openmoji-g1-withheld-calibration-v14-dbfc543-359d06bd-9b9b1699

completed passed

Hypothesis

v13 beat identity without any held-out calibration, 0.6519 against 0.6515, but by only 0.0003, because it estimated the value head's accuracy on icons the model had trained on. There it scores 0.2751 against 0.1233 held out, so the derived threshold came out at 0.784 where the held-out accuracy implies 0.890 - too permissive, so the model edits more than pays. Estimating that accuracy on 256 train icons withheld from training should give a threshold near 0.89 and a margin closer to the 0.0065 a fitted threshold achieved, while still never touching held-out data.

Why run it

It closes the last gap between the project's first win over the trivial policy and a win that needs no asterisk at all: no sweep, no held-out calibration, and a threshold the model derives from data it never trained on.

Scope limits

Fixed-topology and geometry-only, single seed, marginal-respecting corruption at 0.35, token accuracy rather than renders. Withholding 256 icons also removes them from training, so the model sees 2,425 rather than 2,681 - a 9.6% reduction that works against this run and is disclosed rather than hidden.

Measured behaviour

Held-out loss

Table view

optimizer step	held-out loss
0	1.7456
60	1.3075
120	1.2794
180	1.2665
240	1.2555
300	1.2464
360	1.2209
420	1.1997
480	1.1843
540	1.1718
600	1.1595
660	1.1461
720	1.1376
780	1.1233
840	1.1101
900	1.1022
960	1.0932
1020	1.0862
1080	1.0772
1140	1.0720
1200	1.0660
1260	1.0607
1320	1.0617
1380	1.0585
1440	1.0535
1500	1.0497
1560	1.0458
1620	1.0483
1680	1.0430
1740	1.0468
1800	1.0412
1860	1.0391
1920	1.0386
1980	1.0386
2040	1.0403
2100	1.0407
2160	1.0364
2220	1.0360
2280	1.0392
2340	1.0411
2400	1.0364
2460	1.0385
2520	1.0424
2580	1.0408
2640	1.0397
2700	1.0440

Held-out token accuracy

aggregatechanged fieldsretained fields

Table view

optimizer step	aggregate	changed fields	retained fields
0	0.3339	0.0011	0.5118
60	0.6254	0.0012	0.9592
120	0.6437	0.0016	0.9871
180	0.6484	0.0019	0.9942
240	0.6465	0.0031	0.9906
300	0.6461	0.0034	0.9898
360	0.6358	0.0042	0.9736
420	0.6065	0.0110	0.9249
480	0.5725	0.0216	0.8671
540	0.5869	0.0213	0.8894
600	0.5828	0.0268	0.8802
660	0.5868	0.0326	0.8833
720	0.5912	0.0346	0.8889
780	0.6057	0.0362	0.9102
840	0.6034	0.0453	0.9018
900	0.6057	0.0506	0.9026
960	0.6120	0.0497	0.9127
1020	0.6071	0.0526	0.9037
1080	0.6172	0.0543	0.9182
1140	0.6101	0.0570	0.9059
1200	0.6169	0.0584	0.9156
1260	0.6143	0.0588	0.9113
1320	0.6036	0.0609	0.8938
1380	0.6019	0.0643	0.8894
1440	0.6047	0.0654	0.8932
1500	0.6071	0.0662	0.8964
1560	0.6028	0.0681	0.8887
1620	0.6044	0.0709	0.8897
1680	0.6098	0.0661	0.9005
1740	0.5953	0.0717	0.8753
1800	0.6127	0.0676	0.9042
1860	0.6086	0.0687	0.8973
1920	0.6003	0.0722	0.8827
1980	0.6055	0.0719	0.8908
2040	0.6051	0.0700	0.8913
2100	0.6049	0.0745	0.8885
2160	0.6047	0.0711	0.8901
2220	0.6049	0.0730	0.8894
2280	0.5994	0.0742	0.8803
2340	0.6038	0.0719	0.8882
2400	0.5973	0.0760	0.8761
2460	0.5983	0.0760	0.8777
2520	0.5975	0.0753	0.8767
2580	0.6041	0.0760	0.8865
2640	0.5997	0.0750	0.8803
2700	0.5950	0.0782	0.8714

Training token accuracy (per batch)

