wuff-mann commited on
Commit
68bca41
·
verified ·
1 Parent(s): 016a566

Qualified Formal bilingual Phase-3A-R3 20260910T135045Z

Browse files
checkpoint/formal_bilingual_continuation/phase3a_r3/20260910T135045Z/qualified/formal_bilingual_phase3a_r3_summary.json ADDED
@@ -0,0 +1,1818 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "run_id": "20260910T135045Z",
3
+ "stage": "formal-bilingual-continuation-phase3a_r3-seq1536-h2112",
4
+ "architecture_freeze": "v6-mature-20260910",
5
+ "parent": {
6
+ "phase2_r2_run": "20260910T052204Z",
7
+ "sidecar_sha256": "96fd0ad71b5e8a681ba0e33e82c9485d44dea01d5a358ccc62d952af4df572c3",
8
+ "optimizer_sha256": "2af311a33963514c27f9de28f03b86a3fe91511caf62bcfe5a436d624ea65ffe",
9
+ "verdict": "FORMAL_BILINGUAL_CONTINUATION_PHASE2_R2_SEQ1024_H1344_DECONFOUNDED_Q2_QUALIFIED_PROMOTE_PHASE3_LONG_CONTEXT_RAMP"
10
+ },
11
+ "data": {
12
+ "train_ranges": {
13
+ "EN": [
14
+ "3300:4222@1536",
15
+ "4322:4629@1536",
16
+ "4729:4883@1536",
17
+ "4983:5290@1536"
18
+ ],
19
+ "ZH": [
20
+ "3360:4282@1536",
21
+ "4382:4689@1536",
22
+ "4789:4942@1536",
23
+ "5042:5350@1536"
24
+ ],
25
+ "CODE": [
26
+ "1700:2160@1536",
27
+ "2240:2394@1536",
28
+ "2474:2551@1536",
29
+ "2631:2784@1536"
30
+ ]
31
+ },
32
+ "domain_counts": {
33
+ "EN": 1690,
34
+ "ZH": 1690,
35
+ "CODE": 844
36
+ },
37
+ "zh_new_id_fraction": 0.6831746101379395,
38
+ "context_preflight": {
39
+ "seq_len": 1536,
40
+ "loss": 4.243109226226807,
41
+ "grad_ok": true,
42
+ "peak_allocated_gb": 8.488282203674316,
43
+ "gpu_vram_gb": 22.0343017578125
44
+ },
45
+ "tokens_per_update": 3072
46
+ },
47
+ "q2_ruler_calibration": {
48
+ "positive": {
49
+ "label": "PHASE2_R2_SELF_POSITIVE",
50
+ "own_gain": 0.09600822659644413,
51
+ "base_full_nll": 8.719880680243174,
52
+ "candidate_full_nll": 8.719880680243174,
53
+ "full_relative_change": 0.0,
54
+ "candidate_target_penalty": 0.16680999100208282,
55
+ "candidate_specificity_abs": 0.20232507586479187,
56
+ "target_delta": 0.0,
57
+ "specificity_delta": 0.0,
58
+ "by_len_gain": {
59
+ "1024": 0.10989073253091672,
60
+ "1536": 0.08396349460015123,
61
+ "2048": 0.09505205061244129
62
+ },
63
+ "legacy_target_attribution": 0.18012232627323804,
64
+ "legacy_specificity_normalized": 0.21847170609655295,
65
+ "by_len_ok": true,
66
+ "ok": true
67
+ },
68
+ "gm_disabled_negative": {
69
+ "label": "GM_DISABLED_NEGATIVE",
70
+ "own_gain": 0.0,
71
+ "base_full_nll": 8.719880680243174,
72
+ "candidate_full_nll": 9.645973488688469,
73
+ "full_relative_change": 0.10620475696915947,
74
+ "candidate_target_penalty": 0.0,
75
+ "candidate_specificity_abs": 0.0,
76
+ "target_delta": -0.16680999100208282,
77
+ "specificity_delta": -0.20232507586479187,
78
+ "by_len_gain": {
79
+ "1024": 0.0,
80
+ "1536": 0.0,
81
+ "2048": 0.0
82
+ },
83
+ "legacy_target_attribution": -Infinity,
84
+ "legacy_specificity_normalized": -Infinity,
85
+ "by_len_ok": false,
86
+ "ok": false
87
+ },
88
+ "identity_erased_negative": {
89
+ "label": "TARGET_IDENTITY_ERASED_NEGATIVE",
90
+ "own_gain": 0.09600822659644413,
91
+ "base_full_nll": 8.719880680243174,
92
+ "candidate_full_nll": 8.719880680243174,
93
+ "full_relative_change": 0.0,
94
+ "candidate_target_penalty": 0.0,
95
+ "candidate_specificity_abs": 0.0,
96
+ "target_delta": -0.16680999100208282,
97
+ "specificity_delta": -0.20232507586479187,
98
+ "by_len_gain": {
99
+ "1024": 0.10989073253091672,
100
+ "1536": 0.08396349460015123,
101
+ "2048": 0.09505205061244129
102
+ },
103
+ "legacy_target_attribution": 0.0,
104
+ "legacy_specificity_normalized": 0.0,
105
+ "by_len_ok": true,
106
+ "ok": false
107
+ },
108
+ "calibrated": true,
109
+ "legacy_parent_ratio_ok": false
110
+ },
111
+ "parent_memory_self": {
112
+ "label": "FORMAL_P3A_PARENT_SELF_DECONFOUNDED",
113
+ "detail": {
114
+ "ZH_ZH": {
115
+ "mode": "ZH_ZH",
116
+ "n": 72,
117
+ "base_full_nll": 7.6896765563223095,
118
+ "candidate_full_nll": 7.6896765563223095,
119
+ "full_relative_change": 0.0,
120
+ "candidate_own_gain": 0.02080420382308312,
121
+ "candidate_target": 0.058420671357048884,
122
+ "candidate_specificity": 0.046388824780782066,
123
+ "target_delta_mean": 0.0,
124
+ "specificity_delta_mean": 0.0,
125
+ "panel_cluster_target_ci95": [
126
+ 0.0,
127
+ 0.0
128
+ ],
129
+ "panel_cluster_specificity_ci95": [
130
+ 0.0,
131
+ 0.0
132
+ ],
133
+ "point_replication_count": 6,
134
+ "replication_ok": true,
135
+ "families": {
136
+ "A": {
137
+ "base_full_nll": 7.562416156133016,
138
+ "candidate_full_nll": 7.562416156133016,
139
+ "full_relative_change": 0.0,
140
+ "candidate_own_gain": 0.015094946306027871,
141
+ "candidate_target": 0.04327989949120416,
142
+ "candidate_specificity": 0.03417368729909261,
143
+ "target_delta": 0.0,
144
+ "specificity_delta": 0.0,
145
+ "gate": true
146
+ },
147
+ "B": {
148
+ "base_full_nll": 7.816936956511603,
149
+ "candidate_full_nll": 7.816936956511603,
150
+ "full_relative_change": 0.0,
151
+ "candidate_own_gain": 0.026264925521636368,
152
+ "candidate_target": 0.07356144322289361,
153
+ "candidate_specificity": 0.05860396226247152,
154
+ "target_delta": 0.0,
155
+ "specificity_delta": 0.0,
156
+ "gate": true
157
+ }
158
+ },
159
+ "panel_detail": [
160
+ {
161
+ "panel_id": 0,
162
+ "candidate_own_gain": 0.019004303125433306,
163
+ "full_relative_change": 0.0,
164
+ "candidate_target": 0.04733586311340332,
165
+ "candidate_specificity": 0.041690826416015625,
166
+ "target_delta_mean": 0.0,
167
+ "specificity_delta_mean": 0.0,
168
+ "point_gate": true
169
+ },
170
+ {
171
+ "panel_id": 1,
172
+ "candidate_own_gain": 0.013636354830070245,
173
+ "full_relative_change": 0.0,
174
+ "candidate_target": 0.04487264156341553,
175
+ "candidate_specificity": 0.030605514844258625,
176
+ "target_delta_mean": 0.0,
177
+ "specificity_delta_mean": 0.0,
178
+ "point_gate": true
179
+ },
180
+ {
181
+ "panel_id": 2,
182
+ "candidate_own_gain": 0.034075647938531836,
183
+ "full_relative_change": 0.0,
184
+ "candidate_target": 0.1000661849975586,
185
+ "candidate_specificity": 0.0844579537709554,
186
+ "target_delta_mean": 0.0,
187
+ "specificity_delta_mean": 0.0,
188
+ "point_gate": true
189
+ },
190
+ {
191
+ "panel_id": 3,
192
+ "candidate_own_gain": 0.014551428370012907,
193
+ "full_relative_change": 0.0,
194
+ "candidate_target": 0.05477726459503174,
195
+ "candidate_specificity": 0.05717186133066813,
196
+ "target_delta_mean": 0.0,
197
+ "specificity_delta_mean": 0.0,
198
+ "point_gate": true
199
+ },
200
+ {
201
+ "panel_id": 4,
202
+ "candidate_own_gain": 0.024038597403966615,
203
+ "full_relative_change": 0.0,
204
+ "candidate_target": 0.06142707665761312,
205
+ "candidate_specificity": 0.04089252154032389,
206
+ "target_delta_mean": 0.0,
207
+ "specificity_delta_mean": 0.0,
208
+ "point_gate": true
209
+ },
210
+ {
211
+ "panel_id": 5,
212
+ "candidate_own_gain": 0.02046405906523933,
213
+ "full_relative_change": 0.0,
214
+ "candidate_target": 0.042044997215270996,
215
+ "candidate_specificity": 0.023514270782470703,
216
+ "target_delta_mean": 0.0,
217
+ "specificity_delta_mean": 0.0,
218
+ "point_gate": true
219
+ }
220
+ ],
221
+ "primary_gate": true
222
+ },
223
+ "ZH_EN": {
224
+ "mode": "ZH_EN",
225
+ "n": 72,
226
+ "base_full_nll": 7.812194380495283,
227
+ "candidate_full_nll": 7.812194380495283,
228
+ "full_relative_change": 0.0,
229
+ "candidate_own_gain": 0.09402039987128685,
230
+ "candidate_target": 0.2181521389219496,
231
+ "candidate_specificity": 0.14704083734088474,
232
+ "target_delta_mean": 0.0,
233
+ "specificity_delta_mean": 0.0,
234
+ "panel_cluster_target_ci95": [
235
+ 0.0,
236
+ 0.0
