),
31
+ "field_mask": np.ones((n, 3), dtype=bool),
32
+ "field_ids": np.tile([0, 1, 2], (n, 1)),
33
+ "unit_ids": np.zeros((n, 3), dtype=int),
34
+ "role_ids": np.tile([0, 1, 2], (n, 1)),
35
+ "future_actions": object(),
36
+ "future_exogenous": object(),
37
+ "target": object(),
38
+ "room_size": object(),
39
+ "simulator_parameters": object(),
40
+ }
41
+
42
+
43
+ class PastControlModel:
44
+ def __init__(self):
45
+ self.called = 0
46
+
47
+ def __call__(self, inputs):
48
+ self.called += 1
49
+ assert set(inputs) == set(INPUT_KEYS)
50
+ assert "target" not in inputs and "room_size" not in inputs
51
+ delta = inputs["future_actions"][:, :, 1].sum(1) + inputs["future_exogenous"][:, :, 2].sum(1)
52
+ mean = torch.zeros_like(inputs["query_horizons"])
53
+ mean[:, 0] = delta
54
+ return mean, torch.zeros_like(mean)
55
+
56
+
57
+ def profile(result):
58
+ return fit_behavior_support(
59
+ result,
60
+ training_system_ids=["train"],
61
+ forbidden_system_ids=["test", "cal"],
62
+ checkpoint_sha256=["model"],
63
+ normalization_sha256="norm",
64
+ transform_id="full-three-fields",
65
+ source_sha256={"source.py": "hash"},
66
+ training_anchor_provenance={"seed": 1, "anchor_sha256": "anchors"},
67
+ )
68
+
69
+
70
+ def assess(result, artifact, **changed):
71
+ kwargs = {
72
+ "checkpoint_sha256": ["model"],
73
+ "normalization_sha256": "norm",
74
+ "transform_id": "full-three-fields",
75
+ "evaluation_system_ids": ["test"],
76
+ }
77
+ kwargs.update(changed)
78
+ return assess_behavior_support(result, artifact, **kwargs)
79
+
80
+
81
+ def test_backtest_alignment_and_only_known_past_plan():
82
+ original = batch(1)
83
+ result = build_history_backtest(original)
84
+ inputs = result["inputs"]
85
+ np.testing.assert_array_equal(inputs["history_dt"], np.arange(-345, 1, 15)[None])
86
+ np.testing.assert_array_equal(inputs["future_dt"], np.arange(15, 121, 15)[None])
87
+ assert not inputs["history_mask"][:, :8].any()
88
+ assert not inputs["history_mask"][:, -1, 1:].any()
89
+ assert not inputs["current_mask"][:, 1:].any()
90
+ np.testing.assert_array_equal(
91
+ inputs["future_actions"][:, :, 1], original["history"][:, 15:23, 1].astype(np.float32)
92
+ )
93
+ np.testing.assert_array_equal(
94
+ inputs["future_exogenous"][:, :, 2], original["history"][:, 15:23, 2].astype(np.float32)
95
+ )
96
+ np.testing.assert_allclose(
97
+ result["observed_delta"][:, 0], original["current"][:, 0] - original["history"][:, 15, 0]
98
+ )
99
+
100
+
101
+ def test_no_future_target_or_hidden_parameter_leak_and_behavior_shift_detection():
102
+ model = PastControlModel()
103
+ training = retrospective_behavior_scores(model, batch())
104
+ artifact = profile(training)
105
+ shifted = batch(response=2)
106
+ scores = retrospective_behavior_scores(model, shifted)
107
+ assessments = assess(scores, artifact)
108
+ assert all(row["behavior_flag"] for row in assessments)
109
+ assert all(
110
+ not row["future_targets_consumed"] and not row["hidden_physical_parameters_consumed"]
111
+ for row in assessments
112
+ )
113
+ # Future plans and targets never enter the backward-facing assay.
114
+ shifted["future_actions"] = np.ones((40, 8, 3)) * 1e9
115
+ shifted["target"] = np.ones((40, 20)) * -1e9
116
+ changed = retrospective_behavior_scores(model, shifted)
117
+ np.testing.assert_array_equal(changed["scores"], scores["scores"])
118
+ assert summarize_behavior(assessments)["behavior_flag_rate_available_windows"] == 1
119
+
120
+
121
+ def test_no_history_or_missing_past_plan_skips_inference():
122
+ model = PastControlModel()
123
+ current_only = batch()
124
+ current_only["history_mask"][:, :-1] = 0
125
+ result = retrospective_behavior_scores(model, current_only)
126
+ assert model.called == 0
127
+ assert set(result["availability_status"]) == {"unavailable"}
128
+ assert all("missing_past_state_history" in codes for codes in result["reasoning_codes"])
129
+ incomplete = batch()
130
+ incomplete["history_mask"][:, 15, 1] = 0
131
+ result = retrospective_behavior_scores(model, incomplete)
132
+ assert model.called == 0
133
+ assert all("missing_known_past_driver_plan" in codes for codes in result["reasoning_codes"])
134
+
135
+
136
+ def test_constant_driver_good_fit_still_flags_low_excitation():
137
+ model = PastControlModel()
138
+ training = retrospective_behavior_scores(model, batch(constant=True))
139
+ result = assess(training, profile(training))
140
+ assert all(not row["behavior_flag"] for row in result)
141
+ assert all("low_excitation_heuristic" in row["reasoning_codes"] for row in result)
142
+ assert all(row["status"] == "limited" for row in result)
143
+
144
+
145
+ def test_matching_schema_and_availability_required_without_pooled_threshold():
146
+ model = PastControlModel()
147
+ artifact = profile(retrospective_behavior_scores(model, batch()))
148
+ sparse = batch()
149
+ sparse["field_mask"][:, 2] = False
150
+ result = assess(retrospective_behavior_scores(model, sparse), artifact)
151
+ assert all(row["thresholds_available"] == 0 for row in result)
152
+ assert all("matching_behavior_threshold_unavailable" in row["reasoning_codes"] for row in result)
153
+ partial = batch()
154
+ partial["history_mask"][:, 3, 0] = 0
155
+ result = assess(retrospective_behavior_scores(model, partial), artifact)
156
+ assert all(row["thresholds_available"] == 0 for row in result)
157
+ small = retrospective_behavior_scores(model, batch(8))
158
+ small_profile = profile(small)
159
+ assert all(entry["threshold"] is None for entry in small_profile["entries"].values())
160
+
161
+
162
+ def test_training_only_thresholds_and_provenance_guard():
163
+ result = retrospective_behavior_scores(PastControlModel(), batch())
164
+ artifact = profile(result)
165
+ with pytest.raises(ValueError, match="provenance"):
166
+ assess(result, artifact, transform_id="different-schema")
167
+ with pytest.raises(ValueError, match="overlap"):
168
+ assess(result, artifact, evaluation_system_ids=["train"])
169
+ with pytest.raises(ValueError, match="training-only"):
170
+ fit_behavior_support(
171
+ result,
172
+ training_system_ids=["test"],
173
+ forbidden_system_ids=["test"],
174
+ checkpoint_sha256=["m"],
175
+ normalization_sha256="n",
176
+ transform_id="t",
177
+ source_sha256={"s": "h"},
178
+ training_anchor_provenance={"a": "h"},
179
+ )
180
+ changed = copy.deepcopy(result)
181
+ changed["policy"]["quantile"] = 0.9
182
+ with pytest.raises(ValueError, match="policy"):
183
+ assess(changed, artifact)
studies/robustness-seed42-20261011/tests/test_challenge_interventions.py ADDED
@@ -0,0 +1,120 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import numpy as np
2
+ import pytest
3
+ import torch
4
+
5
+ from scripts.challenge_interventions import (
6
+ EXOGENOUS_INDICES,
7
+ STATE_INDICES,
8
+ branch_replay,
9
+ observed_input_batch,
10
+ paired_plans,
11
+ replay_snapshot,
12
+ )
13
+ from scripts.thermal_challenges import COHORTS, generate_episode
14
+ from tfm.data import FIELD_INDEX, ROLE_IDS, Normalization, WindowDataset
15
+ from tfm.data.schema import NUM_FIELDS
16
+
17
+
18
+ @pytest.mark.parametrize("cohort", COHORTS)
19
+ def test_observed_plan_replays_independent_hidden_snapshot_exactly(cohort):
20
+ episode = generate_episode(cohort, 177, 277, steps=40)
21
+ parameters, snapshot = replay_snapshot(episode, 24)
22
+ commands = paired_plans(episode, 24)
23
+ disturbances = episode.exogenous[24:32, EXOGENOUS_INDICES]
24
+ truth = branch_replay(parameters, snapshot, commands, disturbances, 24)
25
+ np.testing.assert_allclose(truth[0], episode.values[25:33, STATE_INDICES], rtol=1e-10, atol=1e-10)
26
+ np.testing.assert_array_equal(truth, branch_replay(parameters, snapshot, commands, disturbances, 24))
27
+ np.testing.assert_array_equal(
28
+ truth, branch_replay(parameters, snapshot, commands[::-1], disturbances, 24)[::-1]
29
+ )
30
+ episode.values[26, STATE_INDICES[0]] += 1 # Future labels do not affect reconstructed current snapshot.