Table view

optimizer step	train accuracy
1	0.3187
2	0.3981
3	0.5018
4	0.5079
5	0.5130
6	0.5537
7	0.5227
8	0.5390
9	0.5479
10	0.5150
11	0.5021
12	0.4831
13	0.4697
14	0.4669
15	0.4741
16	0.5191
17	0.5295
18	0.5582
19	0.5428
20	0.5639
21	0.5416
22	0.5687
23	0.5795
24	0.5912
25	0.5681
26	0.5897
27	0.5475
28	0.5880
29	0.5685
30	0.5656
31	0.5759
32	0.5666
33	0.5572
34	0.5758
35	0.5803
36	0.5953
37	0.5757
38	0.5890
39	0.5978
40	0.6176
41	0.6096
42	0.6223
43	0.5844
44	0.6259
45	0.6315
46	0.6153
47	0.6145
48	0.6176
49	0.6037
50	0.6232
51	0.6156
52	0.6204
53	0.6155
54	0.6154
55	0.6114
56	0.6048
57	0.6243
58	0.6248
59	0.6284
60	0.6150
61	0.6211
62	0.6107
63	0.6257
64	0.6367
65	0.6457
66	0.6434
67	0.6212
68	0.6441
69	0.6197
70	0.6365
71	0.6446
72	0.6357
73	0.6472
74	0.6248
75	0.6403
76	0.6306
77	0.6318
78	0.6294
79	0.6009
80	0.6414
81	0.6294
82	0.6347
83	0.6205
84	0.6373
85	0.6382
86	0.6281
87	0.6294
88	0.6236
89	0.6434
90	0.6423
91	0.6403
92	0.6351
93	0.6058
94	0.6489
95	0.6462
96	0.6356
97	0.6361
98	0.6515
99	0.6487
100	0.6404
101	0.6494
102	0.6413
103	0.6472
104	0.6408
105	0.6604
106	0.6432
107	0.6414
108	0.6342
109	0.6615
110	0.6478
111	0.6485
112	0.6466
113	0.6421
114	0.6482
115	0.6451
116	0.6480
117	0.6402
118	0.6437
119	0.6420
120	0.6607
121	0.6438
122	0.6585
123	0.6383
124	0.6464
125	0.6556
126	0.6458
127	0.6416
128	0.6505
129	0.6420
130	0.6459
131	0.6346
132	0.6375
133	0.6361
134	0.6425
135	0.6218
136	0.6359
137	0.6363
138	0.6499
139	0.6431
140	0.6349
141	0.6443
142	0.6471
143	0.6435
144	0.6362
145	0.6439
146	0.6456
147	0.6497
148	0.6376
149	0.6437
150	0.6356
151	0.6362
152	0.6356
153	0.6424
154	0.6489
155	0.6495
156	0.6445
157	0.6367
158	0.6467
159	0.6479
160	0.6505
161	0.6526
162	0.6400
163	0.6472
164	0.6526
165	0.6485
166	0.6518
167	0.6473
168	0.6365
169	0.6382
170	0.6576
171	0.6582
172	0.6498
173	0.6615
174	0.6437
175	0.6479
176	0.6448
177	0.6564
178	0.6350
179	0.6392
180	0.6514
181	0.6613
182	0.6569
183	0.6450
184	0.6476
185	0.6439
186	0.6434
187	0.6375
188	0.6404
189	0.6542
190	0.6515
191	0.6479
192	0.6484
193	0.6550
194	0.6485
195	0.6526
196	0.6441
197	0.6526
198	0.6466
199	0.6517
200	0.6455
201	0.6530
202	0.6424
203	0.6497
204	0.6407
205	0.6464
206	0.6425
207	0.6347
208	0.6428
209	0.6352
210	0.6458
211	0.6385
212	0.6478
213	0.6522
214	0.6503
215	0.6423
216	0.6491
217	0.6533
218	0.6452
219	0.6464
220	0.6449
221	0.6459
222	0.6540
223	0.6598
224	0.6593
225	0.6429
226	0.6553
227	0.6454
228	0.6402
229	0.6555
230	0.6524
231	0.6517
232	0.6409
233	0.6542
234	0.6446
235	0.6452
236	0.6542
237	0.6453
238	0.6432
239	0.6407
240	0.6482
241	0.6501
242	0.6333
243	0.6331
244	0.6465
245	0.6407
246	0.6267
247	0.6487
248	0.6555
249	0.6399
250	0.6582
251	0.6634
252	0.6490
253	0.6440
254	0.6489
255	0.6413
256	0.6492
257	0.6568
258	0.6433
259	0.6444
260	0.6436
261	0.6443
262	0.6497
263	0.6570
264	0.6474
265	0.6326
266	0.6423
267	0.6427
268	0.6494
269	0.6470
270	0.6458
271	0.6500
272	0.6442
273	0.6585
274	0.6533
275	0.6680
276	0.6540
277	0.6404
278	0.6495
279	0.6453
280	0.6545
281	0.6556
282	0.6653
283	0.6401
284	0.6459
285	0.6489
286	0.6547
287	0.6501
288	0.6627
289	0.6505
290	0.6466
291	0.6682
292	0.6470
293	0.6399
294	0.6522
295	0.6425
296	0.6494
297	0.6353
298	0.6486
299	0.6418
300	0.6549
301	0.6406
302	0.6510
303	0.6508
304	0.6452
305	0.6487
306	0.6481
307	0.6602
308	0.6457
309	0.6430
310	0.6543
311	0.6567
312	0.6529
313	0.6535
314	0.6478
315	0.6434
316	0.6388
317	0.6449
318	0.6497
319	0.6403
320	0.6443
321	0.6501
322	0.6529
323	0.6477
324	0.6507
325	0.6397
326	0.6438
327	0.6396
328	0.6645
329	0.6503
330	0.6539
331	0.6540
332	0.6362
333	0.6387
334	0.6463
335	0.6408
336	0.6444
337	0.6538
338	0.6455
339	0.6383
340	0.6428
341	0.6519
342	0.6369
343	0.6495
344	0.6511
345	0.6395
346	0.6392
347	0.6510
348	0.6464
349	0.6431
350	0.6498
351	0.6426
352	0.6525
353	0.6392
354	0.6491
355	0.6404
356	0.6464
357	0.6304
358	0.6408
359	0.6277
360	0.6439
361	0.6380
362	0.6355
363	0.6331
364	0.6318
365	0.6313
366	0.6498
367	0.6252
368	0.6334
369	0.6293
370	0.6221
371	0.6380
372	0.6299
373	0.6237
374	0.6293
375	0.6386
376	0.6253
377	0.6345
378	0.6398
379	0.6282
380	0.6389
381	0.6352
382	0.6333
383	0.6229
384	0.6264
385	0.6179
386	0.6221
387	0.6132
388	0.6203
389	0.6154
390	0.6284
391	0.6287
392	0.6295
393	0.6164
394	0.6313
395	0.6173
396	0.6209
397	0.6143
398	0.6021
399	0.5948
400	0.6215
401	0.6002
402	0.6061
403	0.6223
404	0.6191
405	0.6216
406	0.6211
407	0.6217
408	0.6112
409	0.6215
410	0.6127
411	0.6118
412	0.6152
413	0.6189
414	0.6355
415	0.6152
416	0.6199
417	0.6067
418	0.6029
419	0.5894
420	0.6092
421	0.6130
422	0.6119
423	0.6155
424	0.6221
425	0.6010
426	0.6009
427	0.5849
428	0.6112
429	0.6151
430	0.6161
431	0.6332
432	0.6206
433	0.6354
434	0.6000
435	0.6245
436	0.5925
437	0.5950
438	0.5970
439	0.5993
440	0.6195
441	0.6244
442	0.6180
443	0.6335
444	0.6248
445	0.6226
446	0.6130
447	0.5960
448	0.6192
449	0.5870
450	0.6023
451	0.6191
452	0.5963
453	0.6199
454	0.6248
455	0.6201
456	0.6106
457	0.6180
458	0.6043
459	0.6096
460	0.5890
461	0.5876
462	0.6025
463	0.5905
464	0.6032
465	0.5887
466	0.5891
467	0.6239
468	0.6034
469	0.6049
470	0.6082
471	0.6035
472	0.5927
473	0.5937
474	0.6125
475	0.6033
476	0.6059
477	0.6195
478	0.6190
479	0.5981
480	0.6065
481	0.5718
482	0.5807
483	0.5917
484	0.6012
485	0.6132
486	0.6350
487	0.6256
488	0.6070
489	0.6235
490	0.5973
491	0.6035
492	0.5890
493	0.5898
494	0.5954
495	0.5867
496	0.6181
497	0.6170
498	0.6005
499	0.6143
500	0.5877
501	0.5790
502	0.5863
503	0.5997
504	0.6117
505	0.6038
506	0.6145
507	0.6010
508	0.5902
509	0.5917
510	0.6176
511	0.5875
512	0.5964
513	0.6161
514	0.6083
515	0.6266
516	0.6179
517	0.6214
518	0.6240
519	0.5882
520	0.6000
521	0.6002
522	0.5901
523	0.6273
524	0.6071
525	0.6173
526	0.6230
527	0.5970
528	0.5970
529	0.5874
530	0.5838
531	0.5995
532	0.5916
533	0.6231
534	0.5969
535	0.5974
536	0.5909
537	0.6314
538	0.6088
539	0.6117
540	0.6158
541	0.5919
542	0.6014
543	0.5982
544	0.5942
545	0.6122
546	0.6213
547	0.6088
548	0.5912
549	0.5959
550	0.6047
551	0.5707
552	0.5712
553	0.5899
554	0.5906
555	0.6199
556	0.6248
557	0.6283
558	0.6091
559	0.6188
560	0.6006
561	0.5922
562	0.5899
563	0.6159
564	0.6218
565	0.6099
566	0.6279
567	0.6181
568	0.5976
569	0.6113
570	0.6121
571	0.5978
572	0.5824
573	0.5925
574	0.6368
575	0.6122
576	0.6128