237
+ ],
238
+ "panel_cluster_specificity_ci95": [
239
+ 0.0,
240
+ 0.0
241
+ ],
242
+ "point_replication_count": 5,
243
+ "replication_ok": true,
244
+ "families": {
245
+ "A": {
246
+ "base_full_nll": 7.774835493829515,
247
+ "candidate_full_nll": 7.774835493829515,
248
+ "full_relative_change": 0.0,
249
+ "candidate_own_gain": 0.09739325891861941,
250
+ "candidate_target": 0.25972653759850395,
251
+ "candidate_specificity": 0.19144731097751194,
252
+ "target_delta": 0.0,
253
+ "specificity_delta": 0.0,
254
+ "gate": true
255
+ },
256
+ "B": {
257
+ "base_full_nll": 7.849553267161052,
258
+ "candidate_full_nll": 7.849553267161052,
259
+ "full_relative_change": 0.0,
260
+ "candidate_own_gain": 0.09065470523010799,
261
+ "candidate_target": 0.17657774024539524,
262
+ "candidate_specificity": 0.10263436370425755,
263
+ "target_delta": 0.0,
264
+ "specificity_delta": 0.0,
265
+ "gate": true
266
+ }
267
+ },
268
+ "panel_detail": [
269
+ {
270
+ "panel_id": 0,
271
+ "candidate_own_gain": 0.07209922235225259,
272
+ "full_relative_change": 0.0,
273
+ "candidate_target": 0.19534393151601157,
274
+ "candidate_specificity": 0.13082647323608398,
275
+ "target_delta_mean": 0.0,
276
+ "specificity_delta_mean": 0.0,
277
+ "point_gate": true
278
+ },
279
+ {
280
+ "panel_id": 1,
281
+ "candidate_own_gain": 0.0954328103417744,
282
+ "full_relative_change": 0.0,
283
+ "candidate_target": 0.08984529972076416,
284
+ "candidate_specificity": 0.028788010279337566,
285
+ "target_delta_mean": 0.0,
286
+ "specificity_delta_mean": 0.0,
287
+ "point_gate": true
288
+ },
289
+ {
290
+ "panel_id": 2,
291
+ "candidate_own_gain": 0.11459168011239489,
292
+ "full_relative_change": 0.0,
293
+ "candidate_target": 0.4175782601038615,
294
+ "candidate_specificity": 0.33623091379801434,
295
+ "target_delta_mean": 0.0,
296
+ "specificity_delta_mean": 0.0,
297
+ "point_gate": true
298
+ },
299
+ {
300
+ "panel_id": 3,
301
+ "candidate_own_gain": 0.09210533817946316,
302
+ "full_relative_change": 0.0,
303
+ "candidate_target": 0.12087655067443848,
304
+ "candidate_specificity": 0.010552446047465006,
305
+ "target_delta_mean": 0.0,
306
+ "specificity_delta_mean": 0.0,
307
+ "point_gate": false
308
+ },
309
+ {
310
+ "panel_id": 4,
311
+ "candidate_own_gain": 0.09400773221338901,
312
+ "full_relative_change": 0.0,
313
+ "candidate_target": 0.12044421831766765,
314
+ "candidate_specificity": 0.07453525066375732,
315
+ "target_delta_mean": 0.0,
316
+ "specificity_delta_mean": 0.0,
317
+ "point_gate": true
318
+ },
319
+ {
320
+ "panel_id": 5,
321
+ "candidate_own_gain": 0.0959569173666501,
322
+ "full_relative_change": 0.0,
323
+ "candidate_target": 0.3648245731989543,
324
+ "candidate_specificity": 0.3013119300206502,
325
+ "target_delta_mean": 0.0,
326
+ "specificity_delta_mean": 0.0,
327
+ "point_gate": true
328
+ }
329
+ ],
330
+ "primary_gate": true
331
+ },
332
+ "EN_ZH": {
333
+ "mode": "EN_ZH",
334
+ "n": 96,
335
+ "base_full_nll": 7.621593470374743,
336
+ "candidate_full_nll": 7.621593470374743,
337
+ "full_relative_change": 0.0,
338
+ "candidate_own_gain": 0.028069974330568865,
339
+ "candidate_target": 0.049919510881106056,
340
+ "candidate_specificity": 0.03188669184843699,
341
+ "target_delta_mean": 0.0,
342
+ "specificity_delta_mean": 0.0,
343
+ "panel_cluster_target_ci95": [
344
+ 0.0,
345
+ 0.0
346
+ ],
347
+ "panel_cluster_specificity_ci95": [
348
+ 0.0,
349
+ 0.0
350
+ ],
351
+ "point_replication_count": 6,
352
+ "replication_ok": true,
353
+ "families": {
354
+ "A": {
355
+ "base_full_nll": 7.668041209379832,
356
+ "candidate_full_nll": 7.668041209379832,
357
+ "full_relative_change": 0.0,
358
+ "candidate_own_gain": 0.027243833637697144,
359
+ "candidate_target": 0.05507978796958923,
360
+ "candidate_specificity": 0.031673153241475425,
361
+ "target_delta": 0.0,
362
+ "specificity_delta": 0.0,
363
+ "gate": true
364
+ },
365
+ "B": {
366
+ "base_full_nll": 7.575145731369655,
367
+ "candidate_full_nll": 7.575145731369655,
368
+ "full_relative_change": 0.0,
369
+ "candidate_own_gain": 0.028904818209563816,
370
+ "candidate_target": 0.044759233792622886,
371
+ "candidate_specificity": 0.03210023045539856,
372
+ "target_delta": 0.0,
373
+ "specificity_delta": 0.0,
374
+ "gate": true
375
+ }
376
+ },
377
+ "panel_detail": [
378
+ {
379
+ "panel_id": 0,
380
+ "candidate_own_gain": 0.02956777060336121,
381
+ "full_relative_change": 0.0,
382
+ "candidate_target": 0.08214357495307922,
383
+ "candidate_specificity": 0.06245502829551697,
384
+ "target_delta_mean": 0.0,
385
+ "specificity_delta_mean": 0.0,
386
+ "point_gate": true
387
+ },
388
+ {
389
+ "panel_id": 1,
390
+ "candidate_own_gain": 0.030874593174365226,
391
+ "full_relative_change": 0.0,
392
+ "candidate_target": 0.03337019681930542,
393
+ "candidate_specificity": 0.018990129232406616,
394
+ "target_delta_mean": 0.0,
395
+ "specificity_delta_mean": 0.0,
396
+ "point_gate": true
397
+ },
398
+ {
399
+ "panel_id": 2,
400
+ "candidate_own_gain": 0.02429250127859051,
401
+ "full_relative_change": 0.0,
402
+ "candidate_target": 0.0474545955657959,
403
+ "candidate_specificity": 0.0378088653087616,
404
+ "target_delta_mean": 0.0,
405
+ "specificity_delta_mean": 0.0,
406
+ "point_gate": true
407
+ },
408
+ {
409
+ "panel_id": 3,
410
+ "candidate_own_gain": 0.039882064082457216,
411
+ "full_relative_change": 0.0,
412
+ "candidate_target": 0.08233103156089783,
413
+ "candidate_specificity": 0.05035844445228577,
414
+ "target_delta_mean": 0.0,
415
+ "specificity_delta_mean": 0.0,
416
+ "point_gate": true
417
+ },
418
+ {
419
+ "panel_id": 4,
420
+ "candidate_own_gain": 0.01808965637080427,
421
+ "full_relative_change": 0.0,
422
+ "candidate_target": 0.030673682689666748,
423
+ "candidate_specificity": 0.01895490288734436,
424
+ "target_delta_mean": 0.0,
425
+ "specificity_delta_mean": 0.0,
426
+ "point_gate": true
427
+ },
428
+ {
429
+ "panel_id": 5,
430
+ "candidate_own_gain": 0.025830958206635283,
431
+ "full_relative_change": 0.0,
432
+ "candidate_target": 0.023543983697891235,
433
+ "candidate_specificity": 0.0027527809143066406,
434
+ "target_delta_mean": 0.0,
435
+ "specificity_delta_mean": 0.0,
436
+ "point_gate": true
437
+ }
438
+ ],
439
+ "primary_gate": true
440
+ }
441
+ },
442
+ "primary_all": true,
443
+ "replication_all": true,
444
+ "ok": true
445
+ },
446
+ "parent_q2": {
447
+ "n": 96,
448
+ "beta0_nll": 9.645973488688469,
449
+ "full_nll": 8.719880680243174,
450
+ "target_ko_nll": 8.886690671245256,
451
+ "control_ko_nll": 8.684365595380465,
452
+ "gain_abs": 0.9260928084452953,
453
+ "gain": 0.09600822659644413,
454
+ "target_penalty": 0.16680999100208282,
455
+ "specificity_abs": 0.20232507586479187,
456
+ "target_attribution": 0.18012232627323804,
457
+ "specificity_normalized": 0.21847170609655295,
458
+ "by_len": {
459
+ "1024": {
460
+ "gain": 0.10989073253091672
461
+ },
462
+ "1536": {
463
+ "gain": 0.08396349460015123
464
+ },
465
+ "2048": {
466
+ "gain": 0.09505205061244129
467
+ }
468
+ }
469
+ },
470
+ "branch": {
471
+ "branch": "FORMAL_BILINGUAL_CONTINUATION_PHASE3A_R3_SEQ1536_H2112",
472
+ "finite_ok": true,
473
+ "grad_ok": true,
474
+ "gradient_audit": {
475
+ "L23_selector": {
476
+ "mean_norm": 0.0020665300945645576,
477
+ "max_norm": 0.013878784142434597,
478
+ "nonzero_tensors": 14,
479
+ "tensor_count": 14
480
+ },
481
+ "L23_qk": {
482
+ "mean_norm": 0.11805308610200882,
483
+ "max_norm": 0.12682388722896576,
484
+ "nonzero_tensors": 2,
485
+ "tensor_count": 2
486
+ },
487
+ "L23_propagation": {
488
+ "mean_norm": 0.18262008260935544,
489
+ "max_norm": 0.46417051553726196,
490