31
+ np.testing.assert_array_equal(snapshot, replay_snapshot(episode, 24)[1])
32
+ episode.values[24, STATE_INDICES[0]] += 1
33
+ with pytest.raises(ValueError, match="snapshot replay"):
34
+ replay_snapshot(episode, 24)
35
+
36
+
37
+ def test_changed_plan_same_exogenous_and_no_hidden_metadata_features():
38
+ episode = generate_episode("small_room_3node", 177, 277, steps=40)
39
+ norm = Normalization(np.ones(NUM_FIELDS), np.full(NUM_FIELDS, 2.0), ["original-training"])
40
+ dataset = WindowDataset([episode], norm, history_steps=24, samples=None)
41
+ sample = dataset[0]
42
+ commands = paired_plans(episode, dataset.indices[0][1])
43
+ batch = observed_input_batch(sample, episode, commands, norm)
44
+ assert set(batch) == set(sample) - {"target", "target_mask"}
45
+ for key in batch.keys() - {"future_actions"}:
46
+ torch.testing.assert_close(batch[key][0], batch[key][1])
47
+ assert not torch.equal(batch["future_actions"][0], batch["future_actions"][1])
48
+ active = (ROLE_IDS == 1) & episode.field_mask
49
+ assert torch.count_nonzero(batch["future_actions"][:, :, ~active]) == 0
50
+ episode.hidden_parameters = {"leak": 1e30}
51
+ episode.seed += 999999
52
+ episode.system_id = "changed-id"
53
+ source = {key: value.clone() for key, value in sample.items()}
54
+ source["target"] += 1e30
55
+ source["hidden_parameters"] = torch.tensor(1e30)
56
+ again = observed_input_batch(source, episode, commands, norm)
57
+ for key in batch:
58
+ torch.testing.assert_close(batch[key], again[key])
59
+
60
+
61
+ def test_public_clipping_still_changes_fully_saturated_requested_plan():
62
+ episode = generate_episode("parameter_shift_2node", 177, 277, steps=40)
63
+ field = FIELD_INDEX["hvac_power"]
64
+ episode.actions[24:32, field] = 15
65
+ commands = paired_plans(episode, 24)
66
+ np.testing.assert_array_equal(commands[0], np.full(8, 15.0))
67
+ np.testing.assert_array_equal(commands[1], np.full(8, 9.0))
68
+ assert commands.min() >= -15 and commands.max() <= 15
69
+
70
+
71
+ def test_intervention_report_end_to_end_is_bound_to_unseen_cohorts(tmp_path):
72
+ import json
73
+
74
+ from scripts.challenge_interventions import parser, run
75
+ from scripts.thermal_challenges import generate_dataset, sha256
76
+ from tfm.models import build_model
77
+
78
+ data, directory, output = tmp_path / "data", tmp_path / "model", tmp_path / "report"
79
+ generate_dataset(data, systems_per_cohort=1, steps=40)
80
+ directory.mkdir()
81
+ norm = Normalization(np.zeros(NUM_FIELDS), np.ones(NUM_FIELDS), ["original-train"])
82
+ norm.save(directory / "normalization.json")
83
+ model = build_model(
84
+ "transformer", {"history_steps": 24, "d_model": 16, "heads": 4, "layers": 1, "ff_dim": 32}
85
+ )
86
+ torch.save(
87
+ {
88
+ "model_kind": "transformer",
89
+ "model_config": model.config,
90
+ "model_state": model.state_dict(),
91
+ "training_system_ids": ["original-train"],
92
+ "normalization_sha256": sha256(directory / "normalization.json"),
93
+ },
94
+ directory / "checkpoint.pt",
95
+ )
96
+ args = parser().parse_args(
97
+ [
98
+ "--manifest",
99
+ str(data / "manifest.json"),
100
+ "--checkpoint",
101
+ str(directory / "checkpoint.pt"),
102
+ "--output",
103
+ str(output),
104
+ "--samples-per-cohort",
105
+ "1",
106
+ ]
107
+ )
108
+ report = run(args)
109
+ assert report["status"] == "completed"
110
+ assert report["external_real_world_validation"] is False
111
+ assert report["observed_plan_max_replay_error"] < 1e-10
112
+ assert len(report["records"]) == 9
113
+ assert len(report["summary"]) == 9
114
+ assert len(report["truth_effects"]) == 3
115
+ assert all(row["nonzero_effect_snapshots"] <= 1 for row in report["true_effect_summary"])
116
+ receipt = json.loads((output / "provenance.json").read_text())
117
+ assert receipt["checkpoint_sha256"] == sha256(directory / "checkpoint.pt")
118
+ assert receipt["generator_sha256"] == sha256("scripts/thermal_challenges.py")
119
+ assert len(receipt["anchors"]) == 3
120
+ assert not any("hidden" in key for key in receipt["anchors"][0])
studies/robustness-seed42-20261011/tests/test_citylearn_validation.py ADDED
@@ -0,0 +1,133 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Execute actual CityLearn Battery code and verify canonical input semantics."""