577	0.6099
578	0.5897
579	0.6001
580	0.5906
581	0.6067
582	0.6088
583	0.6219
584	0.6112
585	0.6356
586	0.5753
587	0.6323
588	0.5947
589	0.5867
590	0.5844
591	0.6121
592	0.5823
593	0.6068
594	0.5892
595	0.5958
596	0.6150
597	0.6230
598	0.6115
599	0.5932
600	0.6003
601	0.5954
602	0.6066
603	0.6083
604	0.5872
605	0.6265
606	0.6242
607	0.5972
608	0.6101
609	0.5942
610	0.6347
611	0.5905
612	0.5701
613	0.5738
614	0.6050
615	0.6144
616	0.6163
617	0.5987
618	0.6030
619	0.6237
620	0.6110
621	0.6266
622	0.6069
623	0.5956
624	0.6077
625	0.6201
626	0.6349
627	0.6071
628	0.6111
629	0.6256
630	0.5970
631	0.6175
632	0.6013
633	0.6246
634	0.5905
635	0.6011
636	0.6086
637	0.6022
638	0.6070
639	0.6075
640	0.6147
641	0.6000
642	0.6136
643	0.6139
644	0.6147
645	0.6175
646	0.6038
647	0.6008
648	0.6251
649	0.5828
650	0.6198
651	0.5988
652	0.6014
653	0.5813
654	0.6071
655	0.6240
656	0.6144
657	0.6001
658	0.6028
659	0.5800
660	0.5873
661	0.6086
662	0.5938
663	0.6267
664	0.5979
665	0.6253
666	0.6132
667	0.6167
668	0.6175
669	0.6166
670	0.5743
671	0.6173
672	0.6117
673	0.6071
674	0.5972
675	0.6243
676	0.6060
677	0.6148
678	0.6381
679	0.6154
680	0.6251
681	0.5899
682	0.5889
683	0.6217
684	0.6161
685	0.5971
686	0.6004
687	0.6095
688	0.6027
689	0.6262
690	0.5941
691	0.6281
692	0.6230
693	0.6387
694	0.6103
695	0.6089
696	0.6324
697	0.6479
698	0.6150
699	0.6046
700	0.5673
701	0.6013
702	0.5990
703	0.6160
704	0.6247
705	0.6241
706	0.6359
707	0.6381
708	0.6360
709	0.6180
710	0.6299
711	0.6073
712	0.6058
713	0.5978
714	0.6033
715	0.6263
716	0.6157
717	0.6192
718	0.6288
719	0.6127
720	0.6131
721	0.5933
722	0.6083
723	0.5876
724	0.6180
725	0.5926
726	0.6116
727	0.6188
728	0.5946
729	0.6270
730	0.6084
731	0.6143
732	0.6127
733	0.6091
734	0.6292
735	0.6186
736	0.6265
737	0.6142
738	0.6174
739	0.6220
740	0.6116
741	0.6065
742	0.6386
743	0.6141
744	0.6134
745	0.6146
746	0.5853
747	0.6247
748	0.6233
749	0.6278
750	0.6001
751	0.6208
752	0.6151
753	0.6191
754	0.6039
755	0.6228
756	0.6041
757	0.6266
758	0.6317
759	0.5942
760	0.6176
761	0.6360
762	0.6460
763	0.6125
764	0.6042
765	0.6225
766	0.6366
767	0.6212
768	0.6074
769	0.5721
770	0.6313
771	0.6075
772	0.6348
773	0.6493
774	0.6107
775	0.6309
776	0.6246
777	0.6613
778	0.6145
779	0.6030
780	0.6008
781	0.6221
782	0.6275
783	0.6171
784	0.6286
785	0.6253
786	0.6401
787	0.6025
788	0.6151
789	0.5997
790	0.6541
791	0.6198
792	0.6485
793	0.6190
794	0.6254
795	0.6125
796	0.6336
797	0.6121
798	0.5768
799	0.6285
800	0.6289
801	0.5999
802	0.6375
803	0.6186
804	0.6221
805	0.6036
806	0.6303
807	0.6288
808	0.6150
809	0.6306
810	0.6108
811	0.5962
812	0.6193
813	0.6417
814	0.5851
815	0.6327
816	0.6290
817	0.6465
818	0.6170
819	0.6387
820	0.6099
821	0.6466
822	0.6040
823	0.6330
824	0.6390
825	0.5939
826	0.6285
827	0.6332
828	0.6360
829	0.6535
830	0.6421
831	0.6272
832	0.6421
833	0.5994
834	0.6316
835	0.5992
836	0.6575
837	0.6189
838	0.6062
839	0.6117
840	0.6295
841	0.6351
842	0.6254
843	0.6193
844	0.6300
845	0.6032
846	0.6307
847	0.6393
848	0.6409
849	0.6526
850	0.6369
851	0.6027
852	0.5881
853	0.6386
854	0.6127
855	0.6488
856	0.6279
857	0.6132
858	0.6609
859	0.6433
860	0.6519
861	0.6087
862	0.6520
863	0.6404
864	0.6328
865	0.6171
866	0.6477
867	0.6470
868	0.5980
869	0.6467
870	0.6252
871	0.6257
872	0.6282
873	0.6345
874	0.6143
875	0.6246
876	0.6381
877	0.6713
878	0.6036
879	0.6403
880	0.6373
881	0.6266
882	0.6080
883	0.6376
884	0.6333
885	0.6349
886	0.6556
887	0.6422
888	0.6553
889	0.5948
890	0.6485
891	0.6035
892	0.6268
893	0.6221
894	0.6587
895	0.6087
896	0.6536
897	0.6286
898	0.6468
899	0.6306
900	0.6355
901	0.6305
902	0.6104
903	0.6282
904	0.6271
905	0.6249
906	0.6477
907	0.6183
908	0.6579
909	0.6432
910	0.6279
911	0.6365
912	0.6073
913	0.6407
914	0.6609
915	0.6052
916	0.6200
917	0.6401
918	0.6417
919	0.6478
920	0.6139
921	0.6240
922	0.6506
923	0.6325
924	0.6700
925	0.6435
926	0.6285
927	0.6322
928	0.6537
929	0.6761
930	0.6225
931	0.6172
932	0.6360
933	0.6261
934	0.6402
935	0.6442
936	0.6501
937	0.6418
938	0.6465
939	0.6269
940	0.6469
941	0.6386
942	0.6376
943	0.6084
944	0.6514
945	0.6314
946	0.6396
947	0.6386
948	0.6288
949	0.6237
950	0.6294
951	0.6647
952	0.6164
953	0.6452
954	0.6431
955	0.6230
956	0.6132
957	0.6345
958	0.6320
959	0.6460
960	0.6355
961	0.6307
962	0.6047
963	0.6162
964	0.6328
965	0.6366
966	0.6310
967	0.6333
968	0.6486
969	0.6193
970	0.6420
971	0.6375
972	0.6363
973	0.6244
974	0.6377
975	0.6426
976	0.6276
977	0.6084
978	0.6536
979	0.6194
980	0.6388
981	0.6731
982	0.6328
983	0.6795
984	0.6331
985	0.6243
986	0.6292
987	0.6147
988	0.6150
989	0.6170
990	0.6251
991	0.6127
992	0.6606
993	0.6163
994	0.6513
995	0.6498
996	0.6628
997	0.6251
998	0.6481
999	0.6526
1000	0.6448
1001	0.6650
1002	0.6144
1003	0.5961
1004	0.6017
1005	0.6512
1006	0.6336
1007	0.6314
1008	0.6238
1009	0.6523
1010	0.6370
1011	0.6207
1012	0.6371
1013	0.6536
1014	0.6408
1015	0.6328
1016	0.6105
1017	0.6204
1018	0.6584
1019	0.6381
1020	0.6353
1021	0.6544
1022	0.6255
1023	0.6256
1024	0.6367
1025	0.6392
1026	0.6297
1027	0.6428
1028	0.6131
1029	0.6468
1030	0.6296
1031	0.6322
1032	0.6600
1033	0.6421
1034	0.6343
1035	0.6282
1036	0.6314
1037	0.6335
1038	0.6397
1039	0.6653
1040	0.6354
1041	0.6250
1042	0.6314
1043	0.6379
1044	0.6273
1045	0.6703
1046	0.6286
1047	0.6574
1048	0.6487
1049	0.6122
1050	0.6625
1051	0.6145
1052	0.6345
1053	0.6089
1054	0.6450
1055	0.6193
1056	0.6498
1057	0.6206
1058	0.6483
1059	0.6413
1060	0.6593
1061	0.6621
1062	0.6137
1063	0.6337
1064	0.6303
1065	0.6595
1066	0.6169
1067	0.6027
1068	0.6457
1069	0.6430
1070	0.6317
1071	0.6369
1072	0.6130
1073	0.6601
1074	0.6222
1075	0.6641
1076	0.6627
1077	0.6410
1078	0.6515
1079	0.6081
1080	0.6809
1081	0.6615
1082	0.6313
1083	0.6362
1084	0.6510
1085	0.6320
1086	0.6546
1087	0.6537
1088	0.6347
1089	0.6790
1090	0.6148
1091	0.6579
1092	0.6318
1093	0.6667
1094	0.6276
1095	0.6689
1096	0.6421
1097	0.6572
1098	0.6271
1099	0.6501
1100	0.6277
1101	0.6139
1102	0.6640
1103	0.6607
1104	0.6230
1105	0.6634
1106	0.6298
1107	0.6171
1108	0.6202
1109	0.6367
1110	0.6667
1111	0.6084
1112	0.6461
1113	0.6341
1114	0.6153
1115	0.6374
1116	0.6547
1117	0.6124
1118	0.6345
1119	0.6453
1120	0.6600
1121	0.6343
1122	0.6591
1123	0.6203
1124	0.6692
1125	0.6238
1126	0.6619
1127	0.6496
1128	0.5746
1129	0.6491
1130	0.6109
1131	0.6532
1132	0.6676
1133	0.6574
