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

completed partially-falsified

Hypothesis

v12 beat the identity baseline, 0.6567 against 0.6502, but only with a decode threshold fitted on a disjoint half of the held-out icons; at its own argmax decode it lost, 0.6086 against 0.6515. That threshold is derivable rather than arbitrary: keeping a retained field is always right and changing one is right only if the value head re-predicts the same token, so changing pays exactly above p = 1/(1+q) in the value head's accuracy q on corrupted fields. Estimating q on TRAIN icons and decoding at that threshold should beat identity with no calibration on held-out data at all.

Why run it

It converts the project's first win over the trivial policy from one that needed a post-hoc threshold into one the model simply produces. That is the difference between a result that holds up and a result that needs an asterisk, and it costs one run.

Scope limits

Fixed-topology and geometry-only, single seed, marginal-respecting corruption at 0.35, token accuracy rather than renders. The threshold is derived from a first-order break-even argument that assumes changing a retained field is never right; that is nearly but not exactly true, so the derived threshold is principled rather than optimal.

Measured behaviour

Held-out loss

Table view

optimizer step	held-out loss
0	1.7456
60	1.3075
120	1.2794
180	1.2635
240	1.2579
300	1.2418
360	1.2186
420	1.1976
480	1.1826
540	1.1721
600	1.1582
660	1.1461
720	1.1345
780	1.1225
840	1.1049
900	1.0950
960	1.0905
1020	1.0784
1080	1.0782
1140	1.0685
1200	1.0643
1260	1.0622
1320	1.0564
1380	1.0533
1440	1.0512
1500	1.0468
1560	1.0484
1620	1.0512
1680	1.0444
1740	1.0429
1800	1.0435
1860	1.0400
1920	1.0370
1980	1.0373
2040	1.0380
2100	1.0378
2160	1.0326
2220	1.0353
2280	1.0389
2340	1.0352
2400	1.0401
2460	1.0406
2520	1.0370
2580	1.0357
2640	1.0419

Held-out token accuracy

aggregatechanged fieldsretained fields

Table view

optimizer step	aggregate	changed fields	retained fields
0	0.3341	0.0013	0.5121
60	0.6255	0.0012	0.9594
120	0.6437	0.0016	0.9871
180	0.6473	0.0027	0.9921
240	0.6482	0.0022	0.9937
300	0.6476	0.0030	0.9923
360	0.6289	0.0063	0.9619
420	0.6096	0.0143	0.9279
480	0.5906	0.0167	0.8975
540	0.5757	0.0240	0.8708
600	0.5827	0.0260	0.8804
660	0.5867	0.0346	0.8819
720	0.6004	0.0336	0.9035
780	0.5949	0.0411	0.8910
840	0.6087	0.0440	0.9107
900	0.6067	0.0464	0.9064
960	0.6029	0.0515	0.8978
1020	0.6138	0.0514	0.9146
1080	0.6070	0.0543	0.9026
1140	0.6108	0.0569	0.9070
1200	0.6086	0.0600	0.9019
1260	0.6069	0.0621	0.8983
1320	0.6081	0.0644	0.8989
1380	0.6042	0.0643	0.8930
1440	0.6055	0.0673	0.8933
1500	0.6116	0.0659	0.9035
1560	0.6080	0.0657	0.8981
1620	0.6087	0.0703	0.8967
1680	0.6020	0.0715	0.8857
1740	0.6100	0.0715	0.8979
1800	0.6058	0.0728	0.8909
1860	0.6033	0.0730	0.8869
1920	0.6100	0.0718	0.8978
1980	0.6066	0.0712	0.8929
2040	0.5992	0.0753	0.8793
2100	0.6008	0.0745	0.8823
2160	0.6086	0.0738	0.8947
2220	0.6071	0.0748	0.8917
2280	0.6006	0.0770	0.8806
2340	0.6005	0.0750	0.8815
2400	0.6021	0.0740	0.8845
2460	0.6018	0.0768	0.8826
2520	0.6029	0.0773	0.8840
2580	0.6044	0.0741	0.8880
2640	0.5991	0.0818	0.8757

Training token accuracy (per batch)