+ "nonzero_tensors": 5,
491
+ "tensor_count": 5
492
+ },
493
+ "L24_selector": {
494
+ "mean_norm": 0.004680941084058888,
495
+ "max_norm": 0.029577363282442093,
496
+ "nonzero_tensors": 14,
497
+ "tensor_count": 14
498
+ },
499
+ "L24_qk": {
500
+ "mean_norm": 1.1244141459465027,
501
+ "max_norm": 1.1580697298049927,
502
+ "nonzero_tensors": 2,
503
+ "tensor_count": 2
504
+ },
505
+ "L24_propagation": {
506
+ "mean_norm": 0.25356794744729994,
507
+ "max_norm": 0.5262930393218994,
508
+ "nonzero_tensors": 5,
509
+ "tensor_count": 5
510
+ },
511
+ "L25_selector": {
512
+ "mean_norm": 0.001510086879404428,
513
+ "max_norm": 0.006543087773025036,
514
+ "nonzero_tensors": 14,
515
+ "tensor_count": 14
516
+ },
517
+ "L25_qk": {
518
+ "mean_norm": 0.17664532363414764,
519
+ "max_norm": 0.2344283163547516,
520
+ "nonzero_tensors": 2,
521
+ "tensor_count": 2
522
+ },
523
+ "L25_propagation": {
524
+ "mean_norm": 0.17275904091075062,
525
+ "max_norm": 0.3970559239387512,
526
+ "nonzero_tensors": 5,
527
+ "tensor_count": 5
528
+ }
529
+ },
530
+ "protected_diff": {
531
+ "reader_q": 0.0,
532
+ "reader_k": 0.0,
533
+ "source_writer": 0.0,
534
+ "embedding_sample": 0.0,
535
+ "v_proj": {
536
+ "23": 0.0,
537
+ "24": 0.0,
538
+ "25": 0.0
539
+ }
540
+ },
541
+ "protected_ok": true,
542
+ "dtn_param_maxdiff": 0.0,
543
+ "dtn_params_exact": true,
544
+ "phase_audits": [
545
+ {
546
+ "update": 192,
547
+ "phase": 1,
548
+ "opt_weighted": 0.0008381095203276497,
549
+ "real_composite": -4.195342858744178e-05,
550
+ "min_real_improvement": -0.001025599186985094,
551
+ "target_hinge_rate_to_date": 0.2708333333333333,
552
+ "specificity_hinge_rate_to_date": 0.3541666666666667
553
+ },
554
+ {
555
+ "update": 384,
556
+ "phase": 2,
557
+ "opt_weighted": 0.0014358412280924477,
558
+ "real_composite": 0.0007723087440813732,
559
+ "min_real_improvement": -0.0009293115105575257,
560
+ "target_hinge_rate_to_date": 0.2708333333333333,
561
+ "specificity_hinge_rate_to_date": 0.3020833333333333
562
+ },
563
+ {
564
+ "update": 576,
565
+ "phase": 3,
566
+ "opt_weighted": 0.001640359055106617,
567
+ "real_composite": 0.0006406577754614002,
568
+ "min_real_improvement": -0.00173041091423547,
569
+ "target_hinge_rate_to_date": 0.2847222222222222,
570
+ "specificity_hinge_rate_to_date": 0.3055555555555556
571
+ },
572
+ {
573
+ "update": 768,
574
+ "phase": 4,
575
+ "opt_weighted": 0.0026395329356910264,
576
+ "real_composite": 0.0011122472765670555,
577
+ "min_real_improvement": -0.0017151095861583274,
578
+ "target_hinge_rate_to_date": 0.296875,
579
+ "specificity_hinge_rate_to_date": 0.3020833333333333
580
+ },
581
+ {
582
+ "update": 960,
583
+ "phase": 5,
584
+ "opt_weighted": 0.0031223201380795843,
585
+ "real_composite": 0.0014877711318855228,
586
+ "min_real_improvement": -0.0017966884540215277,
587
+ "target_hinge_rate_to_date": 0.2875,
588
+ "specificity_hinge_rate_to_date": 0.29583333333333334
589
+ },
590
+ {
591
+ "update": 1152,
592
+ "phase": 6,
593
+ "opt_weighted": 0.0033789557840021535,
594
+ "real_composite": 0.0021414301379028655,
595
+ "min_real_improvement": -0.000967773189260369,
596
+ "target_hinge_rate_to_date": 0.2847222222222222,
597
+ "specificity_hinge_rate_to_date": 0.2986111111111111
598
+ },
599
+ {
600
+ "update": 1344,
601
+ "phase": 7,
602
+ "opt_weighted": 0.003798681527255504,
603
+ "real_composite": 0.002050242924327196,
604
+ "min_real_improvement": -0.0020720407361722874,
605
+ "target_hinge_rate_to_date": 0.2976190476190476,
606
+ "specificity_hinge_rate_to_date": 0.3005952380952381
607
+ },
608
+ {
609
+ "update": 1536,
610
+ "phase": 8,
611
+ "opt_weighted": 0.0041757792453815355,
612
+ "real_composite": 0.0022594447212905006,
613
+ "min_real_improvement": -0.0018178954436161036,
614
+ "target_hinge_rate_to_date": 0.296875,
615
+ "specificity_hinge_rate_to_date": 0.2916666666666667
616
+ },
617
+ {
618
+ "update": 1728,
619
+ "phase": 9,
620
+ "opt_weighted": 0.00456536152337141,
621
+ "real_composite": 0.0031024169595494583,
622
+ "min_real_improvement": -0.0014139479908267976,
623
+ "target_hinge_rate_to_date": 0.29398148148148145,
624
+ "specificity_hinge_rate_to_date": 0.2962962962962963
625
+ },
626
+ {
627
+ "update": 1920,
628
+ "phase": 10,
629
+ "opt_weighted": 0.005042930208916899,
630
+ "real_composite": 0.0030527464300098743,
631
+ "min_real_improvement": -0.0018344839134158593,
632
+ "target_hinge_rate_to_date": 0.3,
633
+ "specificity_hinge_rate_to_date": 0.3020833333333333
634
+ },
635
+ {
636
+ "update": 2112,
637
+ "phase": 11,
638
+ "opt_weighted": 0.005583029538533179,
639
+ "real_composite": 0.0031049427428773762,
640
+ "min_real_improvement": -0.0018872135250855256,
641
+ "target_hinge_rate_to_date": 0.29545454545454547,
642
+ "specificity_hinge_rate_to_date": 0.29545454545454547
643
+ }
644
+ ],
645
+ "phase_stable": true,
646
+ "final_opt_post": {
647
+ "EN": 4.026940981547038,
648
+ "ZH": 6.593978722890218,
649
+ "CODE": 1.728573203086853
650
+ },
651
+ "final_opt_relative": {
652
+ "EN": 0.001783439856842553,
653
+ "ZH": 0.011357889190639731,
654
+ "CODE": 0.001632489597701324
655
+ },
656
+ "final_opt_weighted": 0.005583029538533179,
657
+ "real": {
658
+ "general": 5.571628451347351,
659
+ "en_stream": 4.071645379066467,
660
+ "zh": 6.149386823177338,
661
+ "code": 2.415623217821121,
662
+ "reasoning": 4.054291804631551,
663
+ "long_en": 4.172328591346741,
664
+ "long_zh": 6.144298235575358,
665
+ "code_stream": 1.5288460552692413
666
+ },
667
+ "real_relative": {
668
+ "general": -0.0012157218805931296,
669
+ "en_stream": 0.002218192049361158,
670
+ "zh": 0.012649208142147803,
671
+ "code": 0.000972398345418651,
672
+ "reasoning": 0.0004925456754831082,
673
+ "long_en": 0.0016712295007618669,
674
+ "long_zh": 0.013326061691282882,
675
+ "code_stream": 4.6742854400550277e-05
676
+ },
677
+ "real_composite": 0.0037230346826467534,
678
+ "dtn_profile": {
679
+ "domain_nll": {
680
+ "EN": {
681
+ "0": 4.561229109764099,
682
+ "1": 4.609660983085632,
683
+ "2": 4.062892258167267,
684
+ "3": 4.007568120956421
685
+ },
686
+ "ZH": {
687
+ "0": 6.203189373016357,
688
+ "1": 6.131345272064209,
689
+ "2": 5.793992042541504,
690
+ "3": 5.77936577796936
691
+ },
692
+ "CODE": {
693
+ "0": 2.355822205543518,
694
+ "1": 2.3604418635368347,
695
+ "2": 1.831506997346878,
696
+ "3": 1.7631154358386993
697
+ },
698
+ "REASON": {
699
+ "0": 4.157577693462372,
700
+ "1": 4.384354472160339,
701
+ "2": 3.7294474244117737,
702
+ "3": 3.6861140727996826
703
+ }
704
+ },
705
+ "mean_d2_over_d1": 0.13677284322422062,
706
+ "mean_d3_over_d2": 0.01627556203234641,
707
+ "mean_d3_gain_vs_no_dtn": 0.1386746531792376,
708
+ "router_distribution": [
709
+ 0.20756186098556062,
710
+ 0.0007349875108866958,
711
+ 0.29022395829383796,
712
+ 0.4948861763957392,
713
+ 0.001319938994106205,
714
+ 0.005273077818869212
715
+ ],
716
+ "router_entropy_norm": 0.6000922922497945,
717
+ "router_max_share": 0.4948861763957392,
718
+ "output_scale": 0.9994887113571167
719
+ },
720
+ "target_hinge_rate": 0.29545454545454547,
721
+ "specificity_hinge_rate": 0.29545454545454547,
722
+ "optimizer_resume_audit": {
723
+ "state_count": 63,
724
+ "selected_param_count": 63,
725
+ "min_step": 2496.0,
726
+ "max_step": 2496.0
727
+ },
728
+ "state_path": "/content/mann_formal_bilingual_phase3a_r3_20260910T135045Z/branch_states/FORMAL_PHASE3A_R3_SEQ1536_H2112.pt",
729
+ "optimizer_path": "/content/mann_formal_bilingual_phase3a_r3_20260910T135045Z/branch_states/FORMAL_PHASE3A_R3_SEQ1536_H2112_optimizer.pt"
730
+ },
731
+ "adjudication": {
732
+ "candidate": {
733
+ "real": {
734