1134	0.6414
1135	0.6452
1136	0.6283
1137	0.6464
1138	0.6202
1139	0.6583
1140	0.6158
1141	0.6126
1142	0.6425
1143	0.6591
1144	0.6364
1145	0.6455
1146	0.6486
1147	0.6773
1148	0.6166
1149	0.6356
1150	0.6476
1151	0.6618
1152	0.6616
1153	0.6414
1154	0.6168
1155	0.6258
1156	0.6362
1157	0.6491
1158	0.6490
1159	0.6258
1160	0.6068
1161	0.6527
1162	0.6481
1163	0.6368
1164	0.6259
1165	0.6575
1166	0.6453
1167	0.6303
1168	0.6334
1169	0.6577
1170	0.6548
1171	0.6141
1172	0.6789
1173	0.6147
1174	0.6251
1175	0.6180
1176	0.6512
1177	0.6264
1178	0.6407
1179	0.6344
1180	0.6541
1181	0.6172
1182	0.6193
1183	0.6582
1184	0.6228
1185	0.6117
1186	0.6372
1187	0.6252
1188	0.6370
1189	0.6591
1190	0.6476
1191	0.6847
1192	0.6023
1193	0.6577
1194	0.6379
1195	0.6347
1196	0.6273
1197	0.6845
1198	0.6340
1199	0.6566
1200	0.6418
1201	0.6551
1202	0.6573
1203	0.6645
1204	0.6417
1205	0.6188
1206	0.6430
1207	0.6379
1208	0.6494
1209	0.6516
1210	0.6174
1211	0.6550
1212	0.6589
1213	0.6344
1214	0.6327
1215	0.6310
1216	0.6712
1217	0.6610
1218	0.6160
1219	0.6232
1220	0.6413
1221	0.6325
1222	0.6642
1223	0.6055
1224	0.6105
1225	0.6798
1226	0.6380
1227	0.6732
1228	0.6629
1229	0.6358
1230	0.6442
1231	0.6557
1232	0.7008
1233	0.6177
1234	0.6404
1235	0.6548
1236	0.6473
1237	0.6506
1238	0.6669
1239	0.6317
1240	0.6718
1241	0.6544
1242	0.6189
1243	0.6412
1244	0.6557
1245	0.6425
1246	0.6212
1247	0.6746
1248	0.6543
1249	0.6552
1250	0.6485
1251	0.6314
1252	0.6386
1253	0.6262
1254	0.6697
1255	0.6305
1256	0.6438
1257	0.6570
1258	0.6349
1259	0.6372
1260	0.6477
1261	0.6549
1262	0.6627
1263	0.6313
1264	0.6513
1265	0.6337
1266	0.6466
1267	0.6581
1268	0.6484
1269	0.6401
1270	0.6473
1271	0.6293
1272	0.6530
1273	0.6328
1274	0.6489
1275	0.6435
1276	0.6363
1277	0.6428
1278	0.6563
1279	0.6386
1280	0.5926
1281	0.6657
1282	0.6214
1283	0.6313
1284	0.6902
1285	0.6335
1286	0.6696
1287	0.6427
1288	0.6407
1289	0.6409
1290	0.6248
1291	0.6411
1292	0.6302
1293	0.6287
1294	0.6057
1295	0.6846
1296	0.6480
1297	0.6420
1298	0.6796
1299	0.6602
1300	0.6295
1301	0.6495
1302	0.6519
1303	0.6516
1304	0.6800
1305	0.6303
1306	0.6120
1307	0.6171
1308	0.6498
1309	0.6311
1310	0.6563
1311	0.6412
1312	0.6607
1313	0.6362
1314	0.6532
1315	0.6659
1316	0.6375
1317	0.6663
1318	0.6563
1319	0.6196
1320	0.6459
1321	0.6670
1322	0.6583
1323	0.6334
1324	0.6727
1325	0.6196
1326	0.6409
1327	0.6293
1328	0.6613
1329	0.6256
1330	0.6194
1331	0.6250
1332	0.6596
1333	0.6313
1334	0.6459
1335	0.6520
1336	0.6207
1337	0.6414
1338	0.6295
1339	0.6274
1340	0.6377
1341	0.6447
1342	0.6661
1343	0.6651
1344	0.6287
1345	0.6404
1346	0.6405
1347	0.6285
1348	0.6723
1349	0.6436
1350	0.6506
1351	0.6440
1352	0.6342
1353	0.6655
1354	0.6469
1355	0.6851
1356	0.6144
1357	0.6277
1358	0.6166
1359	0.6521
1360	0.6528
1361	0.6496
1362	0.6409
1363	0.6740
1364	0.6548
1365	0.6018
1366	0.6706
1367	0.6406
1368	0.6790
1369	0.6155
1370	0.5865
1371	0.6304
1372	0.6424
1373	0.6562
1374	0.6450
1375	0.6130
1376	0.6457
1377	0.6625
1378	0.6640
1379	0.6614
1380	0.6443
1381	0.6493
1382	0.6525
1383	0.6778
1384	0.6790
1385	0.6334
1386	0.6321
1387	0.6578
1388	0.6230
1389	0.6709
1390	0.6631
1391	0.6350
1392	0.6831
1393	0.6477
1394	0.6557
1395	0.6642
1396	0.6565
1397	0.6276
1398	0.6844
1399	0.6534
1400	0.6667
1401	0.6312
1402	0.6461
1403	0.6219
1404	0.6175
1405	0.6661
1406	0.6542
1407	0.5909
1408	0.6671
1409	0.6613
1410	0.6213
1411	0.6115
1412	0.6434
1413	0.6734
1414	0.6326
1415	0.6534
1416	0.6279
1417	0.6182
1418	0.6615
1419	0.6671
1420	0.6022
1421	0.6552
1422	0.6403
1423	0.6680
1424	0.6490
1425	0.6474
1426	0.6549
1427	0.6580
1428	0.6307
1429	0.6777
1430	0.6656
1431	0.6154
1432	0.6703
1433	0.6418
1434	0.6556
1435	0.6688
1436	0.6757
1437	0.6585
1438	0.6659
1439	0.6247
1440	0.6474
1441	0.6095
1442	0.6695
1443	0.6234
1444	0.6173
1445	0.6382
1446	0.6522
1447	0.6630
1448	0.6366
1449	0.6689
1450	0.6726
1451	0.6316
1452	0.6375
1453	0.6262
1454	0.6883
1455	0.6736
1456	0.6744
1457	0.6038
1458	0.6024
1459	0.6437
1460	0.6657
1461	0.6440
1462	0.6286
1463	0.6352
1464	0.6607
1465	0.6634
1466	0.6687
1467	0.6340
1468	0.6562
1469	0.6551
1470	0.6363
1471	0.6427
1472	0.6451
1473	0.6574
1474	0.6513
1475	0.6597
1476	0.6302
1477	0.6208
1478	0.6301
1479	0.6343
1480	0.6341
1481	0.6395
1482	0.6465
1483	0.6566
1484	0.6069
1485	0.6288
1486	0.6820
1487	0.6454
1488	0.6147
1489	0.6530
1490	0.6114
1491	0.6358
1492	0.6905
1493	0.6478
1494	0.6973
1495	0.6011
1496	0.6460
1497	0.6466
1498	0.6424
1499	0.6298
1500	0.6788
1501	0.6247
1502	0.6725
1503	0.6268
1504	0.6675
1505	0.6571
1506	0.6652
1507	0.6535
1508	0.6456
1509	0.6543
1510	0.6427
1511	0.6397
1512	0.6506
1513	0.6587
1514	0.6444
1515	0.6740
1516	0.6540
1517	0.6202
1518	0.6398
1519	0.6868
1520	0.6861
1521	0.6255
1522	0.6007
1523	0.6724
1524	0.6249
1525	0.6611
1526	0.6242
1527	0.6162
1528	0.6874
1529	0.6337
1530	0.6993
1531	0.6867
1532	0.6468
1533	0.6550
1534	0.6403
1535	0.7156
1536	0.6484
1537	0.6332
1538	0.6308
1539	0.6640
1540	0.6733
1541	0.6724
1542	0.6482
1543	0.6723
1544	0.6653
1545	0.6440
1546	0.6639
1547	0.6684
1548	0.6530
1549	0.6358
1550	0.6878
1551	0.6308
1552	0.6694
1553	0.6665
1554	0.6317
1555	0.6461
1556	0.5912
1557	0.6774
1558	0.6501
1559	0.6199
1560	0.6824
1561	0.6336
1562	0.6279
1563	0.6201
1564	0.6650
1565	0.6551
1566	0.6051
1567	0.6560
1568	0.6307
1569	0.6244
1570	0.6412
1571	0.6680
1572	0.6291
1573	0.6047
1574	0.6358
1575	0.6639
1576	0.6461
1577	0.6417
1578	0.6548
1579	0.6508
1580	0.6617
1581	0.6641
1582	0.6644
1583	0.6119
1584	0.6915
1585	0.6299
1586	0.6633
1587	0.6993
1588	0.6565
1589	0.6646
1590	0.6417
1591	0.6313
1592	0.6466
1593	0.6279
1594	0.6488
1595	0.6413
1596	0.6463
1597	0.6145
1598	0.6868
1599	0.6531
1600	0.6551
1601	0.6894
1602	0.6341
1603	0.6248
1604	0.6342
1605	0.6326
1606	0.6771
1607	0.6900
1608	0.6371
1609	0.6105
1610	0.6166
1611	0.6767
1612	0.6407
1613	0.6496
1614	0.6442
1615	0.6391
1616	0.6634
1617	0.6669
1618	0.6772
1619	0.6292
1620	0.6543
1621	0.6674
1622	0.6266
1623	0.6256
1624	0.6875
1625	0.6693
1626	0.6380
1627	0.6883
1628	0.6285
1629	0.6343
1630	0.6403
1631	0.6551
1632	0.6424
1633	0.6367
1634	0.6316
1635	0.6646
1636	0.6010
1637	0.6608
1638	0.6506
1639	0.6388
1640	0.6503
1641	0.6220
1642	0.6324
1643	0.6259
1644	0.6764
1645	0.6902
1646	0.6755
1647	0.6226
1648	0.6824
1649	0.6558
1650	0.6134
1651	0.6417
1652	0.6816
1653	0.6531
1654	0.6556
1655	0.6415
1656	0.6588