Table view

optimizer step	train accuracy
1	0.3187
2	0.3975
3	0.5010
4	0.5082
5	0.5148
6	0.5540
7	0.5224
8	0.5394
9	0.5473
10	0.5155
11	0.5012
12	0.4833
13	0.4709
14	0.4672
15	0.4734
16	0.5177
17	0.5300
18	0.5586
19	0.5445
20	0.5629
21	0.5410
22	0.5685
23	0.5791
24	0.5906
25	0.5677
26	0.5893
27	0.5475
28	0.5895
29	0.5687
30	0.5671
31	0.5765
32	0.5660
33	0.5575
34	0.5747
35	0.5800
36	0.5951
37	0.5767
38	0.5887
39	0.5978
40	0.6188
41	0.6092
42	0.6228
43	0.5846
44	0.6257
45	0.6311
46	0.6147
47	0.6138
48	0.6178
49	0.6034
50	0.6230
51	0.6152
52	0.6210
53	0.6163
54	0.6163
55	0.6118
56	0.6051
57	0.6248
58	0.6245
59	0.6274
60	0.6143
61	0.6211
62	0.6114
63	0.6262
64	0.6372
65	0.6455
66	0.6428
67	0.6226
68	0.6447
69	0.6204
70	0.6365
71	0.6446
72	0.6361
73	0.6472
74	0.6248
75	0.6399
76	0.6305
77	0.6321
78	0.6294
79	0.6015
80	0.6416
81	0.6292
82	0.6351
83	0.6209
84	0.6371
85	0.6386
86	0.6285
87	0.6286
88	0.6233
89	0.6438
90	0.6423
91	0.6414
92	0.6347
93	0.6048
94	0.6499
95	0.6464
96	0.6358
97	0.6359
98	0.6515
99	0.6491
100	0.6406
101	0.6494
102	0.6413
103	0.6472
104	0.6406
105	0.6602
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	0.6479
115	0.6456
116	0.6482
117	0.6404
118	0.6439
119	0.6418
120	0.6604
121	0.6438
122	0.6583
123	0.6383
124	0.6464
125	0.6556
126	0.6460
127	0.6414
128	0.6505
129	0.6420
130	0.6457
131	0.6346
132	0.6373
133	0.6361
134	0.6430
135	0.6218
136	0.6357
137	0.6368
138	0.6501
139	0.6434
140	0.6349
141	0.6443
142	0.6471
143	0.6435
144	0.6362
145	0.6439
146	0.6452
147	0.6495
148	0.6373
149	0.6441
150	0.6354
151	0.6454
152	0.6394
153	0.6402
154	0.6369
155	0.6645
156	0.6427
157	0.6382
158	0.6474
159	0.6401
160	0.6490
161	0.6432
162	0.6414
163	0.6487
164	0.6432
165	0.6422
166	0.6432
167	0.6506
168	0.6417
169	0.6573
170	0.6531
171	0.6518
172	0.6556
173	0.6488
174	0.6426
175	0.6489
176	0.6377
177	0.6438
178	0.6396
179	0.6562
180	0.6434
181	0.6417
182	0.6476
183	0.6464
184	0.6535
185	0.6613
186	0.6541
187	0.6443
188	0.6389
189	0.6374
190	0.6475
191	0.6467
192	0.6393
193	0.6539
194	0.6381
195	0.6528
196	0.6527
197	0.6411
198	0.6402
199	0.6429
200	0.6552
201	0.6453
202	0.6438
203	0.6591
204	0.6495
205	0.6439
206	0.6491
207	0.6530
208	0.6627
209	0.6542
210	0.6547
211	0.6369
212	0.6559
213	0.6491
214	0.6395
215	0.6512
216	0.6355
217	0.6463
218	0.6392
219	0.6558
220	0.6407
221	0.6478
222	0.6429
223	0.6524
224	0.6577
225	0.6609
226	0.6431
227	0.6381
228	0.6413
229	0.6524
230	0.6485
231	0.6343
232	0.6448
233	0.6515
234	0.6494
235	0.6339
236	0.6527
237	0.6366
238	0.6522
239	0.6582
240	0.6537
241	0.6395
242	0.6450
243	0.6617
244	0.6402
245	0.6516
246	0.6493
247	0.6386
248	0.6547
249	0.6453
250	0.6488
251	0.6569
252	0.6536
253	0.6530
254	0.6598
255	0.6522
256	0.6527
257	0.6570
258	0.6451
259	0.6629
260	0.6416
261	0.6313
262	0.6474
263	0.6548
264	0.6458
265	0.6580
266	0.6440
267	0.6545
268	0.6448
269	0.6489
270	0.6504
271	0.6365
272	0.6579
273	0.6431
274	0.6350
275	0.6283
276	0.6441
277	0.6509
278	0.6390
279	0.6597
280	0.6395
281	0.6499
282	0.6586
283	0.6427
284	0.6561
285	0.6520