2
+
3
+ import importlib.util
4
+
5
+ import numpy as np
6
+ import pytest
7
+
8
+ from scripts.citylearn_validation import (
9
+ COHORTS,
10
+ HOURS,
11
+ SCHEMA,
12
+ advance_battery,
13
+ generate_dataset,
14
+ generate_episode,
15
+ load_citylearn_windows,
16
+ make_battery,
17
+ parameters,
18
+ requested_commands,
19
+ sha256,
20
+ )
21
+ from tfm.data.normalization import Normalization
22
+ from tfm.data.schema import FIELD_INDEX, NUM_FIELDS
23
+
24
+ NATIVE_AVAILABLE = importlib.util.find_spec("citylearn") is not None and int(np.__version__.split(".")[0]) < 2
25
+ native_only = pytest.mark.skipif(not NATIVE_AVAILABLE, reason="Use isolated CityLearn runtime PYTHONPATH")
26
+
27
+
28
+ def ideal_parameters():
29
+ p = parameters(COHORTS[0], 11)
30
+ p.update(capacity=10.0, nominal_power=3.0, initial_soc=0.5, efficiency=1.0,
31
+ loss_coefficient=0.0, capacity_loss_coefficient=0.0, depth_of_discharge=1.0,
32
+ power_efficiency_curve=[[0.0, 1.0], [1.0, 1.0]],
33
+ capacity_power_curve=[[0.0, 1.0], [1.0, 1.0]])
34
+ return p
35
+
36
+
37
+ @native_only
38
+ def test_actual_native_code_zero_input_and_energy_units():
39
+ p = ideal_parameters()
40
+ battery = make_battery(p, 3)
41
+ assert battery.__class__.__module__ == "citylearn.energy_model"
42
+ assert advance_battery(battery, 0)["soc"] == pytest.approx(0.5, abs=1e-7)
43
+ charged = advance_battery(battery, 1)
44
+ assert charged["requested_energy_kwh"] == HOURS
45
+ assert charged["applied_energy_kwh"] == pytest.approx(HOURS, abs=1e-6)
46
+ assert charged["soc"] == pytest.approx(0.5 + HOURS / 10, abs=1e-7)
47
+ discharged = advance_battery(battery, -1)
48
+ assert discharged["soc"] == pytest.approx(0.5, abs=1e-7)
49
+
50
+
51
+ @native_only
52
+ def test_native_saturation_preserves_requested_action():
53
+ p = ideal_parameters()
54
+ p["nominal_power"] = 0.5
55
+ battery = make_battery(p, 1)
56
+ result = advance_battery(battery, 20)
57
+ assert result["requested_energy_kwh"] == 5
58
+ assert result["applied_energy_kwh"] == pytest.approx(0.5 * HOURS, abs=1e-6)
59
+ assert result["soc"] == pytest.approx(0.5 + 0.5 * HOURS / 10, abs=1e-7)
60
+ with pytest.raises(ValueError, match="canonical bounds"):
61
+ advance_battery(battery, 21)
62
+
63
+
64
+ @native_only
65
+ def test_native_standby_loss_applies_exactly_once():
66
+ p = ideal_parameters()
67
+ p["loss_coefficient"] = 0.004
68
+ battery = make_battery(p, 1)
69
+ result = advance_battery(battery, 0)
70
+ assert result["soc"] == pytest.approx(0.5 * (1 - 0.004 * HOURS), abs=1e-7)
71
+ assert result["applied_energy_kwh"] == pytest.approx(0, abs=1e-7)
72
+
73
+
74
+ @native_only
75
+ @pytest.mark.parametrize("cohort", COHORTS)
76
+ def test_replay_limits_and_observation_availability(cohort):
77
+ a, audit_a = generate_episode(cohort, 111, 211, 40)
78
+ b, audit_b = generate_episode(cohort, 111, 211, 40)
79
+ assert a.schema == SCHEMA
80
+ assert np.array_equal(a.values, b.values)
81
+ assert np.array_equal(a.actions[:, FIELD_INDEX["battery_power"]], requested_commands(40, 211))
82
+ for key in audit_a:
83
+ np.testing.assert_array_equal(audit_a[key], audit_b[key])
84
+ assert np.any(np.abs(audit_a["requested_power_kw"] - audit_a["applied_power_kw"]) > 1e-3)
85
+ assert not np.any(a.observation_mask[:, [FIELD_INDEX[n] for n in
86
+ ("battery_temp", "terminal_voltage", "ambient_temp")]])
87
+ assert not np.any(a.exogenous)
88
+ soc = a.values[:, FIELD_INDEX["state_of_charge"]]
89
+ assert np.all((soc >= 0) & (soc <= 1))
90
+
91
+
92
+ @native_only
93
+ def test_adapter_soc_only_no_hidden_or_future_applied_power(tmp_path):
94
+ manifest = generate_dataset(tmp_path, systems_per_cohort=1, steps=40)
95
+ assert manifest["kind"] == "third_party_synthetic_device_validation"
96
+ assert not manifest["real_world_validation"]
97
+ norm = Normalization(np.zeros(NUM_FIELDS), np.ones(NUM_FIELDS), ["original-training-system"])
98
+ ds = load_citylearn_windows(tmp_path, norm)
99
+ sample = ds[0]
100
+ valid_fields = sample["query_fields"][sample["target_mask"]].numpy()
101
+ assert set(valid_fields) == {FIELD_INDEX["state_of_charge"]}
102
+ ep = ds.episodes[0]
103
+ _, anchor = ds.indices[0]
104
+ np.testing.assert_array_equal(sample["future_actions"][:, FIELD_INDEX["battery_power"]].numpy(),
105
+ ep.actions[anchor:anchor + 8, FIELD_INDEX["battery_power"]].astype(np.float32))
106
+ original = {key: value.numpy().copy() for key, value in sample.items()}
107
+ ep.hidden_parameters = {"future_applied_kw": 100000, "secret_capacity": 1e12}
108
+ ep.system_id = "renamed-system"
109
+ for key, value in ds[0].items():
110
+ np.testing.assert_array_equal(value.numpy(), original[key])
111
+ # Applied-power audit is never consumed by the canonical loader.
112
+ audit = tmp_path / manifest["episodes"][0]["native_audit_path"]
113
+ audit.write_bytes(b"unreadable-audit")
114
+ ds_again = load_citylearn_windows(tmp_path, norm)
115
+ for key, value in ds_again[0].items():
116
+ np.testing.assert_array_equal(value.numpy(), original[key])
117
+ with pytest.raises(ValueError, match="only training"):
118
+ Normalization.fit(ds_again.episodes)
119
+ contaminated = Normalization(np.zeros(NUM_FIELDS), np.ones(NUM_FIELDS), [manifest["episodes"][0]["system_id"]])
120
+ with pytest.raises(ValueError, match="overlap"):
121
+ load_citylearn_windows(tmp_path, contaminated)
122
+ file = tmp_path / manifest["episodes"][0]["path"]
123
+ assert sha256(file) == manifest["episodes"][0]["sha256"]
124
+ file.write_bytes(file.read_bytes() + b"tamper")
125
+ with pytest.raises(ValueError, match="hash mismatch"):
126
+ load_citylearn_windows(tmp_path, norm)
127
+
128
+
129
+ def test_precommitted_public_request_schedule():
130
+ for seed in (1, 17, 100):
131
+ commands = requested_commands(192, seed)
132
+ np.testing.assert_array_equal(commands, requested_commands(192, seed))
133
+ assert np.all(np.abs(commands) <= 20)
studies/robustness-seed42-20261011/tests/test_context_uncertainty.py ADDED
@@ -0,0 +1,185 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import copy
2
+
3
+ import numpy as np
4
+ import pytest
5
+
6
+ from scripts.context_uncertainty import (
7
+ apply_context_calibration,
8
+ assess_context,
9
+ context_interval_metrics,
10
+ fit_context_calibration,
11
+ fit_context_support,
12
+ )
13
+
14
+
15
+ def batch(n=40):
16
+ rng = np.random.default_rng(23)
17
+ history = rng.normal(size=(n, 6, 4))
18
+ mask = np.ones_like(history)
19
+ mask[:, -1, 2:] = 0
20
+ return {
21
+ "history": history,
22
+ "history_mask": mask,
23
+ "current": history[:, -1].copy(),
24
+ "current_mask": mask[:, -1].copy(),
25