1657	0.6341
1658	0.6849
1659	0.6485
1660	0.6246
1661	0.6438
1662	0.6554
1663	0.6220
1664	0.6555
1665	0.6544
1666	0.6804
1667	0.6715
1668	0.6145
1669	0.6489
1670	0.6362
1671	0.7033
1672	0.6444
1673	0.6106
1674	0.6332
1675	0.6594
1676	0.6767
1677	0.6711
1678	0.6204
1679	0.6537
1680	0.6622
1681	0.6851
1682	0.6848
1683	0.6691
1684	0.6474
1685	0.6616
1686	0.7093
1687	0.6994
1688	0.6337
1689	0.6270
1690	0.6673
1691	0.6415
1692	0.6569
1693	0.6583
1694	0.6574
1695	0.6731
1696	0.6596
1697	0.6694
1698	0.6406
1699	0.6562
1700	0.6509
1701	0.6907
1702	0.6561
1703	0.6874
1704	0.6377
1705	0.6476
1706	0.6464
1707	0.6332
1708	0.6262
1709	0.6941
1710	0.6185
1711	0.6602
1712	0.6549
1713	0.6348
1714	0.6177
1715	0.6413
1716	0.6657
1717	0.6474
1718	0.6744
1719	0.6451
1720	0.6158
1721	0.6512
1722	0.6620
1723	0.6545
1724	0.6457
1725	0.6361
1726	0.6723
1727	0.6402
1728	0.6620
1729	0.6488
1730	0.6783
1731	0.6290
1732	0.6950
1733	0.6684
1734	0.6236
1735	0.6370
1736	0.6822
1737	0.6583
1738	0.6601
1739	0.6951
1740	0.6478
1741	0.6651
1742	0.6000
1743	0.6448
1744	0.6538
1745	0.6332
1746	0.6613
1747	0.6263
1748	0.6401
1749	0.6566
1750	0.6693
1751	0.6438
1752	0.6810
1753	0.6578
1754	0.6336
1755	0.6455
1756	0.6271
1757	0.6614
1758	0.6735
1759	0.6885
1760	0.6125
1761	0.5967
1762	0.6543
1763	0.6406
1764	0.6376
1765	0.6420
1766	0.6293
1767	0.6661
1768	0.6550
1769	0.6679
1770	0.6494
1771	0.6775
1772	0.6533
1773	0.6532
1774	0.6357
1775	0.6677
1776	0.6709
1777	0.6674
1778	0.6426
1779	0.6414
1780	0.6313
1781	0.6525
1782	0.6294
1783	0.6579
1784	0.6410
1785	0.6557
1786	0.6552
1787	0.6530
1788	0.6296
1789	0.6707
1790	0.6409
1791	0.6294
1792	0.6434
1793	0.6241
1794	0.6261
1795	0.6571
1796	0.6543
1797	0.7164
1798	0.6188
1799	0.6460
1800	0.6644
1801	0.6355
1802	0.6503
1803	0.6991
1804	0.6426
1805	0.6662
1806	0.6497
1807	0.6567
1808	0.6484
1809	0.6646
1810	0.6817
1811	0.6185
1812	0.6468
1813	0.6446
1814	0.6685
1815	0.6491
1816	0.6496
1817	0.6752
1818	0.6596
1819	0.6843
1820	0.6274
1821	0.6720
1822	0.6676
1823	0.6810
1824	0.6101
1825	0.6229
1826	0.6638
1827	0.6429
1828	0.6768
1829	0.6328
1830	0.6072
1831	0.7149
1832	0.6501
1833	0.6703
1834	0.6962
1835	0.6470
1836	0.6578
1837	0.6458
1838	0.7085
1839	0.6566
1840	0.6406
1841	0.6162
1842	0.6751
1843	0.6763
1844	0.6534
1845	0.6505
1846	0.6569
1847	0.6924
1848	0.6424
1849	0.6927
1850	0.6464
1851	0.6816
1852	0.6253
1853	0.6831
1854	0.6610
1855	0.6739
1856	0.6582
1857	0.6562
1858	0.6577
1859	0.6038
1860	0.6956
1861	0.6691
1862	0.6269
1863	0.6966
1864	0.6427
1865	0.6414
1866	0.6120
1867	0.6620
1868	0.6775
1869	0.5969
1870	0.6721
1871	0.6419
1872	0.6277
1873	0.6469
1874	0.6798
1875	0.6014
1876	0.6544
1877	0.6333
1878	0.6920
1879	0.6313
1880	0.6789
1881	0.6536
1882	0.6834
1883	0.6572
1884	0.6819
1885	0.6774
1886	0.6128
1887	0.7079
1888	0.6258
1889	0.6438
1890	0.6953
1891	0.6842
1892	0.6547
1893	0.6705
1894	0.6087
1895	0.6556
1896	0.6216
1897	0.6780
1898	0.6319
1899	0.6396
1900	0.6217
1901	0.6862
1902	0.6510
1903	0.6772
1904	0.6903
1905	0.6458
1906	0.6516
1907	0.6484
1908	0.6455
1909	0.6754
1910	0.6830
1911	0.6793
1912	0.6180
1913	0.6305
1914	0.6627
1915	0.6577
1916	0.6580
1917	0.6528
1918	0.6071
1919	0.6763
1920	0.6754
1921	0.6780
1922	0.6387
1923	0.6592
1924	0.6754
1925	0.6513
1926	0.6437
1927	0.6894
1928	0.6711
1929	0.6303
1930	0.6886
1931	0.6278
1932	0.6351
1933	0.6406
1934	0.6694
1935	0.6382
1936	0.6382
1937	0.6403
1938	0.6914
1939	0.6320
1940	0.6611
1941	0.6563
1942	0.6373
1943	0.6306
1944	0.6511
1945	0.6465
1946	0.6224
1947	0.6897
1948	0.6708
1949	0.6935
1950	0.5929
1951	0.6804
1952	0.6576
1953	0.6460
1954	0.6421
1955	0.7019
1956	0.6235
1957	0.6850
1958	0.6279
1959	0.6602
1960	0.6707
1961	0.6849
1962	0.6436
1963	0.6357
1964	0.6471
1965	0.6647
1966	0.6461
1967	0.6710
1968	0.6380
1969	0.6962
1970	0.6648
1971	0.6288
1972	0.6479
1973	0.6414
1974	0.6873
1975	0.6708
1976	0.6242
1977	0.6393
1978	0.6499
1979	0.6590
1980	0.6531
1981	0.6347
1982	0.6404
1983	0.6865
1984	0.6794
1985	0.6912
1986	0.6846
1987	0.6590
1988	0.6732
1989	0.7015
1990	0.7034
1991	0.6172
1992	0.6340
1993	0.6743
1994	0.6267
1995	0.6787
1996	0.6638
1997	0.6796
1998	0.6730
1999	0.6696
2000	0.6490
2001	0.6745
2002	0.6692
2003	0.6651
2004	0.6434
2005	0.6617
2006	0.6687
2007	0.6555
2008	0.6673
2009	0.6449
2010	0.6429
2011	0.6413
2012	0.7043
2013	0.6399
2014	0.6831
2015	0.6626
2016	0.6151
2017	0.6744
2018	0.6390
2019	0.6833
2020	0.6572
2021	0.6404
2022	0.6547
2023	0.6270
2024	0.6567
2025	0.6662
2026	0.6790
2027	0.6450
2028	0.6510
2029	0.6524
2030	0.6729
2031	0.6574
2032	0.6592
2033	0.6861
2034	0.6078
2035	0.6849
2036	0.6749
2037	0.6433
2038	0.6419
2039	0.6840
2040	0.6470
2041	0.6747
2042	0.7136
2043	0.6474
2044	0.7114
2045	0.6358
2046	0.6357
2047	0.6630
2048	0.6478
2049	0.6454
2050	0.6312
2051	0.6448
2052	0.6540
2053	0.6940
2054	0.6417
2055	0.6734
2056	0.6814
2057	0.6617
2058	0.6481
2059	0.6470
2060	0.6629
2061	0.6783
2062	0.7138
2063	0.6385
2064	0.5995
2065	0.6505
2066	0.6631
2067	0.6431
2068	0.6532
2069	0.6424
2070	0.6669
2071	0.6518
2072	0.6533
2073	0.6513
2074	0.6700
2075	0.6636
2076	0.6434
2077	0.6389
2078	0.6648
2079	0.6630
2080	0.6803
2081	0.6510
2082	0.6645
2083	0.6446
2084	0.6627
2085	0.6331
2086	0.6752
2087	0.6322
2088	0.6704
2089	0.6467
2090	0.6441
2091	0.6401
2092	0.6520
2093	0.6695
2094	0.6336
2095	0.6663
2096	0.6486
2097	0.6442
2098	0.6565
2099	0.6764
2100	0.6903
2101	0.6630
2102	0.6452
2103	0.6460
2104	0.6504
2105	0.6303
2106	0.6950
2107	0.6592
2108	0.6726
2109	0.6676
2110	0.6236
2111	0.6969
2112	0.6761
2113	0.6886
2114	0.6280
2115	0.6811
2116	0.6249
2117	0.6857
2118	0.6434
2119	0.6708
2120	0.6701
2121	0.6612
2122	0.7003
2123	0.6159
2124	0.6683
2125	0.6850
2126	0.6827
2127	0.6230
2128	0.6182
2129	0.6621
2130	0.6716
2131	0.6651
2132	0.6472
2133	0.6218
2134	0.6997
2135	0.6402
2136	0.7127
2137	0.7006
2138	0.6762
2139	0.6719
2140	0.6611
2141	0.7246
2142	0.6656
2143	0.6402
2144	0.6288
2145	0.6835
2146	0.6624
2147	0.6860
2148	0.6566
2149	0.6597
2150	0.7077
2151	0.6509
2152	0.6883
2153	0.6491
2154	0.6939
2155	0.6175
2156	0.7007
2157	0.6624
2158	0.6830
2159	0.6483
2160	0.6797
2161	0.6364
2162	0.6433
2163	0.7005
2164	0.6695
2165	0.6318
2166	0.6900
2167	0.6577
2168	0.6645
2169	0.6314
2170	0.6749
2171	0.6751
2172	0.6183
2173	0.6741
2174	0.6302
2175	0.6170
2176	0.6583
2177	0.6989
2178	0.6061