+ "general": 5.571628451347351,
735
+ "en_stream": 4.071645379066467,
736
+ "zh": 6.149386823177338,
737
+ "code": 2.415623217821121,
738
+ "reasoning": 4.054291804631551,
739
+ "long_en": 4.172328591346741,
740
+ "long_zh": 6.144298235575358,
741
+ "code_stream": 1.5288460552692413
742
+ },
743
+ "q2": {
744
+ "n": 96,
745
+ "beta0_nll": 9.184027234713236,
746
+ "full_nll": 8.06357891112566,
747
+ "target_ko_nll": 8.250807446738085,
748
+ "control_ko_nll": 8.026908290882906,
749
+ "gain_abs": 1.120448323587576,
750
+ "gain": 0.12199967344963575,
751
+ "target_penalty": 0.1872285356124248,
752
+ "specificity_abs": 0.22389915585517883,
753
+ "target_attribution": 0.1671014465111034,
754
+ "specificity_normalized": 0.19982997086226487,
755
+ "by_len": {
756
+ "1024": {
757
+ "gain": 0.1386586672807771
758
+ },
759
+ "1536": {
760
+ "gain": 0.10557065230892543
761
+ },
762
+ "2048": {
763
+ "gain": 0.1229383521894124
764
+ }
765
+ }
766
+ },
767
+ "dtn_profile": {
768
+ "domain_nll": {
769
+ "EN": {
770
+ "0": 4.561229109764099,
771
+ "1": 4.609660983085632,
772
+ "2": 4.062892258167267,
773
+ "3": 4.007568120956421
774
+ },
775
+ "ZH": {
776
+ "0": 6.203189373016357,
777
+ "1": 6.131345272064209,
778
+ "2": 5.793992042541504,
779
+ "3": 5.77936577796936
780
+ },
781
+ "CODE": {
782
+ "0": 2.355822205543518,
783
+ "1": 2.3604418635368347,
784
+ "2": 1.831506997346878,
785
+ "3": 1.7631154358386993
786
+ },
787
+ "REASON": {
788
+ "0": 4.157577693462372,
789
+ "1": 4.384354472160339,
790
+ "2": 3.7294474244117737,
791
+ "3": 3.6861140727996826
792
+ }
793
+ },
794
+ "mean_d2_over_d1": 0.13677284322422062,
795
+ "mean_d3_over_d2": 0.01627556203234641,
796
+ "mean_d3_gain_vs_no_dtn": 0.1386746531792376,
797
+ "router_distribution": [
798
+ 0.20756186098556062,
799
+ 0.0007349875108866958,
800
+ 0.29022395829383796,
801
+ 0.4948861763957392,
802
+ 0.001319938994106205,
803
+ 0.005273077818869212
804
+ ],
805
+ "router_entropy_norm": 0.6000922922497945,
806
+ "router_max_share": 0.4948861763957392,
807
+ "output_scale": 0.9994887113571167
808
+ }
809
+ },
810
+ "old_vs_parent_diagnostic": {
811
+ "label": "FORMAL_P3A_R3_CAND_VS_P2R2_OLD_DIAGNOSTIC",
812
+ "detail": {
813
+ "ZH_ZH": {
814
+ "mode": "ZH_ZH",
815
+ "n": 72,
816
+ "gain_delta": 0.011995152635559964,
817
+ "candidate_target": 0.08275654580858019,
818
+ "candidate_specificity": 0.07006898191240099,
819
+ "target_delta_mean": 0.024335874451531306,
820
+ "specificity_delta_mean": 0.023680157131618924,
821
+ "example_target_ci95": [
822
+ 0.007624159422185686,
823
+ 0.04478917221228282
824
+ ],
825
+ "example_specificity_ci95": [
826
+ 0.008477569619814555,
827
+ 0.04248703486389583
828
+ ],
829
+ "panel_cluster_target_ci95": [
830
+ 0.009639037979973687,
831
+ 0.03883338636822171
832
+ ],
833
+ "panel_cluster_specificity_ci95": [
834
+ 0.010125286049313016,
835
+ 0.03754634989632501
836
+ ],
837
+ "panel_target_means": [
838
+ 0.0176846981048584,
839
+ -0.0008048216501871744,
840
+ 0.047187209129333496,
841
+ 0.04520722230275472,
842
+ 0.03052675724029541,
843
+ 0.0062141815821329755
844
+ ],
845
+ "panel_specificity_means": [
846
+ 0.00758818785349528,
847
+ 0.0012011130650838215,
848
+ 0.03958348433176676,
849
+ 0.05053504308064779,
850
+ 0.02650129795074463,
851
+ 0.01667181650797526
852
+ ],
853
+ "point_replication_count": 6,
854
+ "point_replication_required": 4,
855
+ "replication_ok": true,
856
+ "families": {
857
+ "A": {
858
+ "base_gain": 0.015094946306027871,
859
+ "candidate_gain": 0.017061606134190232,
860
+ "gain_delta": 0.0019666598281623607,
861
+ "base_target": 0.04327989949120416,
862
+ "candidate_target": 0.04163148668077257,
863
+ "target_delta": -0.0016484128104315848,
864
+ "base_specificity": 0.03417368729909261,
865
+ "candidate_specificity": 0.04067420959472656,
866
+ "specificity_delta": 0.0065005222956339495,
867
+ "gate": true
868
+ },
869
+ "B": {
870
+ "base_gain": 0.026264925521636368,
871
+ "candidate_gain": 0.04745204581989143,
872
+ "gain_delta": 0.021187120298255063,
873
+ "base_target": 0.07356144322289361,
874
+ "candidate_target": 0.1238816049363878,
875
+ "target_delta": 0.050320161713494196,
876
+ "base_specificity": 0.05860396226247152,
877
+ "candidate_specificity": 0.09946375423007542,
878
+ "specificity_delta": 0.0408597919676039,
879
+ "gate": true
880
+ }
881
+ },
882
+ "panel_detail": [
883
+ {
884
+ "panel_id": 0,
885
+ "gain_delta": 0.008375405701527897,
886
+ "candidate_target": 0.06502056121826172,
887
+ "candidate_specificity": 0.0492790142695109,
888
+ "target_delta_mean": 0.0176846981048584,
889
+ "specificity_delta_mean": 0.00758818785349528,
890
+ "point_gate": true,
891
+ "legacy_target_ci95": [
892
+ -0.005512800812721252,
893
+ 0.03970179756482441
894
+ ],
895
+ "legacy_specificity_ci95": [
896
+ -0.009346456329027811,
897
+ 0.024619630972544337
898
+ ]
899
+ },
900
+ {
901
+ "panel_id": 1,
902
+ "gain_delta": 0.005273130177232497,
903
+ "candidate_target": 0.044067819913228355,
904
+ "candidate_specificity": 0.031806627909342446,
905
+ "target_delta_mean": -0.0008048216501871744,
906
+ "specificity_delta_mean": 0.0012011130650838215,
907
+ "point_gate": true,
908
+ "legacy_target_ci95": [
909
+ -0.027799330155054727,
910
+ 0.024069310228029883
911
+ ],
912
+ "legacy_specificity_ci95": [
913
+ -0.026555181543032328,
914
+ 0.025082755088806117
915
+ ]
916
+ },
917
+ {
918
+ "panel_id": 2,
919
+ "gain_delta": 0.024496106591995997,
920
+ "candidate_target": 0.1472533941268921,
921
+ "candidate_specificity": 0.12404143810272217,
922
+ "target_delta_mean": 0.047187209129333496,
923
+ "specificity_delta_mean": 0.03958348433176676,
924
+ "point_gate": true,
925
+ "legacy_target_ci95": [
926
+ -0.0005786319573720296,
927
+ 0.10865156749884286
928
+ ],
929
+ "legacy_specificity_ci95": [
930
+ -0.002973184982935587,
931
+ 0.09780842661857606
932
+ ]
933
+ },
934
+ {
935
+ "panel_id": 3,
936
+ "gain_delta": 0.008863635515966865,
937
+ "candidate_target": 0.09998448689778645,
938
+ "candidate_specificity": 0.10770690441131592,
939
+ "target_delta_mean": 0.04520722230275472,
940
+ "specificity_delta_mean": 0.05053504308064779,
941
+ "point_gate": true,
942
+ "legacy_target_ci95": [
943
+ -0.007956488927205404,
944
+ 0.1345404167970021
945
+ ],
946
+ "legacy_specificity_ci95": [
947
+ -0.0004405379295349118,
948
+ 0.1354820390542348
949
+ ]
950
+ },
951
+ {
952
+ "panel_id": 4,
953
+ "gain_delta": 0.010912879770837231,
954
+ "candidate_target": 0.09195383389790852,
955
+ "candidate_specificity": 0.06739381949106853,
956
+ "target_delta_mean": 0.03052675724029541,
957
+ "specificity_delta_mean": 0.02650129795074463,
958
+ "point_gate": true,
959
+ "legacy_target_ci95": [
960
+ 0.01127636432647705,
961
+ 0.05695220331350961
962
+ ],
963
+ "legacy_specificity_ci95": [
964
+ 0.009195156892140706,
965
+ 0.04533674418926238
966
+ ]
967
+ },
968
+ {
969
+ "panel_id": 5,
970
+ "gain_delta": 0.014475055068352529,
971
+ "candidate_target": 0.04825917879740397,
972
+ "candidate_specificity": 0.04018608729044596,
973
+ "target_delta_mean": 0.0062141815821329755,
974
+ "specificity_delta_mean": 0.01667181650797526,
975
+ "point_gate": true,
976
+ "legacy_target_ci95": [
977
+ -0.014536238710085551,
978
+ 0.03314209779103597
979
+ ],
980
+ "legacy_specificity_ci95": [
981
+ -0.00517915387948354,
982
+ 0.0434792637825012
983
+ ]
984
+ }
985
+ ],
986
+ "primary_gate": true
987
+ },
988
+ "ZH_EN": {
989
+ "mode": "ZH_EN",
990
+ "n": 72,
991
+ "gain_delta": 0.07425969452573947,
992
+ "candidate_target": 0.2966230246755812,
993
+ "candidate_specificity": 0.18221448858579,
994