+ "future_actions": rng.normal(size=(n, 8, 4)),
26
+ "future_exogenous": rng.normal(size=(n, 8, 4)),
27
+ "field_mask": np.ones((n, 4), dtype=bool),
28
+ "role_ids": np.tile([0, 0, 1, 2], (n, 1)),
29
+ "history_dt": np.tile(np.arange(-75, 1, 15), (n, 1)),
30
+ "future_dt": np.tile(np.arange(15, 121, 15), (n, 1)),
31
+ }
32
+
33
+
34
+ def support(b):
35
+ return fit_context_support(
36
+ b, training_system_ids=["train"], forbidden_system_ids=["cal", "test"], normalization_sha256="norm"
37
+ )
38
+
39
+
40
+ class ForbiddenMetadata:
41
+ def __array__(self, *args, **kwargs):
42
+ raise AssertionError("Hidden metadata/targets were read")
43
+
44
+
45
+ def test_no_hidden_target_leak():
46
+ b = batch()
47
+ artifact = support(b)
48
+ expected = assess_context(b, artifact)
49
+ for key in (
50
+ "target",
51
+ "target_mask",
52
+ "room_size",
53
+ "simulator_params",
54
+ "family_ids",
55
+ "actual_future_temp",
56
+ "system_ids",
57
+ ):
58
+ b[key] = ForbiddenMetadata()
59
+ assert assess_context(b, artifact) == expected
60
+ assert support(b) == artifact
61
+ assert not expected[0]["hidden_physical_parameters_assessed"]
62
+
63
+
64
+ def test_missing_history_still_allows_estimate_and_missing_drivers_are_explicit():
65
+ b = batch(1)
66
+ b["history_mask"][:, :-1] = 0
67
+ row = assess_context(b)[0]
68
+ assert row["estimate_eligible"]
69
+ assert {
70
+ "missing_history",
71
+ "absent_historical_action",
72
+ "absent_historical_exogenous",
73
+ "incomplete_sensors",
74
+ } <= set(row["reasoning_codes"])
75
+ b["history_mask"][:] = 0
76
+ b["current_mask"][:] = 0
77
+ assert not assess_context(b)[0]["estimate_eligible"]
78
+
79
+
80
+ def test_weak_excitation_is_heuristic_and_does_not_reject():
81
+ b = batch(1)
82
+ b["history"][:, :, 2:] = 0.3
83
+ row = assess_context(b)[0]
84
+ assert {"low_observed_action_variation", "weak_excitation_heuristic"} <= set(row["reasoning_codes"])
85
+ assert row["estimate_eligible"]
86
+
87
+
88
+ def test_support_train_only_and_detects_supplied_future_control_shift():
89
+ b = batch()
90
+ profile = support(b)
91
+ assert all("outside_training_support" not in row["reasoning_codes"] for row in assess_context(b, profile))
92
+ b["future_actions"][0, :, 2] = 500
93
+ row = assess_context(b, profile)[0]
94
+ assert "outside_training_support" in row["reasoning_codes"]
95
+ assert row["uncertainty_status"] == "shift_no_coverage_claim"
96
+ with pytest.raises(ValueError, match="training-only"):
97
+ fit_context_support(
98
+ b, training_system_ids=["test"], forbidden_system_ids=["test"], normalization_sha256="norm"
99
+ )
100
+
101
+
102
+ def calibration_inputs(n=40):
103
+ mean = np.zeros((n, 1))
104
+ std = np.ones((n, 1)) * 0.2
105
+ target = np.linspace(0, 0.5, n)[:, None]
106
+ fields = np.zeros((n, 1), dtype=int)
107
+ horizons = np.ones((n, 1)) * 15
108
+ mask = np.ones((n, 1), dtype=bool)
109
+ assessments = assess_context(batch(n))
110
+ kwargs = dict(
111
+ calibration_system_ids=["cal"],
112
+ training_system_ids=["train"],
113
+ validation_selection_system_ids=["val"],
114
+ test_system_ids=["test"],
115
+ checkpoint_sha256=["checkpoint"],
116
+ normalization_sha256="norm",
117
+ transform_id="mask-history-v1",
118
+ min_count=30,
119
+ )
120
+ return (mean, std, target, fields, horizons, mask, assessments), kwargs
121
+
122
+
123
+ def apply(inputs, artifact, **extra):
124
+ mean, std, _, fields, horizons, mask, assessments = inputs
125
+ kwargs = dict(
126
+ checkpoint_sha256=["checkpoint"],
127
+ normalization_sha256="norm",
128
+ transform_id="mask-history-v1",
129
+ evaluation_system_ids=["test"],
130
+ )
131
+ kwargs.update(extra)
132
+ return apply_context_calibration(mean, std, fields, horizons, mask, assessments, artifact, **kwargs)
133
+
134
+
135
+ def test_insufficient_sample_fallback_is_explicit_then_unavailable():
136
+ inputs, kwargs = calibration_inputs()
137
+ assessments = inputs[-1]
138
+ assessments[0]["stratum"] = "rare_missing_context"
139
+ artifact = fit_context_calibration(*inputs, **kwargs)
140
+ result = apply(inputs, artifact)
141
+ assert result["status"][0, 0] == "pooled_fallback_no_conditional_coverage"
142
+ assert result["calibration_count"][0, 0] == 40
143
+ assert result["status"][1, 0] == "stratified"
144
+ assert np.isfinite(result["lower"]).all()
145
+ small_inputs, small_kwargs = calibration_inputs(8)
146
+ small_kwargs["min_count"] = 1
147
+ small = apply(small_inputs, fit_context_calibration(*small_inputs, **small_kwargs))
148
+ assert set(small["status"].ravel()) == {"insufficient_calibration_samples"}
149
+ assert not small["available"].any()
150
+ assert np.isnan(small["lower"]).all()
151
+
152
+
153
+ def test_calibration_provenance_and_separate_systems():
154
+ inputs, kwargs = calibration_inputs()
155
+ artifact = fit_context_calibration(*inputs, **kwargs)
156
+ for changed in (
157
+ {"transform_id": "identity"},
158
+ {"normalization_sha256": "different"},
159
+ {"checkpoint_sha256": ["other"]},
160
+ ):
161
+ with pytest.raises(ValueError, match="provenance"):
162
+ apply(inputs, artifact, **changed)
163
+ with pytest.raises(ValueError, match="leakage"):
164
+ apply(inputs, artifact, evaluation_system_ids=["cal"])
165
+ kwargs["calibration_system_ids"] = ["test"]
166
+ with pytest.raises(ValueError, match="disjoint"):
167
+ fit_context_calibration(*inputs, **kwargs)
168
+ changed_inputs = copy.deepcopy(inputs)
169
+ changed_inputs[-1][0]["support_sha256"] = "different-training-features"
170
+ with pytest.raises(ValueError, match="support provenance"):
171
+ apply(changed_inputs, artifact)
172
+
173
+
174
+ def test_shift_intervals_never_claim_coverage_and_metrics_include_width():
175
+ inputs, kwargs = calibration_inputs()
176
+ artifact = fit_context_calibration(*inputs, **kwargs)
177
+ inputs[-1][0]["reasoning_codes"].append("outside_training_support")
178
+ inputs[-1][0]["stratum"] = "outside_training_support"
179
+ result = apply(inputs, artifact)
180
+ assert result["status"][0, 0].endswith("shift_no_coverage_claim")
181
+ metrics = context_interval_metrics(inputs[0], inputs[2], inputs[5], inputs[-1], result)
182
+ assert metrics["overall"]["calibration_availability"] == 1
183
+ assert 0 <= metrics["overall"]["coverage"] <= 1
184
+ assert metrics["overall"]["mean_width"] > 0
185
+ assert metrics["selective_error_by_flag"]["outside_training_support"]["flagged"]["query_count"] == 1
studies/robustness-seed42-20261011/tests/test_evaluate_citylearn.py ADDED
@@ -0,0 +1,55 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Exact sparse-schema calibration matching and raw SOC decode checks."""