2179	0.6531
2180	0.6522
2181	0.6758
2182	0.6608
2183	0.6877
2184	0.6214
2185	0.6842
2186	0.6611
2187	0.6849
2188	0.6679
2189	0.6028
2190	0.6770
2191	0.6548
2192	0.6580
2193	0.6822
2194	0.6685
2195	0.6716
2196	0.6667
2197	0.6634
2198	0.6640
2199	0.6343
2200	0.6677
2201	0.6679
2202	0.6158
2203	0.6591
2204	0.6775
2205	0.6683
2206	0.6559
2207	0.6690
2208	0.6915
2209	0.6343
2210	0.6431
2211	0.6669
2212	0.6729
2213	0.6914
2214	0.6813
2215	0.5945
2216	0.6071
2217	0.6425
2218	0.6711
2219	0.6510
2220	0.6442
2221	0.6179
2222	0.6736
2223	0.7018
2224	0.6937
2225	0.6320
2226	0.6712
2227	0.6719
2228	0.6412
2229	0.6525
2230	0.6828
2231	0.6916
2232	0.6056
2233	0.7162
2234	0.6392
2235	0.6333
2236	0.6458
2237	0.6765
2238	0.6299
2239	0.6444
2240	0.6434
2241	0.6815
2242	0.6240
2243	0.6606
2244	0.6750
2245	0.6324
2246	0.6297
2247	0.6609
2248	0.6217
2249	0.6427
2250	0.6914
2251	0.6683
2252	0.7272
2253	0.6246
2254	0.6776
2255	0.6350
2256	0.6704
2257	0.6493
2258	0.7190
2259	0.6272
2260	0.6841
2261	0.6485
2262	0.6635
2263	0.6756
2264	0.6929
2265	0.6703
2266	0.6310
2267	0.6336
2268	0.6638
2269	0.6653
2270	0.6708
2271	0.6505
2272	0.6932
2273	0.7071
2274	0.6514
2275	0.6521
2276	0.6619
2277	0.6886
2278	0.6878
2279	0.6472
2280	0.6572
2281	0.6629
2282	0.6623
2283	0.6964
2284	0.6304
2285	0.6406
2286	0.7100
2287	0.6679
2288	0.6980
2289	0.7052
2290	0.6757
2291	0.6804
2292	0.7023
2293	0.7173
2294	0.6282
2295	0.6251
2296	0.6814
2297	0.6667
2298	0.6782
2299	0.6719
2300	0.6380
2301	0.6964
2302	0.7074
2303	0.6543
2304	0.6764
2305	0.6553
2306	0.6672
2307	0.6224
2308	0.7070
2309	0.6592
2310	0.6803
2311	0.6706
2312	0.6475
2313	0.6643
2314	0.6241
2315	0.7088
2316	0.6496
2317	0.6810
2318	0.6842
2319	0.6272
2320	0.6515
2321	0.6325
2322	0.6728
2323	0.6697
2324	0.6597
2325	0.6640
2326	0.6069
2327	0.6557
2328	0.6824
2329	0.6681
2330	0.6507
2331	0.6737
2332	0.6527
2333	0.6705
2334	0.6821
2335	0.6545
2336	0.6730
2337	0.6568
2338	0.6614
2339	0.6988
2340	0.6766
2341	0.5985
2342	0.7010
2343	0.6448
2344	0.6803
2345	0.7037
2346	0.6535
2347	0.7374
2348	0.6379
2349	0.6633
2350	0.6570
2351	0.6424
2352	0.6576
2353	0.6502
2354	0.6440
2355	0.6337
2356	0.6956
2357	0.6542
2358	0.6824
2359	0.6959
2360	0.6931
2361	0.6253
2362	0.6592
2363	0.6718
2364	0.6865
2365	0.7161
2366	0.6444
2367	0.6120
2368	0.6238
2369	0.6870
2370	0.6403
2371	0.6674
2372	0.6370
2373	0.6725
2374	0.6551
2375	0.6398
2376	0.6884
2377	0.6573
2378	0.6903
2379	0.6742
2380	0.6477
2381	0.6577
2382	0.6863
2383	0.6717
2384	0.6336
2385	0.6962
2386	0.6306
2387	0.6489
2388	0.6452
2389	0.6877
2390	0.6474
2391	0.6706
2392	0.6496
2393	0.6749
2394	0.6517
2395	0.6700
2396	0.6769
2397	0.6197
2398	0.6468
2399	0.6353
2400	0.6155
2401	0.6449
2402	0.6749
2403	0.6924
2404	0.6709
2405	0.6566
2406	0.6472
2407	0.6710
2408	0.6427
2409	0.6967
2410	0.6632
2411	0.6827
2412	0.6687
2413	0.6423
2414	0.6828
2415	0.6624
2416	0.7027
2417	0.6265
2418	0.6732
2419	0.6400
2420	0.6609
2421	0.6533
2422	0.6747
2423	0.6445
2424	0.6984
2425	0.6837
2426	0.6462
2427	0.6857
2428	0.6614
2429	0.6987
2430	0.6245
2431	0.6007
2432	0.6729
2433	0.6637
2434	0.6699
2435	0.6531
2436	0.6393
2437	0.6498
2438	0.6822
2439	0.7045
2440	0.7037
2441	0.6969
2442	0.6635
2443	0.6849
2444	0.7031
2445	0.7175
2446	0.6145
2447	0.6463
2448	0.6955
2449	0.6638
2450	0.6964
2451	0.6692
2452	0.6459
2453	0.7104
2454	0.6642
2455	0.6941
2456	0.6549
2457	0.6783
2458	0.6400
2459	0.7214
2460	0.6561
2461	0.6842
2462	0.6268
2463	0.6833
2464	0.6425
2465	0.6531
2466	0.7103
2467	0.6904
2468	0.6164
2469	0.6826
2470	0.6578
2471	0.6433
2472	0.6516
2473	0.6616
2474	0.6987
2475	0.6311
2476	0.6799
2477	0.6503
2478	0.6291
2479	0.6641
2480	0.7040
2481	0.6178
2482	0.6621
2483	0.6518
2484	0.6929
2485	0.6598
2486	0.6815
2487	0.6437
2488	0.6949
2489	0.6621
2490	0.6936
2491	0.6917
2492	0.6255
2493	0.6916
2494	0.6563
2495	0.6917
2496	0.6716
2497	0.6860
2498	0.6714
2499	0.6827
2500	0.6449
2501	0.6717
2502	0.6128
2503	0.6805
2504	0.6427
2505	0.6414
2506	0.6521
2507	0.6664
2508	0.6818
2509	0.6416
2510	0.6970
2511	0.6876
2512	0.6635
2513	0.6576
2514	0.6724
2515	0.6816
2516	0.6988
2517	0.6963
2518	0.6306
2519	0.6051
2520	0.6543
2521	0.6802
2522	0.6742
2523	0.6239
2524	0.6503
2525	0.6801
2526	0.6800
2527	0.6968
2528	0.6723
2529	0.6670
2530	0.6932
2531	0.6448
2532	0.6507
2533	0.6839
2534	0.6873
2535	0.6521
2536	0.7000
2537	0.6516
2538	0.6497
2539	0.6662
2540	0.6924
2541	0.6491
2542	0.6585
2543	0.6593
2544	0.6935
2545	0.6247
2546	0.6490
2547	0.7078
2548	0.6438
2549	0.6349
2550	0.6796
2551	0.6151
2552	0.6286
2553	0.6959
2554	0.6716
2555	0.7228
2556	0.5965
2557	0.6619
2558	0.6513
2559	0.6592
2560	0.6481
2561	0.7358
2562	0.6549
2563	0.7021
2564	0.6540
2565	0.6711
2566	0.6993
2567	0.6789
2568	0.6778
2569	0.6425
2570	0.6508
2571	0.6561
2572	0.6603
2573	0.6781
2574	0.6293
2575	0.6993
2576	0.6805
2577	0.6884
2578	0.6443
2579	0.6823
2580	0.6983
2581	0.6911
2582	0.6325
2583	0.6121
2584	0.6790
2585	0.6456
2586	0.7004
2587	0.6409
2588	0.6068
2589	0.7327
2590	0.6778
2591	0.7184
2592	0.7026
2593	0.6721
2594	0.6809
2595	0.6699
2596	0.7352
2597	0.6592
2598	0.6412
2599	0.6499
2600	0.6771
2601	0.6901
2602	0.6867
2603	0.6506
2604	0.6995
2605	0.6709
2606	0.6402
2607	0.6814
2608	0.6985
2609	0.6608
2610	0.6163
2611	0.7162
2612	0.6552
2613	0.6931
2614	0.6699
2615	0.6549
2616	0.6561
2617	0.6118
2618	0.6962
2619	0.6744
2620	0.6473
2621	0.7246
2622	0.6320
2623	0.6464
2624	0.6496
2625	0.6985
2626	0.6700
2627	0.6115
2628	0.6759
2629	0.6184
2630	0.6548
2631	0.6823
2632	0.6768
2633	0.6425
2634	0.6191
2635	0.6556
2636	0.7042
2637	0.6694
2638	0.6676
2639	0.6860
2640	0.6630
2641	0.6626
2642	0.6863
2643	0.6748
2644	0.6261
2645	0.7129
2646	0.6379
2647	0.6744
2648	0.7344
2649	0.6491
2650	0.6757
2651	0.6704
2652	0.6530
2653	0.6566
2654	0.6543
2655	0.6649
2656	0.6664
2657	0.6588
2658	0.6280
2659	0.6979
2660	0.6754
2661	0.6828
2662	0.7016
2663	0.6292
2664	0.6576
2665	0.6743
2666	0.6689
2667	0.6576
2668	0.7319
2669	0.6744
2670	0.6153
2671	0.6160
2672	0.6650
2673	0.6620
2674	0.6696
2675	0.6549
2676	0.6761
2677	0.6496
2678	0.6888
2679	0.6912
2680	0.6537
2681	0.6991
2682	0.6802
2683	0.6545
2684	0.6429
2685	0.7056
2686	0.6948
2687	0.6502
2688	0.6982
2689	0.6443
2690	0.6394
2691	0.6638
2692	0.6940
2693	0.6516
2694	0.6565
2695	0.6456
2696	0.6823
2697	0.6329
2698	0.6805
2699	0.6750
2700	0.6170