+ "target_delta_mean": 0.07847088575363159,
995
+ "specificity_delta_mean": 0.03517365124490526,
996
+ "example_target_ci95": [
997
+ 0.055551942686239884,
998
+ 0.10359623382488886
999
+ ],
1000
+ "example_specificity_ci95": [
1001
+ 0.0125574824710687,
1002
+ 0.05991005003452301
1003
+ ],
1004
+ "panel_cluster_target_ci95": [
1005
+ 0.04405778646469116,
1006
+ 0.11288398504257204
1007
+ ],
1008
+ "panel_cluster_specificity_ci95": [
1009
+ 0.005836364295747544,
1010
+ 0.0681430564986335
1011
+ ],
1012
+ "panel_target_means": [
1013
+ 0.09079015254974365,
1014
+ 0.016912182172139485,
1015
+ 0.11167379220326741,
1016
+ 0.0726312796274821,
1017
+ 0.03355046113332113,
1018
+ 0.14526744683583578
1019
+ ],
1020
+ "panel_specificity_means": [
1021
+ 0.030432264010111492,
1022
+ -0.01420046885808309,
1023
+ 0.06389741102854411,
1024
+ 0.016755064328511555,
1025
+ 0.0071335236231486005,
1026
+ 0.1070241133371989
1027
+ ],
1028
+ "point_replication_count": 6,
1029
+ "point_replication_required": 4,
1030
+ "replication_ok": true,
1031
+ "families": {
1032
+ "A": {
1033
+ "base_gain": 0.09739325891861941,
1034
+ "candidate_gain": 0.17719632801928212,
1035
+ "gain_delta": 0.0798030691006627,
1036
+ "base_target": 0.25972653759850395,
1037
+ "candidate_target": 0.35266275538338554,
1038
+ "target_delta": 0.09293621778488159,
1039
+ "base_specificity": 0.19144731097751194,
1040
+ "candidate_specificity": 0.2436350319120619,
1041
+ "specificity_delta": 0.05218772093454996,
1042
+ "gate": true
1043
+ },
1044
+ "B": {
1045
+ "base_gain": 0.09065470523010799,
1046
+ "candidate_gain": 0.15971627889655524,
1047
+ "gain_delta": 0.06906157366644725,
1048
+ "base_target": 0.17657774024539524,
1049
+ "candidate_target": 0.24058329396777683,
1050
+ "target_delta": 0.06400555372238159,
1051
+ "base_specificity": 0.10263436370425755,
1052
+ "candidate_specificity": 0.12079394525951809,
1053
+ "specificity_delta": 0.01815958155526054,
1054
+ "gate": true
1055
+ }
1056
+ },
1057
+ "panel_detail": [
1058
+ {
1059
+ "panel_id": 0,
1060
+ "gain_delta": 0.06615701176839013,
1061
+ "candidate_target": 0.2861340840657552,
1062
+ "candidate_specificity": 0.16125873724619547,
1063
+ "target_delta_mean": 0.09079015254974365,
1064
+ "specificity_delta_mean": 0.030432264010111492,
1065
+ "point_gate": true,
1066
+ "legacy_target_ci95": [
1067
+ 0.050472897291183476,
1068
+ 0.131264058748881
1069
+ ],
1070
+ "legacy_specificity_ci95": [
1071
+ -0.005531974633534749,
1072
+ 0.07478085954984028
1073
+ ]
1074
+ },
1075
+ {
1076
+ "panel_id": 1,
1077
+ "gain_delta": 0.08108898648659504,
1078
+ "candidate_target": 0.10675748189290364,
1079
+ "candidate_specificity": 0.014587541421254476,
1080
+ "target_delta_mean": 0.016912182172139485,
1081
+ "specificity_delta_mean": -0.01420046885808309,
1082
+ "point_gate": true,
1083
+ "legacy_target_ci95": [
1084
+ -0.001991110543409983,
1085
+ 0.03426861464977265
1086
+ ],
1087
+ "legacy_specificity_ci95": [
1088
+ -0.04111154476801554,
1089
+ 0.007678879797458647
1090
+ ]
1091
+ },
1092
+ {
1093
+ "panel_id": 2,
1094
+ "gain_delta": 0.07902231182403048,
1095
+ "candidate_target": 0.5292520523071289,
1096
+ "candidate_specificity": 0.4001283248265584,
1097
+ "target_delta_mean": 0.11167379220326741,
1098
+ "specificity_delta_mean": 0.06389741102854411,
1099
+ "point_gate": true,
1100
+ "legacy_target_ci95": [
1101
+ 0.04373111128807068,
1102
+ 0.18313238223393757
1103
+ ],
1104
+ "legacy_specificity_ci95": [
1105
+ 0.01015868882338206,
1106
+ 0.12313870986302691
1107
+ ]
1108
+ },
1109
+ {
1110
+ "panel_id": 3,
1111
+ "gain_delta": 0.07071086189534283,
1112
+ "candidate_target": 0.19350783030192056,
1113
+ "candidate_specificity": 0.027307510375976562,
1114
+ "target_delta_mean": 0.0726312796274821,
1115
+ "specificity_delta_mean": 0.016755064328511555,
1116
+ "point_gate": true,
1117
+ "legacy_target_ci95": [
1118
+ 0.01703299184640249,
1119
+ 0.13958055774370828
1120
+ ],
1121
+ "legacy_specificity_ci95": [
1122
+ -0.048337847987810774,
1123
+ 0.08784319857756297
1124
+ ]
1125
+ },
1126
+ {
1127
+ "panel_id": 4,
1128
+ "gain_delta": 0.06261192016583689,
1129
+ "candidate_target": 0.15399467945098877,
1130
+ "candidate_specificity": 0.08166877428690593,
1131
+ "target_delta_mean": 0.03355046113332113,
1132
+ "specificity_delta_mean": 0.0071335236231486005,
1133
+ "point_gate": true,
1134
+ "legacy_target_ci95": [
1135
+ 0.009774792194366456,
1136
+ 0.056597429513931256
1137
+ ],
1138
+ "legacy_specificity_ci95": [
1139
+ -0.014335973064104715,
1140
+ 0.02628499368826548
1141
+ ]
1142
+ },
1143
+ {
1144
+ "panel_id": 5,
1145
+ "gain_delta": 0.08702881330587786,
1146
+ "candidate_target": 0.51009202003479,
1147
+ "candidate_specificity": 0.4083360433578491,
1148
+ "target_delta_mean": 0.14526744683583578,
1149
+ "specificity_delta_mean": 0.1070241133371989,
1150
+ "point_gate": true,
1151
+ "legacy_target_ci95": [
1152
+ 0.07037907739480337,
1153
+ 0.2336277912060419
1154
+ ],
1155
+ "legacy_specificity_ci95": [
1156
+ 0.03719215591748556,
1157
+ 0.19095355272293085
1158
+ ]
1159
+ }
1160
+ ],
1161
+ "primary_gate": true
1162
+ },
1163
+ "EN_ZH": {
1164
+ "mode": "EN_ZH",
1165
+ "n": 96,
1166
+ "gain_delta": 0.017253679001014263,
1167
+ "candidate_target": 0.07040177782376607,
1168
+ "candidate_specificity": 0.05305696030457815,
1169
+ "target_delta_mean": 0.020482266942660015,
1170
+ "specificity_delta_mean": 0.021170268456141155,
1171
+ "example_target_ci95": [
1172
+ 0.01185892236729463,
1173
+ 0.029881055528918902
1174
+ ],
1175
+ "example_specificity_ci95": [
1176
+ 0.013277628024419149,
1177
+ 0.02962418893973032
1178
+ ],
1179
+ "panel_cluster_target_ci95": [
1180
+ 0.012385209401448568,
1181
+ 0.028008873263994854
1182
+ ],
1183
+ "panel_cluster_specificity_ci95": [
1184
+ 0.015154346823692322,
1185
+ 0.02888934810956319
1186
+ ],
1187
+ "panel_target_means": [
1188
+ 0.0326080322265625,
1189
+ 0.023133963346481323,
1190
+ 0.0035798251628875732,
1191
+ 0.016419291496276855,
1192
+ 0.016476601362228394,
1193
+ 0.030675888061523438
1194
+ ],
1195
+ "panel_specificity_means": [
1196
+ 0.0379844605922699,
1197
+ 0.01248544454574585,
1198
+ 0.01490640640258789,
1199
+ 0.025344550609588623,
1200
+ 0.021552711725234985,
1201
+ 0.014748036861419678
1202
+ ],
1203
+ "point_replication_count": 6,
1204
+ "point_replication_required": 4,
1205
+ "replication_ok": true,
1206
+ "families": {
1207
+ "A": {
1208
+ "base_gain": 0.027243833637697144,
1209
+ "candidate_gain": 0.03870944234537898,
1210
+ "gain_delta": 0.011465608707681835,
1211
+ "base_target": 0.05507978796958923,
1212
+ "candidate_target": 0.06598063309987386,
1213
+ "target_delta": 0.010900845130284623,
1214
+ "base_specificity": 0.031673153241475425,
1215
+ "candidate_specificity": 0.05310255289077759,
1216
+ "specificity_delta": 0.021429399649302162,
1217
+ "gate": true
1218
+ },
1219
+ "B": {
1220
+ "base_gain": 0.028904818209563816,
1221
+ "candidate_gain": 0.05181699390083095,
1222
+ "gain_delta": 0.022912175691267134,
1223
+ "base_target": 0.044759233792622886,
1224
+ "candidate_target": 0.07482292254765828,
1225
+ "target_delta": 0.030063688755035393,
1226
+ "base_specificity": 0.03210023045539856,
1227
+ "candidate_specificity": 0.0530113677183787,
1228
+ "specificity_delta": 0.02091113726298014,
1229
+ "gate": true
1230
+ }
1231
+ },
1232
+ "panel_detail": [
1233
+ {
1234
+ "panel_id": 0,
1235
+ "gain_delta": 0.015983013181004216,
1236
+ "candidate_target": 0.11475160717964172,
1237
+ "candidate_specificity": 0.10043948888778687,
1238
+ "target_delta_mean": 0.0326080322265625,
1239