2
+
3
+ import numpy as np
4
+
5
+ from scripts.evaluate_citylearn import POWER, SOC, prediction_summary, sparse_battery_episode
6
+ from tfm.data import Normalization, WindowDataset
7
+ from tfm.data.episodes import Episode
8
+ from tfm.data.schema import NUM_FIELDS, SCHEMAS
9
+
10
+
11
+ def test_sparse_original_calibration_matches_external_two_field_schema():
12
+ steps = 40
13
+ values = np.ones((steps + 1, NUM_FIELDS)) * 77
14
+ values[:, SOC] = 0.5
15
+ values[:, POWER] = 1.5
16
+ actions = np.ones((steps, NUM_FIELDS)) * 8
17
+ actions[:, POWER] = 1.5
18
+ masks = np.zeros_like(values, dtype=bool)
19
+ from tfm.data.schema import field_mask
20
+ masks[:, field_mask(SCHEMAS["battery"])] = True
21
+ ep = Episode("reserved-original-cal", "battery", SCHEMAS["battery"], np.arange(steps + 1) * 15,
22
+ values, masks, actions, np.ones_like(actions) * 12, {"capacity": 999}, 42, "validation")
23
+ transformed = sparse_battery_episode(ep)
24
+ assert transformed.schema == ["state_of_charge", "battery_power"]
25
+ assert ep.schema == SCHEMAS["battery"]
26
+ assert np.array_equal(ep.values, values)
27
+ omitted = np.ones(NUM_FIELDS, dtype=bool)
28
+ omitted[[SOC, POWER]] = False
29
+ assert not transformed.values[:, omitted].any()
30
+ assert not transformed.observation_mask[:, omitted].any()
31
+ assert not transformed.actions[:, omitted].any()
32
+ assert not transformed.exogenous.any()
33
+ norm = Normalization(np.ones(NUM_FIELDS) * 2, np.ones(NUM_FIELDS) * 3, ["original-train"])
34
+ dataset = WindowDataset([transformed], norm, samples=1)
35
+ sample = dataset[0]
36
+ assert sample["field_mask"].nonzero().flatten().tolist() == [SOC, POWER]
37
+ for key in ("history", "history_mask", "future_actions", "future_exogenous"):
38
+ assert not sample[key][:, omitted].any()
39
+ for key in ("current", "current_mask"):
40
+ assert not sample[key][omitted].any()
41
+ assert not sample["future_exogenous"].any()
42
+ assert set(sample["query_fields"][sample["target_mask"]].tolist()) == {SOC}
43
+ assert transformed.split == "validation" and transformed.system_id == ep.system_id
44
+
45
+
46
+ def test_soc_mae_and_bounds_are_raw_unclipped():
47
+ norm = Normalization(np.zeros(NUM_FIELDS), np.ones(NUM_FIELDS), ["train"])
48
+ arrays = {"current": np.zeros((1, NUM_FIELDS)), "target": np.array([[0.1, -0.1]]),
49
+ "query_fields": np.array([[SOC, SOC]]), "target_mask": np.ones((1, 2), bool),
50
+ "query_horizons": np.array([[15., 30.]])}
51
+ arrays["current"][:, SOC] = 0.8
52
+ result = prediction_summary(np.array([[0.4, -0.9]]), arrays, norm)
53
+ assert result["raw_soc_bounds_violation_rate"] == 1
54
+ assert result["soc_mae_fraction"] == 0.55
55
+ assert not result["clipping_applied"]
studies/robustness-seed42-20261011/tests/test_history_diagnostics.py ADDED
@@ -0,0 +1,210 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import numpy as np
2
+ import pytest
3
+ import torch
4
+
5
+ from scripts.history_diagnostics import (
6
+ CONDITIONS,
7
+ HistoryView,
8
+ ModelView,
9
+ match_donors,
10
+ perturb,
11
+ validate_partitions,
12
+ )
13
+ from tfm.data import FIELD_INDEX, Normalization, WindowDataset, generate_episode
14
+ from tfm.models import build_model
15
+
16
+
17
+ def dataset():
18
+ train = generate_episode("thermal", "train", days=0.5, seed=31, split="train", missing_probability=0)
19
+ episodes = [
20
+ generate_episode("thermal", f"test-{i}", days=0.5, seed=33 + i, split="test", missing_probability=0)
21
+ for i in range(3)
22
+ ]
23
+ for episode in episodes:
24
+ episode.observation_mask[:, :3] = True
25
+ return WindowDataset(episodes, Normalization.fit([train]), history_steps=12, samples=32, seed=42)
26
+
27
+
28
+ def test_history_conditions_preserve_labels_current_and_future_plan():
29
+ source = dataset()[0]
30
+ for condition in CONDITIONS:
31
+ if condition == "thermal_temperature_only":
32
+ continue
33
+ changed = perturb(source, condition, seed=42, donor=dataset()[1])
34
+ for key in source.keys() - {"history", "history_mask"}:
35
+ torch.testing.assert_close(changed[key], source[key], rtol=0, atol=0)
36
+ torch.testing.assert_close(changed["history"][-1], source["history"][-1], rtol=0, atol=0)
37
+ torch.testing.assert_close(changed["history_mask"][-1], source["history_mask"][-1], rtol=0, atol=0)
38
+ removed = perturb(source, "current_only")
39
+ assert torch.count_nonzero(removed["history_mask"][:-1]) == 0
40
+ state = perturb(source, "past_state_only")
41
+ assert torch.count_nonzero(state["history_mask"][:-1, source["role_ids"] != 0]) == 0
42
+ torch.testing.assert_close(
43
+ state["history"][:, source["role_ids"] == 0], source["history"][:, source["role_ids"] == 0]
44
+ )
45
+ for length in (4, 8):
46
+ short = perturb(source, f"truncate_{length}")
47
+ assert torch.count_nonzero(short["history_mask"][:-length]) == 0
48
+ torch.testing.assert_close(short["history"][-length:], source["history"][-length:])
49
+ # No mutation of the source, including tensors sharing a last history/current view.
50
+ torch.testing.assert_close(source["history"], dataset()[0]["history"])
51
+
52
+
53
+ def test_temporal_shuffle_joint_order_and_target_independence():
54
+ source = dataset()[0]
55
+ shuffled = perturb(source, "temporal_shuffle", seed=44)
56
+ order = np.random.default_rng(44).permutation(len(source["history"]) - 1)
57
+ torch.testing.assert_close(shuffled["history"][:-1], source["history"][:-1][order])
58
+ torch.testing.assert_close(shuffled["history_mask"][:-1], source["history_mask"][:-1][order])
59
+ corrupted = {key: value.clone() for key, value in source.items()}
60
+ corrupted["target"] += 1e8
61
+ again = perturb(corrupted, "temporal_shuffle", seed=44)
62
+ torch.testing.assert_close(again["history"], shuffled["history"])
63
+
64
+
65
+ def test_donors_are_other_system_same_family_and_ignore_targets():
66
+ base = dataset()
67
+ donors, distances = match_donors(base)
68
+ assert len(distances) == len(base)
69
+ for index, donor in enumerate(donors):
70
+ recipient_ep = base.episodes[base.indices[index][0]]
71
+ donor_ep = base.episodes[base.indices[donor][0]]
72
+ assert recipient_ep.system_id != donor_ep.system_id
73
+ assert recipient_ep.family_id == donor_ep.family_id
74
+ changed = HistoryView(base, "matched_family_swap", donors=donors)[index]
75
+ torch.testing.assert_close(changed["history"][:-1], base[donor]["history"][:-1])
76
+ torch.testing.assert_close(changed["history"][-1], base[index]["history"][-1])
77
+
78
+ class ChangedLabels(HistoryView):
79
+ def __getitem__(self, index):
80
+ item = super().__getitem__(index)
81
+ item["target"] += 1e8
82
+ item["future_actions"] += 1e8
83
+ return item
84
+
85
+ assert match_donors(ChangedLabels(base, "full_history"))[0] == donors
86
+ single = WindowDataset([base.episodes[0]], base.normalization, history_steps=12, samples=4)
87
+ with pytest.raises(ValueError, match="No same-family"):
88
+ match_donors(single)
89
+
90
+
91
+ def test_temperature_only_retains_plans_and_scores_only_observed_temperature():
92
+ source = dataset()[0]
93
+ changed = perturb(source, "thermal_temperature_only", family="thermal")
94
+ indoor = FIELD_INDEX["indoor_temp"]
95
+ torch.testing.assert_close(changed["history"][:, indoor], source["history"][:, indoor])
96