Measured result

Verbatim from the run's summary.json.

bucket

bucket-p32-t128

bucket_icons

3359

calibration_samples

256

calibration_withheld_from_training

true

checkpoint_bytes

6463494

checkpoint_round_trip

true

checkpoint_sha256

90c15badb537045b…

config_sha256

359d06bd2530822e…

corruption

factorized_marginal_respecting_geometry

cuda_version

13.4

decision_threshold

0.818313113807048

deterministic_algorithms

true

device

cuda

eval_every

final_train

loss

0.964240550994873

step

2700

train_token_accuracy

0.6169747081712063

group_vocabulary_size

locked_path_exact

true

metrics_sha256

6697f23f15c849e0…

model_parameters

525152

schema_version

scope

dominant-bucket fixed-topology geometry pilot; not unconditional generation

selected_train_rows

color/svg/1F3F4-E0064-E0065-E0062-E0065-E007F.svg, color/svg/1F468-1F3FC-200D-1F9BC.svg, color/svg/1F561.svg, color/svg/1F469-1F3FE-200D-1F9BC.svg, color/svg/1F469-1F3FD-200D-2764-FE0F-200D-1F48B-200D-1F469-1F3FE.svg, color/svg/1F6B5-1F3FB-200D-2642-FE0F.svg … and 2419 more

selected_validation_rows

color/svg/1F994.svg, color/svg/E30A.svg, color/svg/1F6BE.svg, color/svg/1F3CA-1F3FB-200D-2642-FE0F.svg, color/svg/1F3CB-1F3FC-200D-2642-FE0F.svg, color/svg/1F93D-1F3FC.svg … and 122 more

selection

completed_steps

2700

evals_without_improvement

min_delta

0.0

objective

held_out_loss

patience_evals

selected_held_out_loss

1.0360467433929443

selected_step

2220

stopped_early

true

steps

6300

study_version

openmoji-g1-withheld-calibration-v14

subgroup_vocabulary_size

118

torch_version

2.14.0a0+4fdf77b940.nv26.08

train_value_accuracy

0.2220261207200847

validation

accuracy

0.6542691461707658

changed_accuracy

0.04232121081701456

changed_total

13941

loss

1.0360467433929443

retained_accuracy

0.9815475505428319

retained_total

26067

validation_final_step

accuracy

0.5949810037992401

changed_accuracy

0.07818664371278962

changed_total

13941

loss

1.0439848899841309

retained_accuracy

0.8713699313308014

retained_total

26067

validation_trace_sha256

4484c0d829990c73…

validation_untrained

accuracy

0.33385822835432916

changed_accuracy

0.0011476938526648016

changed_total

13941

loss

1.745607614517212

retained_accuracy

0.511796524341121

retained_total

26067

Written result

v12 was the first model here to beat the identity baseline — emit the corrupted input unchanged — but only with a decode threshold fitted on held-out icons. Its own argmax decode lost by 0.0429. Two runs removed that asterisk.