+ "specificity_delta_mean": 0.0379844605922699,
1240
+ "point_gate": true,
1241
+ "legacy_target_ci95": [
1242
+ 0.010603100806474686,
1243
+ 0.062475830316543544
1244
+ ],
1245
+ "legacy_specificity_ci95": [
1246
+ 0.009382635354995728,
1247
+ 0.07110699862241744
1248
+ ]
1249
+ },
1250
+ {
1251
+ "panel_id": 1,
1252
+ "gain_delta": 0.013917236585406837,
1253
+ "candidate_target": 0.05650416016578674,
1254
+ "candidate_specificity": 0.031475573778152466,
1255
+ "target_delta_mean": 0.023133963346481323,
1256
+ "specificity_delta_mean": 0.01248544454574585,
1257
+ "point_gate": true,
1258
+ "legacy_target_ci95": [
1259
+ 0.010964906215667725,
1260
+ 0.03518813252449036
1261
+ ],
1262
+ "legacy_specificity_ci95": [
1263
+ 0.0020817264914512637,
1264
+ 0.02212553173303604
1265
+ ]
1266
+ },
1267
+ {
1268
+ "panel_id": 2,
1269
+ "gain_delta": 0.016885502511979743,
1270
+ "candidate_target": 0.05103442072868347,
1271
+ "candidate_specificity": 0.05271527171134949,
1272
+ "target_delta_mean": 0.0035798251628875732,
1273
+ "specificity_delta_mean": 0.01490640640258789,
1274
+ "point_gate": true,
1275
+ "legacy_target_ci95": [
1276
+ -0.023739128559827804,
1277
+ 0.026083828508853914
1278
+ ],
1279
+ "legacy_specificity_ci95": [
1280
+ -0.008255751430988311,
1281
+ 0.0381250813603401
1282
+ ]
1283
+ },
1284
+ {
1285
+ "panel_id": 3,
1286
+ "gain_delta": 0.017118117265295087,
1287
+ "candidate_target": 0.09875032305717468,
1288
+ "candidate_specificity": 0.07570299506187439,
1289
+ "target_delta_mean": 0.016419291496276855,
1290
+ "specificity_delta_mean": 0.025344550609588623,
1291
+ "point_gate": true,
1292
+ "legacy_target_ci95": [
1293
+ -0.0007519803941249847,
1294
+ 0.03528958410024642
1295
+ ],
1296
+ "legacy_specificity_ci95": [
1297
+ 0.008917194604873658,
1298
+ 0.042491755634546276
1299
+ ]
1300
+ },
1301
+ {
1302
+ "panel_id": 4,
1303
+ "gain_delta": 0.015759277051559707,
1304
+ "candidate_target": 0.04715028405189514,
1305
+ "candidate_specificity": 0.040507614612579346,
1306
+ "target_delta_mean": 0.016476601362228394,
1307
+ "specificity_delta_mean": 0.021552711725234985,
1308
+ "point_gate": true,
1309
+ "legacy_target_ci95": [
1310
+ 0.0015961535274982461,
1311
+ 0.03137609213590622
1312
+ ],
1313
+ "legacy_specificity_ci95": [
1314
+ 0.00772199034690857,
1315
+ 0.03601809665560722
1316
+ ]
1317
+ },
1318
+ {
1319
+ "panel_id": 5,
1320
+ "gain_delta": 0.023763140922858726,
1321
+ "candidate_target": 0.05421987175941467,
1322
+ "candidate_specificity": 0.01750081777572632,
1323
+ "target_delta_mean": 0.030675888061523438,
1324
+ "specificity_delta_mean": 0.014748036861419678,
1325
+ "point_gate": true,
1326
+ "legacy_target_ci95": [
1327
+ 0.011259471625089645,
1328
+ 0.05978693440556521
1329
+ ],
1330
+ "legacy_specificity_ci95": [
1331
+ 0.002710422873497009,
1332
+ 0.026434226334095
1333
+ ]
1334
+ }
1335
+ ],
1336
+ "primary_gate": true
1337
+ }
1338
+ },
1339
+ "primary_all": true,
1340
+ "replication_all": true,
1341
+ "ok": true
1342
+ },
1343
+ "deconfounded_vs_parent": {
1344
+ "label": "FORMAL_P3A_R3_CAND_VS_P2R2_DECONFOUNDED",
1345
+ "detail": {
1346
+ "ZH_ZH": {
1347
+ "mode": "ZH_ZH",
1348
+ "n": 72,
1349
+ "base_full_nll": 7.6896765563223095,
1350
+ "candidate_full_nll": 6.923334346877204,
1351
+ "full_relative_change": -0.09965857521211775,
1352
+ "candidate_own_gain": 0.032799356458643084,
1353
+ "candidate_target": 0.08275654580858019,
1354
+ "candidate_specificity": 0.07006898191240099,
1355
+ "target_delta_mean": 0.024335874451531306,
1356
+ "specificity_delta_mean": 0.023680157131618924,
1357
+ "panel_cluster_target_ci95": [
1358
+ 0.0098383625348409,
1359
+ 0.03883338636822171
1360
+ ],
1361
+ "panel_cluster_specificity_ci95": [
1362
+ 0.010125286049313016,
1363
+ 0.03754634989632501
1364
+ ],
1365
+ "point_replication_count": 6,
1366
+ "replication_ok": true,
1367
+ "families": {
1368
+ "A": {
1369
+ "base_full_nll": 7.562416156133016,
1370
+ "candidate_full_nll": 6.784773945808411,
1371
+ "full_relative_change": -0.1028298620797201,
1372
+ "candidate_own_gain": 0.017061606134190232,
1373
+ "candidate_target": 0.04163148668077257,
1374
+ "candidate_specificity": 0.04067420959472656,
1375
+ "target_delta": -0.0016484128104315848,
1376
+ "specificity_delta": 0.0065005222956339495,
1377
+ "gate": true
1378
+ },
1379
+ "B": {
1380
+ "base_full_nll": 7.816936956511603,
1381
+ "candidate_full_nll": 7.061894747945997,
1382
+ "full_relative_change": -0.09659054598574532,
1383
+ "candidate_own_gain": 0.04745204581989143,
1384
+ "candidate_target": 0.1238816049363878,
1385
+ "candidate_specificity": 0.09946375423007542,
1386
+ "target_delta": 0.050320161713494196,
1387
+ "specificity_delta": 0.0408597919676039,
1388
+ "gate": true
1389
+ }
1390
+ },
1391
+ "panel_detail": [
1392
+ {
1393
+ "panel_id": 0,
1394
+ "candidate_own_gain": 0.027379708826961204,
1395
+ "full_relative_change": -0.11519044594639936,
1396
+ "candidate_target": 0.06502056121826172,
1397
+ "candidate_specificity": 0.0492790142695109,
1398
+ "target_delta_mean": 0.0176846981048584,
1399
+ "specificity_delta_mean": 0.00758818785349528,
1400
+ "point_gate": true
1401
+ },
1402
+ {
1403
+ "panel_id": 1,
1404
+ "candidate_own_gain": 0.018909485007302742,
1405
+ "full_relative_change": -0.08677216472331792,
1406
+ "candidate_target": 0.044067819913228355,
1407
+ "candidate_specificity": 0.031806627909342446,
1408
+ "target_delta_mean": -0.0008048216501871744,
1409
+ "specificity_delta_mean": 0.0012011130650838215,
1410
+ "point_gate": true
1411
+ },
1412
+ {
1413
+ "panel_id": 2,
1414
+ "candidate_own_gain": 0.058571754530527834,
1415
+ "full_relative_change": -0.08878443333387936,
1416
+ "candidate_target": 0.1472533941268921,
1417
+ "candidate_specificity": 0.12404143810272217,
1418
+ "target_delta_mean": 0.047187209129333496,
1419
+ "specificity_delta_mean": 0.03958348433176676,
1420
+ "point_gate": true
1421
+ },
1422
+ {
1423
+ "panel_id": 3,
1424
+ "candidate_own_gain": 0.023415063885979772,
1425
+ "full_relative_change": -0.09604346664607702,
1426
+ "candidate_target": 0.09998448689778645,
1427
+ "candidate_specificity": 0.10770690441131592,
1428
+ "target_delta_mean": 0.04520722230275472,
1429
+ "specificity_delta_mean": 0.05053504308064779,
1430
+ "point_gate": true
1431
+ },
1432
+ {
1433
+ "panel_id": 4,
1434
+ "candidate_own_gain": 0.03495147717480385,
1435
+ "full_relative_change": -0.1071506209364316,
1436
+ "candidate_target": 0.09195383389790852,
1437
+ "candidate_specificity": 0.06739381949106853,
1438
+ "target_delta_mean": 0.03052675724029541,
1439
+ "specificity_delta_mean": 0.02650129795074463,
1440
+ "point_gate": true
1441
+ },
1442
+ {
1443
+ "panel_id": 5,
1444
+ "candidate_own_gain": 0.03493911413359186,
1445
+ "full_relative_change": -0.10284010152087279,
1446
+ "candidate_target": 0.04825917879740397,
1447
+ "candidate_specificity": 0.04018608729044596,
1448
+ "target_delta_mean": 0.0062141815821329755,
1449
+ "specificity_delta_mean": 0.01667181650797526,
1450
+ "point_gate": true
1451
+ }
1452
+ ],
1453
+ "primary_gate": true
1454
+ },
1455
+ "ZH_EN": {
1456
+ "mode": "ZH_EN",
1457
+ "n": 72,
1458
+ "base_full_nll": 7.812194380495283,
1459
+ "candidate_full_nll": 6.1164188418123455,
1460
+ "full_relative_change": -0.21706776048952173,
1461
+ "candidate_own_gain": 0.16828009439702632,
1462
+ "candidate_target": 0.2966230246755812,
1463
+ "candidate_specificity": 0.18221448858579,
1464
+ "target_delta_mean": 0.07847088575363159,
1465
+ "specificity_delta_mean": 0.03517365124490526,
1466
+ "panel_cluster_target_ci95": [
1467
+ 0.04405778646469116,
1468