+ assert changed["current_mask"].sum() == source["current_mask"][indoor]
97
+ assert torch.all(changed["query_fields"][changed["target_mask"]] == indoor)
98
+ for key in ("field_mask", "future_actions", "future_exogenous", "target"):
99
+ torch.testing.assert_close(changed[key], source[key])
100
+ for field in (FIELD_INDEX["mass_temp"], FIELD_INDEX["energy_kwh"]):
101
+ assert changed["history_mask"][:, field].sum() == 0
102
+ assert changed["current"][field] == 0
103
+
104
+
105
+ def test_history_branch_disabled_restores_flag_and_labels_never_enter_model():
106
+ source = dataset()[0]
107
+ batch = {key: value.unsqueeze(0) for key, value in source.items() if key not in ("target", "target_mask")}
108
+ model = build_model(
109
+ "transformer", {"history_steps": 12, "d_model": 16, "layers": 1, "heads": 4, "ff_dim": 32}
110
+ ).eval()
111
+ wrapped = ModelView(model, disable_history=True)
112
+ with torch.inference_mode():
113
+ expected = wrapped(batch)
114
+ batch["history"] += 1e6
115
+ actual = wrapped(batch)
116
+ torch.testing.assert_close(expected[0], actual[0])
117
+ assert model.encoder.use_history is True
118
+ with pytest.raises(ValueError, match="Targets"):
119
+ wrapped({**batch, "target": source["target"]})
120
+
121
+
122
+ def test_reserved_calibration_partition_leakage_is_rejected():
123
+ base = dataset()
124
+ train = generate_episode("thermal", "train", days=0.5, seed=31, split="train")
125
+ calibration = generate_episode("thermal", "cal", days=0.5, seed=32, split="validation")
126
+ splits = {"train": [train], "validation": [calibration], "test": base.episodes}
127
+ checkpoint = {
128
+ "training_system_ids": ["train"],
129
+ "validation_selection_system_ids": [],
130
+ "calibration_system_ids": ["cal"],
131
+ "test_system_ids": [e.system_id for e in base.episodes],
132
+ }
133
+ validate_partitions(splits, checkpoint, base.normalization)
134
+ checkpoint["validation_selection_system_ids"] = ["cal"]
135
+ with pytest.raises(ValueError, match="overlap"):
136
+ validate_partitions(splits, checkpoint, base.normalization)
137
+
138
+
139
+ def test_diagnostic_end_to_end_binds_exact_anchors_and_reserved_calibration(tmp_path):
140
+ import json
141
+
142
+ from scripts.history_diagnostics import parser, run
143
+ from tfm.data import generate_dataset
144
+ from tfm.evaluation.benchmark import sha256
145
+ from tfm.training.train import load_manifest_episodes, split_validation_systems
146
+
147
+ data, directory, output = tmp_path / "data", tmp_path / "checkpoint", tmp_path / "report"
148
+ generate_dataset(data, systems=30, days=0.5, seed=10)
149
+ _, splits = load_manifest_episodes(data)
150
+ selection, calibration = split_validation_systems(splits["validation"], 42)
151
+ directory.mkdir()
152
+ norm = Normalization.fit(splits["train"])
153
+ norm.save(directory / "normalization.json")
154
+ model = build_model(
155
+ "transformer", {"history_steps": 12, "d_model": 16, "layers": 1, "heads": 4, "ff_dim": 32}
156
+ )
157
+
158
+ def ids(episodes):
159
+ return sorted({str(episode.system_id) for episode in episodes})
160
+
161
+ torch.save(
162
+ {
163
+ "model_kind": "transformer",
164
+ "model_config": model.config,
165
+ "model_state": model.state_dict(),
166
+ "training_system_ids": ids(splits["train"]),
167
+ "validation_selection_system_ids": ids(selection),
168
+ "calibration_system_ids": ids(calibration),
169
+ "test_system_ids": ids(splits["test"]),
170
+ "normalization_sha256": sha256(directory / "normalization.json"),
171
+ "manifest_sha256": sha256(data / "manifest.json"),
172
+ },
173
+ directory / "checkpoint.pt",
174
+ )
175
+ args = parser().parse_args(
176
+ [
177
+ "--data",
178
+ str(data),
179
+ "--checkpoint",
180
+ str(directory / "checkpoint.pt"),
181
+ "--output",
182
+ str(output),
183
+ "--samples",
184
+ "24",
185
+ "--calibration-samples",
186
+ "64",
187
+ "--counterfactual-samples",
188
+ "0",
189
+ "--conditions",
190
+ "full_history",
191
+ "current_only",
192
+ "thermal_temperature_only",
193
+ ]
194
+ )
195
+ result = run(args)
196
+ assert result["status"] == "completed"
197
+ full = np.load(output / "full_history-test.npz")
198
+ removed = np.load(output / "current_only-test.npz")
199
+ for key in ("target", "target_mask", "query_fields", "query_horizons", "current_mask", "system_id"):
200
+ np.testing.assert_array_equal(full[key], removed[key])
201
+ receipt = json.loads((output / "provenance.json").read_text())
202
+ assert receipt["test"]["window_sampling"]["anchor_sha256"]
203
+ assert len(receipt["test"]["anchors"]) == 24
204
+ assert not set(receipt["partitions"]["test_system_ids"]) & set(
205
+ receipt["partitions"]["calibration_system_ids"]
206
+ )
207
+ for condition in args.conditions:
208
+ bound = json.loads((output / f"calibration-{condition}.json").read_text())
209
+ assert bound["calibration_system_ids"] == ids(calibration)
210
+ assert bound["checkpoint_sha256"] == [sha256(directory / "checkpoint.pt")]
studies/robustness-seed42-20261011/tests/test_robustness_release.py ADDED
@@ -0,0 +1,53 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import json
2
+
3
+ import pytest
4
+
5
+ from scripts.build_robustness_release import build, check_hashes, copy_files, sha256
6
+
7
+
8
+ def test_provenance_rejects_changed_bytes(tmp_path):
9
+ source = tmp_path / "source.py"
10
+ source.write_text("original")
11
+ digest = sha256(source)
12
+ source.write_text("changed")
13
+ with pytest.raises(ValueError, match="Changed provenance"):
14
+ check_hashes(tmp_path, {"source.py": digest})
15
+
16
+
17
+ def test_provenance_rejects_external_and_symlink_paths(tmp_path):
18
+ root = tmp_path / "root"
19
+ root.mkdir()
20
+ outside = tmp_path / "outside"
21
+ outside.write_text("external")
22
+ with pytest.raises(ValueError, match="Unsafe"):
23
+ check_hashes(root, {"../outside": sha256(outside)})
24
+ (root / "link").symlink_to(outside)
25
+ with pytest.raises(ValueError, match="Unsafe"):
26
+ check_hashes(root, {"link": sha256(outside)})
27
+
28
+
29
+ def test_copy_refuses_training_resume_artifact(tmp_path):
30
+ source = tmp_path / "source"
31
+ source.mkdir()
32
+ (source / "last.pt").write_bytes(b"unfinished")
33
+ with pytest.raises(ValueError, match="Unfinished"):
34
+ copy_files(source, tmp_path / "release")
35
+
36
+
37
+ def test_running_diagnostic_blocks_publication_before_output_created(tmp_path):
38
+ root = tmp_path / "root"
39
+ reports = (
40
+ "history/history-diagnostics.json",
41
+ "thermal-evaluation/thermal-evaluation.json",
42
+ "challenge-interventions/interventions.json",
43
+ "citylearn/evaluation/citylearn-evaluation.json",
44
+ "behavior/behavior-diagnostic.json",
45
+ )
46
+ for i, name in enumerate(reports):
47
+ path = root / "artifacts/robustness" / name
48
+ path.parent.mkdir(parents=True, exist_ok=True)
49
+ path.write_text(json.dumps({"status": "running" if i == 0 else "completed"}))
50
+ output = tmp_path / "release"
51
+ with pytest.raises(ValueError, match="unfinished"):
52
+ build(root, output, tmp_path / "license")
53
+ assert not output.exists()
studies/robustness-seed42-20261011/tests/test_thermal_challenges.py ADDED
@@ -0,0 +1,122 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Physics and leakage checks for the independent synthetic holdouts."""