The threshold is derivable rather than arbitrary. Keeping a retained field is always right; changing one is right only if the value head re-predicts the same token, which is negligible. So changing at confidence p has expected gain p·q − (1 − p), positive exactly when

p > 1 / (1 + q)

with q the value head's accuracy on genuinely corrupted fields.

v13 estimated q on train icons and got 0.2751, against 0.1233 held out — the value head is much better on what it trained on. That gave a threshold of 0.784 where the held-out accuracy implies 0.890, so the model edited more than paid and beat identity by only 0.0003. This run estimates q on 256 icons withheld from training.

Result

	q	threshold	aggregate	identity	margin
v12, argmax decode	—	0.500	0.6086	0.6515	−0.0429
v12, threshold fitted on held-out icons	—	0.924	0.6567	0.6502	+0.0065 *
v13, q from icons it trained on	0.2751	0.784	0.6519	0.6515	+0.0003
v14, q from withheld icons	0.2220	0.818	0.6543	0.6515	+0.0027

\* fitted on held-out data, which is why it needed an asterisk.

criterion	threshold	observed	outcome
beats_identity_without_touching_held_out_data	> identity	+0.0027	pass
improves_on_the_trained_on_estimate	> 0.0003	+0.0027	pass
threshold_reflects_an_honest_estimate	0.7842 < t < 0.95	0.8183	pass
structural_safety	locked-path exact, round-trips	both true	pass
reproducibility	identical rerun	identical	pass

All four pass, and the run is handicapped: withholding 256 icons also removes them from training, so it learned from 2,425 rather than 2,681 — 9.6% less data than every run it is compared against. The margin is conservative for that reason.

Retained-token accuracy is 0.9815, the highest of any model in this project and close to identity's 1.0, while changed-token recovery stays at 0.0423. The model has learned to leave things alone unless it is nearly certain, which is exactly what the break-even rule asks of it.

What is now established

Nothing in this run is fitted on held-out data. The threshold comes from icons the model never trained on, the model selects its checkpoint on held-out loss as every run here does, and the identity comparison is a plain read of the result. The project has a denoiser that does better than doing nothing, with no caveat about how the number was obtained.

What is not

The margin is +0.0027 on token accuracy — about 0.4% relative. This is not a good model; it is a model that has stopped being worse than useless. The value head still reaches only 0.222 on withheld corrupted fields, which is why the break-even threshold sits at 0.818 and the model edits just 4% of fields.

The measured next step is unchanged and remains the task-formulation lens's: the value head predicts an exact bin on a quarter-unit metric lattice through a categorical softmax, so being close earns nothing. A distance-kernel target would give partial credit for proximity, which should lift q, which lowers the break-even threshold, which lets the model edit more of what it correctly detects. Every term in that chain is now measured.

Scope

Fixed-topology and geometry-only, single seed, marginal-respecting corruption at 0.35, token accuracy rather than renders. The project's primary/test split remains untouched. No render in this project has yet produced a recognizable icon, and this run does not change that.

State transitions

planned2026-09-21T03:30:00Z
completed2026-09-21T03:40:00Zbeats_the_identity_baseline_with_nothing_fitted_on_held_out_data

Run record

Verbatim from runs/openmoji-g1-withheld-calibration-v14-dbfc543-359d06bd-9b9b1699/run.yaml, the record committed before launch.

schema_version

run_id

openmoji-g1-withheld-calibration-v14-dbfc543-359d06bd-9b9b1699

state

completed

parent_run

openmoji-g1-derived-threshold-v13-7c73fab-1687c462-9b9b1699

hypothesis

v13 beat identity without any held-out calibration, 0.6519 against 0.6515, but by only 0.0003, because it estimated the value head's accuracy on icons the model had trained on. There it scores 0.2751 against 0.1233 held out, so the derived threshold came out at 0.784 where the held-out accuracy implies 0.890 - too permissive, so the model edits more than pays. Estimating that accuracy on 256 train icons withheld from training should give a threshold near 0.89 and a margin closer to the 0.0065 a fitted threshold achieved, while still never touching held-out data.

expected_information_gain

scope_limits

changed_factors

primary

training.calibration_samples 0 -> 256. The last 256 icons of the deterministic train selection are withheld from training and used only to estimate the value accuracy that sets the decode threshold.

disclosed_side_effect

Training data drops from 2,681 to 2,425 icons. That is a real 9.6% reduction and it works against the hypothesis, so a pass is conservative and a narrow failure should be read with it in mind.

parameters

unchanged at 525,152

held_fixed

v13 exactly otherwise.

code

git_commit

dbfc543

execution_mode

native-local

config

path

configs/learning/openmoji-g1-withheld-calibration-v14.yaml

sha256

359d06bd2530822e…

dataset

hybrid_sha256

9b9b1699677a6f97…

bucket

bucket-p32-t128

train_samples

2681

trained_on

2425

calibration_samples

256

validation_samples

128

baselines

identity

0.6515

v12_argmax

0.6086

v12_threshold_fitted_on_held_out_icons

evaluation_half

0.6567

identity_there

0.6502

margin

0.0065

v13_threshold_from_trained_on_icons

aggregate

0.6519

threshold

0.7842

0.2751

margin

0.0003

held_out_value_accuracy_implying_threshold_0.890

0.1233

predeclared_criteria

evaluated_at

the checkpoint selected by held-out loss, decoded at the derived threshold

primary

beats_identity_without_touching_held_out_data

statement

Held-out aggregate token accuracy exceeds the identity policy's, computed from this run's own totals. Nothing in this run fits anything on held-out data.

improves_on_the_trained_on_estimate

statement

The margin over identity exceeds v13's 0.0003, despite training on 9.6% fewer icons.

threshold_reflects_an_honest_estimate

statement

The derived threshold exceeds v13's 0.7842 and lies below 0.95. A withheld estimate should be lower than the trained-on one and therefore a higher, more cautious threshold.

structural_safety

statement

Locked-path exactness holds and the canonical checkpoint round-trips.

reproducibility

statement

A second complete invocation returns an identical JSON result.

reported_not_gated

the withheld value accuracy against v13's trained-on 0.2751 and the held-out 0.1233, changed and retained accuracy, which move in opposite directions by design

falsification_meaning

If the withheld estimate does not improve the margin, then the gap between the trained-on and held-out value accuracy is not what limits the threshold, and the remaining explanation is that the break-even rule itself is too crude - it assumes changing a retained field is never right. The fix would then be to estimate the full 2x2 outcome table on the withheld slice rather than the single scalar q.

outputs

local_metadata

runs/openmoji-g1-withheld-calibration-v14-dbfc543-359d06bd-9b9b1699

report_root

reports/learning/openmoji-g1-withheld-calibration-v14

durable_artifacts

/home/dev/.cache/openmoji-g1-withheld-calibration-v14-dbfc543-359d06bd-9b9b1699

planned_at

2026-09-21 03:30:00+00:00

completed_at

2026-09-21 03:40:00+00:00

result

device

cuda

trained_on_icons

2425

calibration_icons_withheld

256

withheld_value_accuracy

0.222

derived_threshold

0.8183

held_out

aggregate

0.6543

changed

0.0423

retained

0.9815

identity_baseline

0.6515

margin

0.0027

identical_rerun

true

predeclared_outcome

overall

passed

beats_identity_without_touching_held_out_data

passed

true

observed

0.6543

threshold

0.6515

improves_on_the_trained_on_estimate

passed

true

observed

0.0027

threshold

0.0003

threshold_reflects_an_honest_estimate

passed

true

observed

0.8183

bounds

0.7842, 0.95

structural_safety

passed

true

reproducibility

passed

true

conclusion

The project beats the trivial policy with nothing fitted on held-out data: the decode threshold is derived from 256 icons withheld from training, the checkpoint is selected on held-out loss as every run here does, and the identity comparison is a plain read. Retained-token accuracy is 0.9815, the highest of any model here. The run is handicapped by its own design, training on 2,425 icons rather than 2,681, so the margin is conservative.

honest_limits

The margin is +0.0027 on token accuracy, about 0.4% relative. This is not a good model; it is a model that has stopped being worse than useless. The value head reaches only 0.222 on withheld corrupted fields, which is why the break-even threshold sits at 0.818 and the model edits just 4% of fields. No render in this project has produced a recognizable icon and this run does not change that.