+ 0.11288398504257202
1469
+ ],
1470
+ "panel_cluster_specificity_ci95": [
1471
+ 0.005828177846140338,
1472
+ 0.06814305649863349
1473
+ ],
1474
+ "point_replication_count": 6,
1475
+ "replication_ok": true,
1476
+ "families": {
1477
+ "A": {
1478
+ "base_full_nll": 7.774835493829515,
1479
+ "candidate_full_nll": 5.928857227166493,
1480
+ "full_relative_change": -0.23742988107312102,
1481
+ "candidate_own_gain": 0.17719632801928212,
1482
+ "candidate_target": 0.35266275538338554,
1483
+ "candidate_specificity": 0.2436350319120619,
1484
+ "target_delta": 0.09293621778488159,
1485
+ "specificity_delta": 0.05218772093454996,
1486
+ "gate": true
1487
+ },
1488
+ "B": {
1489
+ "base_full_nll": 7.849553267161052,
1490
+ "candidate_full_nll": 6.303980456458198,
1491
+ "full_relative_change": -0.19689946142143214,
1492
+ "candidate_own_gain": 0.15971627889655524,
1493
+ "candidate_target": 0.24058329396777683,
1494
+ "candidate_specificity": 0.12079394525951809,
1495
+ "target_delta": 0.06400555372238159,
1496
+ "specificity_delta": 0.01815958155526054,
1497
+ "gate": true
1498
+ }
1499
+ },
1500
+ "panel_detail": [
1501
+ {
1502
+ "panel_id": 0,
1503
+ "candidate_own_gain": 0.13825623412064272,
1504
+ "full_relative_change": -0.18475207917586287,
1505
+ "candidate_target": 0.2861340840657552,
1506
+ "candidate_specificity": 0.16125873724619547,
1507
+ "target_delta_mean": 0.09079015254974365,
1508
+ "specificity_delta_mean": 0.030432264010111492,
1509
+ "point_gate": true
1510
+ },
1511
+ {
1512
+ "panel_id": 1,
1513
+ "candidate_own_gain": 0.17652179682836944,
1514
+ "full_relative_change": -0.24605724999945147,
1515
+ "candidate_target": 0.10675748189290364,
1516
+ "candidate_specificity": 0.014587541421254476,
1517
+ "target_delta_mean": 0.016912182172139485,
1518
+ "specificity_delta_mean": -0.01420046885808309,
1519
+ "point_gate": true
1520
+ },
1521
+ {
1522
+ "panel_id": 2,
1523
+ "candidate_own_gain": 0.19361399193642537,
1524
+ "full_relative_change": -0.2134143481147252,
1525
+ "candidate_target": 0.5292520523071289,
1526
+ "candidate_specificity": 0.4001283248265584,
1527
+ "target_delta_mean": 0.11167379220326741,
1528
+ "specificity_delta_mean": 0.06389741102854411,
1529
+ "point_gate": true
1530
+ },
1531
+ {
1532
+ "panel_id": 3,
1533
+ "candidate_own_gain": 0.162816200074806,
1534
+ "full_relative_change": -0.20079180418565673,
1535
+ "candidate_target": 0.19350783030192056,
1536
+ "candidate_specificity": 0.027307510375976562,
1537
+ "target_delta_mean": 0.0726312796274821,
1538
+ "specificity_delta_mean": 0.016755064328511555,
1539
+ "point_gate": true
1540
+ },
1541
+ {
1542
+ "panel_id": 4,
1543
+ "candidate_own_gain": 0.1566196523792259,
1544
+ "full_relative_change": -0.19332720375119486,
1545
+ "candidate_target": 0.15399467945098877,
1546
+ "candidate_specificity": 0.08166877428690593,
1547
+ "target_delta_mean": 0.03355046113332113,
1548
+ "specificity_delta_mean": 0.0071335236231486005,
1549
+ "point_gate": true
1550
+ },
1551
+ {
1552
+ "panel_id": 5,
1553
+ "candidate_own_gain": 0.18298573067252796,
1554
+ "full_relative_change": -0.2656090769950541,
1555
+ "candidate_target": 0.51009202003479,
1556
+ "candidate_specificity": 0.4083360433578491,
1557
+ "target_delta_mean": 0.14526744683583578,
1558
+ "specificity_delta_mean": 0.1070241133371989,
1559
+ "point_gate": true
1560
+ }
1561
+ ],
1562
+ "primary_gate": true
1563
+ },
1564
+ "EN_ZH": {
1565
+ "mode": "EN_ZH",
1566
+ "n": 96,
1567
+ "base_full_nll": 7.621593470374743,
1568
+ "candidate_full_nll": 6.793611114223798,
1569
+ "full_relative_change": -0.10863638415894608,
1570
+ "candidate_own_gain": 0.04532365333158313,
1571
+ "candidate_target": 0.07040177782376607,
1572
+ "candidate_specificity": 0.05305696030457815,
1573
+ "target_delta_mean": 0.020482266942660015,
1574
+ "specificity_delta_mean": 0.021170268456141155,
1575
+ "panel_cluster_target_ci95": [
1576
+ 0.012375657757123312,
1577
+ 0.0276964008808136
1578
+ ],
1579
+ "panel_cluster_specificity_ci95": [
1580
+ 0.015154346823692322,
1581
+ 0.02888934810956319
1582
+ ],
1583
+ "point_replication_count": 6,
1584
+ "replication_ok": true,
1585
+ "families": {
1586
+ "A": {
1587
+ "base_full_nll": 7.668041209379832,
1588
+ "candidate_full_nll": 6.777597904205322,
1589
+ "full_relative_change": -0.11612395928249404,
1590
+ "candidate_own_gain": 0.03870944234537898,
1591
+ "candidate_target": 0.06598063309987386,
1592
+ "candidate_specificity": 0.05310255289077759,
1593
+ "target_delta": 0.010900845130284623,
1594
+ "specificity_delta": 0.021429399649302162,
1595
+ "gate": true
1596
+ },
1597
+ "B": {
1598
+ "base_full_nll": 7.575145731369655,
1599
+ "candidate_full_nll": 6.809624324242274,
1600
+ "full_relative_change": -0.10105698745269778,
1601
+ "candidate_own_gain": 0.05181699390083095,
1602
+ "candidate_target": 0.07482292254765828,
1603
+ "candidate_specificity": 0.0530113677183787,
1604
+ "target_delta": 0.030063688755035393,
1605
+ "specificity_delta": 0.02091113726298014,
1606
+ "gate": true
1607
+ }
1608
+ },
1609
+ "panel_detail": [
1610
+ {
1611
+ "panel_id": 0,
1612
+ "candidate_own_gain": 0.045550783784365424,
1613
+ "full_relative_change": -0.10261674949627236,
1614
+ "candidate_target": 0.11475160717964172,
1615
+ "candidate_specificity": 0.10043948888778687,
1616
+ "target_delta_mean": 0.0326080322265625,
1617
+ "specificity_delta_mean": 0.0379844605922699,
1618
+ "point_gate": true
1619
+ },
1620
+ {
1621
+ "panel_id": 1,
1622
+ "candidate_own_gain": 0.04479182975977206,
1623
+ "full_relative_change": -0.10096751789049366,
1624
+ "candidate_target": 0.05650416016578674,
1625
+ "candidate_specificity": 0.031475573778152466,
1626
+ "target_delta_mean": 0.023133963346481323,
1627
+ "specificity_delta_mean": 0.01248544454574585,
1628
+ "point_gate": true
1629
+ },
1630
+ {
1631
+ "panel_id": 2,
1632
+ "candidate_own_gain": 0.04117800379057025,
1633
+ "full_relative_change": -0.11477270208316427,
1634
+ "candidate_target": 0.05103442072868347,
1635
+ "candidate_specificity": 0.05271527171134949,
1636
+ "target_delta_mean": 0.0035798251628875732,
1637
+ "specificity_delta_mean": 0.01490640640258789,
1638
+ "point_gate": true
1639
+ },
1640
+ {
1641
+ "panel_id": 3,
1642
+ "candidate_own_gain": 0.0570001813477523,
1643
+ "full_relative_change": -0.12268851050449688,
1644
+ "candidate_target": 0.09875032305717468,
1645
+ "candidate_specificity": 0.07570299506187439,
1646
+ "target_delta_mean": 0.016419291496276855,
1647
+ "specificity_delta_mean": 0.025344550609588623,
1648
+ "point_gate": true
1649
+ },
1650
+ {
1651
+ "panel_id": 4,
1652
+ "candidate_own_gain": 0.033848933422363975,
1653
+ "full_relative_change": -0.1012425382284625,
1654
+ "candidate_target": 0.04715028405189514,
1655
+ "candidate_specificity": 0.040507614612579346,
1656
+ "target_delta_mean": 0.016476601362228394,
1657
+ "specificity_delta_mean": 0.021552711725234985,
1658
+ "point_gate": true
1659
+ },
1660
+ {
1661
+ "panel_id": 5,
1662
+ "candidate_own_gain": 0.04959409912949401,
1663
+ "full_relative_change": -0.10951061990321055,
1664
+ "candidate_target": 0.05421987175941467,
1665
+ "candidate_specificity": 0.01750081777572632,
1666
+ "target_delta_mean": 0.030675888061523438,
1667
+ "specificity_delta_mean": 0.014748036861419678,
1668
+ "point_gate": true
1669
+ }
1670
+ ],
1671
+ "primary_gate": true
1672
+ }
1673
+ },
1674
+ "primary_all": true,
1675
+ "replication_all": true,
1676
+ "ok": true
1677
+ },
1678
+ "q2_deconfounded": {
1679
+ "label": "FORMAL_P3A_R3_CAND_VS_P2R2_Q2_DECONFOUNDED",
1680
+ "own_gain": 0.12199967344963575,