2
+
3
+ from dataclasses import replace
4
+
5
+ import numpy as np
6
+ import pytest
7
+
8
+ from scripts.thermal_challenges import (
9
+ COHORTS,
10
+ generate_dataset,
11
+ generate_episode,
12
+ integrate_interval,
13
+ load_challenge_windows,
14
+ sample_parameters,
15
+ sha256,
16
+ thermal_rhs,
17
+ )
18
+ from tfm.data.normalization import Normalization
19
+ from tfm.data.schema import FIELD_INDEX, FIELD_SPECS, NUM_FIELDS
20
+
21
+
22
+ @pytest.mark.parametrize("cohort", COHORTS)
23
+ def test_zero_input_equilibrium(cohort):
24
+ p = sample_parameters(cohort, 351)
25
+ state = np.array([21.0, 21.0, 21.0, 0.0, 5.0])
26
+ result = integrate_interval(p, state, 0, [21, 0, 0], 0, 2)
27
+ np.testing.assert_allclose(result, state, atol=1e-10)
28
+
29
+
30
+ @pytest.mark.parametrize("cohort", COHORTS)
31
+ def test_instantaneous_heat_accounting(cohort):
32
+ p = sample_parameters(cohort, 991)
33
+ state = np.array([20.0, 23.0, 19.0, 0.25, 0.0])
34
+ exog = [8.0, 0.4, 0.2]
35
+ rhs = thermal_rhs(p, state, 0.7, exog, True)
36
+ stored_rate = np.dot([p.air_capacity, p.wall_capacity, p.furniture_capacity], rhs[:3])
37
+ delivered = state[3] if p.actuator_lag_hours else min(0.7, p.max_heating_power)
38
+ expected = ((p.envelope_conductance + p.ventilation_conductance) * (exog[0] - state[0])
39
+ + p.wall_outside_conductance * (exog[0] - state[1])
40
+ + p.hvac_efficiency * delivered + exog[1] + exog[2])
41
+ assert stored_rate == pytest.approx(expected, abs=1e-12)
42
+ assert rhs[-1] == abs(delivered)
43
+
44
+
45
+ def test_integrated_meter_and_storage_with_lag():
46
+ p = sample_parameters("small_room_3node", 10)
47
+ p = replace(p, envelope_conductance=0, wall_outside_conductance=0,
48
+ ventilation_conductance=0, max_heating_power=0.5, actuator_lag_hours=0.1)
49
+ start = np.array([20.0, 20.0, 20.0, 0.0, 0.0])
50
+ result = integrate_interval(p, start, 15, [20, 0, 0], 0, 0.25)
51
+ expected_energy = 0.5 * (0.25 - 0.1 * (1 - np.exp(-0.25 / 0.1)))
52
+ assert result[-1] == pytest.approx(expected_energy, abs=1e-9)
53
+ capacities = np.array([p.air_capacity, p.wall_capacity, p.furniture_capacity])
54
+ stored_change = capacities @ (result[:3] - start[:3])
55
+ assert stored_change == pytest.approx(expected_energy * p.hvac_efficiency, abs=1e-9)
56
+ assert result[3] < p.max_heating_power
57
+
58
+
59
+ def test_closed_energy_meter_without_lag():
60
+ p = sample_parameters("parameter_shift_2node", 45)
61
+ start = np.array([20.0, 20.0, 20.0, 0.0, 2.0])
62
+ result = integrate_interval(p, start, -15, [20, 0, 0], 0, 0.25)
63
+ assert result[-1] - 2 == pytest.approx(p.max_cooling_power * 0.25, abs=1e-11)
64
+
65
+
66
+ @pytest.mark.parametrize("cohort", COHORTS)
67
+ def test_replay_physical_bounds_and_mapping(cohort):
68
+ a = generate_episode(cohort, 177, 277, steps=48)
69
+ b = generate_episode(cohort, 177, 277, steps=48)
70
+ for name in ("values", "actions", "exogenous", "observation_mask"):
71
+ assert np.array_equal(getattr(a, name), getattr(b, name))
72
+ assert np.all(np.diff(a.values[:, FIELD_INDEX["energy_kwh"]]) >= -1e-10)
73
+ for name in a.schema:
74
+ field = FIELD_INDEX[name]
75
+ values = a.values[a.observation_mask[:, field], field]
76
+ spec = FIELD_SPECS[field]
77
+ assert np.all((values >= spec.minimum) & (values <= spec.maximum))
78
+ assert a.split == "test"
79
+ assert a.hidden_parameters["mass_temp_definition"] == "wall temperature degC"
80
+ with pytest.raises(ValueError, match="train or tune"):
81
+ generate_episode(cohort, 177, 277, split="train")
82
+
83
+
84
+ def test_adapter_is_train_free_and_does_not_leak_hidden_features(tmp_path):
85
+ manifest = generate_dataset(tmp_path, systems_per_cohort=1, calibration_per_cohort=1, steps=40)
86
+ ids_by_split = {s: {e["system_id"] for e in manifest["episodes"] if e["split"] == s}
87
+ for s in ("test", "challenge_calibration")}
88
+ assert not ids_by_split["test"].intersection(ids_by_split["challenge_calibration"])
89
+ assert manifest["frozen_sources_unchanged_during_generation"]
90
+ assert manifest["external_real_world_validation"] is False
91
+ assert manifest["generator_sha256"] == sha256("scripts/thermal_challenges.py")
92
+ norm = Normalization(np.zeros(NUM_FIELDS), np.ones(NUM_FIELDS), ["original-train-system"])
93
+ ds = load_challenge_windows(tmp_path, norm, cohort="small_room_3node")
94
+ assert len(ds) == 10
95
+ first = ds[0]
96
+ ds.episodes[0].hidden_parameters = {"leak": 1e30}
97
+ ds.episodes[0].system_id = "totally-different-id"
98
+ for key, tensor in first.items():
99
+ assert np.array_equal(tensor.numpy(), ds[0][key].numpy())
100
+ assert not any("hidden" in key or "system" in key for key in first)
101
+ assert first["history"].shape == (24, NUM_FIELDS)
102
+ assert first["future_actions"].shape == (8, NUM_FIELDS)
103
+ with pytest.raises(ValueError, match="only training"):
104
+ Normalization.fit(ds.episodes)
105
+ contaminated = Normalization(np.zeros(NUM_FIELDS), np.ones(NUM_FIELDS), list(ids_by_split["test"]))
106
+ with pytest.raises(ValueError, match="overlap"):
107
+ load_challenge_windows(tmp_path, contaminated)
108
+ with pytest.raises(ValueError, match="test/calibration"):
109
+ load_challenge_windows(tmp_path, norm, split="train")
110
+ episode_path = tmp_path / manifest["episodes"][0]["path"]
111
+ episode_path.write_bytes(episode_path.read_bytes() + b"tamper")
112
+ with pytest.raises(ValueError, match="hash mismatch"):
113
+ load_challenge_windows(tmp_path, norm)
114
+
115
+
116
+ def test_capacity_and_heater_shifts_are_outside_training_bounds():
117
+ small = sample_parameters("small_room_3node", 14)
118
+ large = sample_parameters("auditorium_3node", 14)
119
+ assert small.air_capacity < 0.5 and small.wall_capacity < 3
120
+ assert small.max_heating_power < 4 and small.max_cooling_power < 3
121
+ assert large.air_capacity > 3 and large.wall_capacity > 20
122
+ assert large.max_heating_power > 15 and large.max_cooling_power > 14
studies/robustness-seed42-20261011/tests/test_thermal_evaluation.py ADDED
@@ -0,0 +1,160 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Small observation/mask/unit integration checks; no simulator or training."""
2
+
3
+ import numpy as np
4
+ import pytest
5
+ import torch
6
+
7
+ from scripts.context_uncertainty import (
8
+ apply_context_calibration,
9
+ assess_context,
10
+ fit_context_calibration,
11
+ fit_context_support,
12
+ )
13
+ from scripts.evaluate_thermal_challenges import air_summary, observations
14
+ from scripts.history_diagnostics import HistoryView
15
+ from tfm.data import Normalization, WindowDataset
16
+ from tfm.data.episodes import Episode
17
+ from tfm.data.schema import FIELD_INDEX, NUM_FIELDS, SCHEMAS
18
+ from tfm.evaluation.benchmark import collect_predictions, summarize
19
+
20
+
21
+ def episode(system, split):
22
+ steps = 48
23
+ values = np.zeros((steps + 1, NUM_FIELDS))
24
+ mask = np.zeros_like(values, dtype=bool)
25
+ actions = np.zeros((steps, NUM_FIELDS))
26
+ exogenous = np.zeros_like(actions)
27
+ t = np.arange(steps + 1)
28
+ for field, values_field in (
29
+ ("indoor_temp", 18 + 0.25 * t),
30
+ ("mass_temp", 17 + 0.1 * t),
31
+ ("energy_kwh", 0.5 * t),
32
+ ):
33
+ values[:, FIELD_INDEX[field]] = values_field
34
+ mask[:, FIELD_INDEX[field]] = True
35
+ actions[:, FIELD_INDEX["hvac_power"]] = 2 * np.sin(t[:-1])
36
+ for field, source in (
37
+ ("outdoor_temp", 9 + np.sin(t[:-1])),
38
+ ("solar_gain", np.maximum(0, np.sin(t[:-1]))),
39
+ ("occupancy_gain", 0.2 + 0.1 * np.cos(t[:-1])),
40
+ ):
41
+ exogenous[:, FIELD_INDEX[field]] = source
42
+ for field in SCHEMAS["thermal"]:
43
+ index = FIELD_INDEX[field]
44
+ if index >= FIELD_INDEX["hvac_power"]:
45
+ values[:-1, index] = actions[:, index] + exogenous[:, index]
46
+ mask[:-1, index] = True
47
+ return Episode(
48
+ system,
49
+ "thermal",
50
+ SCHEMAS["thermal"],
51
+ t * 15,
52
+ values,
53
+ mask,
54
+ actions,
55
+ exogenous,
56
+ {"room_size": "hidden_unused"},
57
+ 1,
58
+ split,
59
+ )
60
+
61
+
62
+ class LabelRejectingPersistence:
63
+ def __call__(self, inputs):
64
+ assert "target" not in inputs and "target_mask" not in inputs
65
+ assert "room_size" not in inputs and "hidden_parameters" not in inputs
66
+ values = torch.zeros_like(inputs["query_horizons"])
67
+ return values, torch.zeros_like(values)
68
+
69
+
70
+ def datasets():
71
+ train = episode("train", "train")
72
+ normalization = Normalization.fit([train])
73
+ return normalization, {
74
+ name: WindowDataset([ep], normalization, samples=40, history_steps=24, seed=i)
75
+ for i, (name, ep) in enumerate(
76
+ (
77
+ ("train", train),
78
+ ("cal", episode("cal", "validation")),
79
+ ("test", episode("challenge", "test")),
80
+ )
81
+ )
82
+ }
83
+
84
+
85
+ def test_temperature_only_scores_air_and_preserves_physical_delta_units():
86
+ norm, data = datasets()
87
+ base = data["test"]
88
+ original = base[0]
89
+ view = HistoryView(base, "thermal_temperature_only")
90
+ masked = view[0]
91
+ np.testing.assert_array_equal(masked["target"], original["target"])
92
+ np.testing.assert_array_equal(base[0]["history_mask"], original["history_mask"])
93
+ assert masked["current_mask"][FIELD_INDEX["indoor_temp"]]
94
+ assert not masked["current_mask"][FIELD_INDEX["mass_temp"]]
95
+ assert not masked["current_mask"][FIELD_INDEX["energy_kwh"]]
96
+ mean, log_std, arrays = collect_predictions(LabelRejectingPersistence(), view, "cpu", 20)
97
+ active_fields = arrays["query_fields"][arrays["target_mask"].astype(bool)]
98
+ assert set(active_fields.tolist()) == {FIELD_INDEX["indoor_temp"]}
99
+ rows = summarize(mean, log_std, arrays, ["thermal"] * len(view), norm)
100
+ summary = air_summary(rows)
101
+ assert summary["mae_degC"] == pytest.approx(0.9375)
102
+ assert {r["minutes"]: r["mae_degC"] for r in summary["by_horizon"]} == pytest.approx(
103
+ {15: 0.25, 30: 0.5, 60: 1.0, 120: 2.0}
104
+ )
105
+
106
+
107
+ def test_matched_mask_calibration_and_train_support_provenance_integration():
108
+ norm, data = datasets()
109
+ condition = "current_only"
110
+ views = {name: HistoryView(dataset, condition) for name, dataset in data.items()}
111
+ support = fit_context_support(
112
+ observations(views["train"], 20),
113
+ training_system_ids=["train"],
114
+ forbidden_system_ids=["cal", "challenge"],
115
+ normalization_sha256="norm",
116
+ transform_id=condition,
117
+ )
118
+ model = LabelRejectingPersistence()
119
+ cal_mean, cal_log, cal = collect_predictions(model, views["cal"], "cpu", 20)
120
+ cal_context = assess_context(cal, support)
121
+ assert all(row["estimate_eligible"] for row in cal_context)
122
+ assert all("missing_history" in row["reasoning_codes"] for row in cal_context)
123
+ artifact = fit_context_calibration(
124
+ cal_mean,
125
+ np.exp(cal_log),
126
+ cal["target"],
127
+ cal["query_fields"],
128
+ cal["query_horizons"],
129
+ cal["target_mask"],
130
+ cal_context,
131
+ calibration_system_ids=["cal"],
132
+ training_system_ids=["train"],
133
+ test_system_ids=["challenge"],
134
+ checkpoint_sha256=["frozen"],
135
+ normalization_sha256="norm",
136
+ transform_id=condition,
137
+ )
138
+ mean, log_std, arrays = collect_predictions(model, views["test"], "cpu", 20)
139
+ assessed = assess_context(arrays, support)
140
+ intervals = apply_context_calibration(
141
+ mean,
142
+ np.exp(log_std),
143
+ arrays["query_fields"],
144
+ arrays["query_horizons"],
145
+ arrays["target_mask"],
146
+ assessed,
147
+ artifact,
148
+ checkpoint_sha256=["frozen"],
149
+ normalization_sha256="norm",
150
+ transform_id=condition,
151
+ evaluation_system_ids=["challenge"],
152
+ )
153
+ air = arrays["target_mask"].astype(bool) & (arrays["query_fields"] == FIELD_INDEX["indoor_temp"])
154
+ assert intervals["available"][air].all()
155
+ # Normalized delta interval widths convert with scale, not mean or anchor.
156
+ physical_width = (intervals["upper"] - intervals["lower"])[air] * norm.scale[0]
157
+ expected_width = 2 * arrays["query_horizons"][air] / 15 * 0.25
158
+ np.testing.assert_allclose(physical_width, expected_width, rtol=1e-6)
159
+ assert artifact["support_sha256"] == support["sha256"]
160
+ assert artifact["transform_id"] == condition