1681
+ "base_full_nll": 8.719880680243174,
1682
+ "candidate_full_nll": 8.06357891112566,
1683
+ "full_relative_change": -0.07526499423375266,
1684
+ "candidate_target_penalty": 0.1872285356124248,
1685
+ "candidate_specificity_abs": 0.22389915585517883,
1686
+ "target_delta": 0.020418544610341982,
1687
+ "specificity_delta": 0.021574079990386963,
1688
+ "by_len_gain": {
1689
+ "1024": 0.1386586672807771,
1690
+ "1536": 0.10557065230892543,
1691
+ "2048": 0.1229383521894124
1692
+ },
1693
+ "legacy_target_attribution": 0.1671014465111034,
1694
+ "legacy_specificity_normalized": 0.19982997086226487,
1695
+ "by_len_ok": true,
1696
+ "ok": true
1697
+ },
1698
+ "q2_legacy_ratio_diagnostic": {
1699
+ "gain_retention_ok": true,
1700
+ "target_attribution_ok": false,
1701
+ "specificity_normalized_ok": true,
1702
+ "ok": false
1703
+ },
1704
+ "real_relative": {
1705
+ "general": -0.0012157218805931296,
1706
+ "en_stream": 0.002218192049361158,
1707
+ "zh": 0.012649208142147803,
1708
+ "code": 0.000972398345418651,
1709
+ "reasoning": 0.0004925456754831082,
1710
+ "long_en": 0.0016712295007618669,
1711
+ "long_zh": 0.013326061691282882,
1712
+ "code_stream": 4.6742854400550277e-05
1713
+ },
1714
+ "real_composite": 0.0037230346826467534,
1715
+ "value_ok": true,
1716
+ "q2_ok": true,
1717
+ "dtn_structure_ok": true,
1718
+ "strict_causality_ok": true,
1719
+ "structural_safety_ok": true,
1720
+ "deconfounded_memory_ok": true,
1721
+ "qualified": true,
1722
+ "causality": [
1723
+ {
1724
+ "kind": "OLD",
1725
+ "seq_len": 512,
1726
+ "max_abs_prefix_logit_diff": 0.0,
1727
+ "mean_abs_prefix_logit_diff": 0.0
1728
+ },
1729
+ {
1730
+ "kind": "OLD",
1731
+ "seq_len": 1024,
1732
+ "max_abs_prefix_logit_diff": 0.0,
1733
+ "mean_abs_prefix_logit_diff": 0.0
1734
+ },
1735
+ {
1736
+ "kind": "OLD",
1737
+ "seq_len": 1536,
1738
+ "max_abs_prefix_logit_diff": 0.0,
1739
+ "mean_abs_prefix_logit_diff": 0.0
1740
+ },
1741
+ {
1742
+ "kind": "OLD",
1743
+ "seq_len": 2048,
1744
+ "max_abs_prefix_logit_diff": 0.0,
1745
+ "mean_abs_prefix_logit_diff": 0.0
1746
+ },
1747
+ {
1748
+ "kind": "CHINESE",
1749
+ "seq_len": 512,
1750
+ "max_abs_prefix_logit_diff": 0.0,
1751
+ "mean_abs_prefix_logit_diff": 0.0
1752
+ },
1753
+ {
1754
+ "kind": "CHINESE",
1755
+ "seq_len": 1024,
1756
+ "max_abs_prefix_logit_diff": 0.0,
1757
+ "mean_abs_prefix_logit_diff": 0.0
1758
+ },
1759
+ {
1760
+ "kind": "CHINESE",
1761
+ "seq_len": 1536,
1762
+ "max_abs_prefix_logit_diff": 0.0,
1763
+ "mean_abs_prefix_logit_diff": 0.0
1764
+ },
1765
+ {
1766
+ "kind": "CHINESE",
1767
+ "seq_len": 2048,
1768
+ "max_abs_prefix_logit_diff": 0.0,
1769
+ "mean_abs_prefix_logit_diff": 0.0
1770
+ },
1771
+ {
1772
+ "kind": "MIXED",
1773
+ "seq_len": 512,
1774
+ "max_abs_prefix_logit_diff": 0.0,
1775
+ "mean_abs_prefix_logit_diff": 0.0
1776
+ },
1777
+ {
1778
+ "kind": "MIXED",
1779
+ "seq_len": 1024,
1780
+ "max_abs_prefix_logit_diff": 0.0,
1781
+ "mean_abs_prefix_logit_diff": 0.0
1782
+ },
1783
+ {
1784
+ "kind": "MIXED",
1785
+ "seq_len": 1536,
1786
+ "max_abs_prefix_logit_diff": 0.0,
1787
+ "mean_abs_prefix_logit_diff": 0.0
1788
+ },
1789
+ {
1790
+ "kind": "MIXED",
1791
+ "seq_len": 2048,
1792
+ "max_abs_prefix_logit_diff": 0.0,
1793
+ "mean_abs_prefix_logit_diff": 0.0
1794
+ }
1795
+ ]
1796
+ },
1797
+ "gates": {
1798
+ "phase2_r2_parent_verdict_exact": true,
1799
+ "phase2_r2_parent_loadback_ok": true,
1800
+ "q2_ruler_calibrated": true,
1801
+ "context_preflight_ok": true,
1802
+ "phase3a_r3_parent_valid": true,
1803
+ "branch_run": true,
1804
+ "optimizer_resumed": true,
1805
+ "value_ok": true,
1806
+ "q2_ok": true,
1807
+ "dtn_structure_ok": true,
1808
+ "strict_causality_ok": true,
1809
+ "structural_safety_ok": true,
1810
+ "deconfounded_memory_ok": true,
1811
+ "qualified": true,
1812
+ "loadback_ok": true
1813
+ },
1814
+ "sidecar_sha256": "b7218920916bee1f071eeebbd51c285145d98ba023cf2ecb8f54d2a6c57c2dbf",
1815
+ "optimizer_sha256": "5de46e986123f303a1fa5d328492f8a1516dec223ac9dfde62bd144749fff4dd",
1816
+ "loadback_maxdiff": 0.0,
1817
+ "VERDICT": "FORMAL_BILINGUAL_CONTINUATION_PHASE3A_R3_SEQ1536_H2112_QUALIFIED_PROMOTE_PHASE3B_SEQ2048_RAMP"
1818
+ }
checkpoint/formal_bilingual_continuation/phase3a_r3/20260910T135045Z/qualified/formal_phase3a_r3_state.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b7218920916bee1f071eeebbd51c285145d98ba023cf2ecb8f54d2a6c57c2dbf
3
+ size 184204534
checkpoint/formal_bilingual_continuation/phase3a_r3/20260910T135045Z/qualified/optimizer_state.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5de46e986123f303a1fa5d328492f8a1516dec223ac9dfde62bd144749fff4dd
3
+ size 368424107
checkpoint/formal_bilingual_continuation/phase3a_r3/20260910T135045Z/qualified/sha256.txt ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ formal_phase3a_r3_state.pt b7218920916bee1f071eeebbd51c285145d98ba023cf2ecb8f54d2a6c57c2dbf
2
+ optimizer_state.pt 5de46e986123f303a1fa5d328492f8a1516dec223ac9dfde62bd144749fff4dd
checkpoint/formal_bilingual_continuation/phase3a_r3/20260910T135045Z/qualified/training_contract.json ADDED
@@ -0,0 +1,85 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "stage": "formal-bilingual-continuation-phase3a_r3-seq1536-h2112",
3
+ "run_id": "20260910T135045Z",
4
+ "stable_parent": {
5
+ "formal_phase2_r2_run": "20260910T052204Z",
6
+ "sidecar_sha256": "96fd0ad71b5e8a681ba0e33e82c9485d44dea01d5a358ccc62d952af4df572c3",
7
+ "optimizer_sha256": "2af311a33963514c27f9de28f03b86a3fe91511caf62bcfe5a436d624ea65ffe",
8
+ "verdict": "FORMAL_BILINGUAL_CONTINUATION_PHASE2_R2_SEQ1024_H1344_DECONFOUNDED_Q2_QUALIFIED_PROMOTE_PHASE3_LONG_CONTEXT_RAMP"
9
+ },
10
+ "architecture_freeze": "v6-mature-20260910",
11
+ "seq_len": 1536,
12
+ "updates": 2112,
13
+ "grad_accum": 2,
14
+ "tokens_per_update": 3072,
15
+ "tail_lr": 5e-06,
16
+ "scope": "NEIGHBOR23_25",
17
+ "reader_trainable": false,
18
+ "optimizer_resume": true,
19
+ "mixture": {
20
+ "EN": 0.4,
21
+ "ZH": 0.4,
22
+ "CODE": 0.2
23
+ },
24
+ "train_ranges": {
25
+ "EN": [
26
+ "3300:4222@1536",
27
+ "4322:4629@1536",
28
+ "4729:4883@1536",
29
+ "4983:5290@1536"
30
+ ],
31
+ "ZH": [
32
+ "3360:4282@1536",
33
+ "4382:4689@1536",
34
+ "4789:4942@1536",
35
+ "5042:5350@1536"
36
+ ],
37
+ "CODE": [
38
+ "1700:2160@1536",
39
+ "2240:2394@1536",
40
+ "2474:2551@1536",
41
+ "2631:2784@1536"
42
+ ]
43
+ },
44
+ "replay": {
45
+ "interval": 8,
46
+ "events": 264,
47
+ "records": 528,
48
+ "seq_len": 1024,
49
+ "causal_weight_each": 1.0,
50
+ "teacher": "fixed qualified Formal Phase-2-R2 parent"
51
+ },
52
+ "promotion_value_contract": {
53
+ "opt_weighted_floor": 0.005,
54
+ "real_composite_floor": 0.0025,
55
+ "per_panel_regression_floor": -0.005,
56
+ "zh_gain_floor": 0.0025,
57
+ "long_zh_gain_floor": 0.0025
58
+ },
59
+ "q2_deconfounded_contract": {
60
+ "own_gain_floor": 0.005,
61
+ "full_path_regression_tolerance": 0.005,
62
+ "target_abs_floor": 0.1,
63
+ "specificity_abs_floor": 0.1,
64
+ "target_delta_floor": -0.01,
65
+ "specificity_delta_floor": -0.015,
66
+ "by_length_gain_floor": 0.005,
67
+ "legacy_target_attribution": "diagnostic_only",
68
+ "legacy_specificity_normalized": "diagnostic_only"
69
+ },
70
+ "deconfounded_global_memory_contract": {
71
+ "own_gain_floor": 0.005,
72
+ "full_path_regression_tolerance": 0.005,
73
+ "cluster_ci_floor": -0.03,
74
+ "family_abs_floor": -0.02,
75
+ "family_delta_floor": -0.03,
76
+ "min_point_replication": 4
77
+ },
78
+ "strict_causality_lengths": [
79
+ 512,
80
+ 1024,
81
+ 1536,
82
+ 2048
83
+ ],
84
+ "next_stage": "FORMAL_BILINGUAL_CONTINUATION_PHASE3B_SEQ2048_RAMP"
85
+ }