286	0.6547
287	0.6445
288	0.6415
289	0.6566
290	0.6591
291	0.6352
292	0.6493
293	0.6465
294	0.6558
295	0.6481
296	0.6538
297	0.6540
298	0.6587
299	0.6468
300	0.6510
301	0.6561
302	0.6544
303	0.6421
304	0.6473
305	0.6356
306	0.6414
307	0.6411
308	0.6489
309	0.6479
310	0.6419
311	0.6359
312	0.6619
313	0.6522
314	0.6480
315	0.6393
316	0.6447
317	0.6509
318	0.6482
319	0.6468
320	0.6596
321	0.6446
322	0.6413
323	0.6371
324	0.6434
325	0.6374
326	0.6471
327	0.6476
328	0.6454
329	0.6531
330	0.6393
331	0.6389
332	0.6440
333	0.6495
334	0.6471
335	0.6406
336	0.6564
337	0.6383
338	0.6539
339	0.6480
340	0.6379
341	0.6412
342	0.6358
343	0.6405
344	0.6461
345	0.6478
346	0.6421
347	0.6486
348	0.6305
349	0.6499
350	0.6558
351	0.6528
352	0.6479
353	0.6370
354	0.6538
355	0.6460
356	0.6421
357	0.6506
358	0.6341
359	0.6221
360	0.6362
361	0.6373
362	0.6279
363	0.6372
364	0.6228
365	0.6354
366	0.6364
367	0.6482
368	0.6297
369	0.6328
370	0.6286
371	0.6480
372	0.6228
373	0.6278
374	0.6237
375	0.6256
376	0.6301
377	0.6346
378	0.6193
379	0.6316
380	0.6099
381	0.6068
382	0.6230
383	0.6216
384	0.6298
385	0.6127
386	0.6176
387	0.6186
388	0.6223
389	0.6090
390	0.6100
391	0.6097
392	0.6217
393	0.5968
394	0.6377
395	0.6175
396	0.6363
397	0.6287
398	0.6283
399	0.6195
400	0.6139
401	0.6268
402	0.6226
403	0.6290
404	0.6285
405	0.6164
406	0.6163
407	0.6118
408	0.6125
409	0.6154
410	0.6002
411	0.6048
412	0.6168
413	0.6171
414	0.6079
415	0.6112
416	0.6178
417	0.6119
418	0.6261
419	0.6057
420	0.6187
421	0.6151
422	0.6278
423	0.6094
424	0.6299
425	0.6256
426	0.6253
427	0.6169
428	0.6020
429	0.5979
430	0.5960
431	0.5924
432	0.6372
433	0.6148
434	0.6149
435	0.6209
436	0.6154
437	0.6260
438	0.6062
439	0.6065
440	0.6135
441	0.5979
442	0.6021
443	0.6060
444	0.6230
445	0.6142
446	0.6177
447	0.6096
448	0.5968
449	0.5864
450	0.5799
451	0.5773
452	0.5916
453	0.6036
454	0.6119
455	0.6306
456	0.6277
457	0.6202
458	0.6159
459	0.5774
460	0.5975
461	0.5778
462	0.6000
463	0.6188
464	0.6116
465	0.6297
466	0.6024
467	0.6014
468	0.6048
469	0.6019
470	0.6006
471	0.6162
472	0.6042
473	0.6358
474	0.5887
475	0.6237
476	0.6146
477	0.5970
478	0.6002
479	0.5698
480	0.6034
481	0.5902
482	0.6255
483	0.6316
484	0.6263
485	0.6140
486	0.6225
487	0.6309
488	0.5935
489	0.5972
490	0.5950
491	0.5924
492	0.5808
493	0.5930
494	0.5970
495	0.6084
496	0.5943
497	0.5976
498	0.5854
499	0.5990
500	0.6236
501	0.6017
502	0.6078
503	0.6062
504	0.6080
505	0.6012
506	0.5905
507	0.5975
508	0.5867
509	0.5812
510	0.5994
511	0.5950
512	0.6279
513	0.6043
514	0.5731
515	0.6248
516	0.5985
517	0.6142
518	0.6125
519	0.6017
520	0.6160
521	0.5882
522	0.6310
523	0.6176
524	0.5982
525	0.6037
526	0.6257
527	0.5895
528	0.6281
529	0.6040
530	0.6089
531	0.6031
532	0.6094
533	0.6096
534	0.6259
535	0.6206
536	0.5942
537	0.6204
538	0.5971
539	0.5931
540	0.5704
541	0.5971
542	0.5949
543	0.5880
544	0.6356
545	0.6086
546	0.5814
547	0.6029
548	0.5568
549	0.5951
550	0.5802
551	0.5887
552	0.6059
553	0.5799
554	0.6116
555	0.5823
556	0.5639
557	0.5957
558	0.5804
559	0.6037
560	0.5885
561	0.6264
562	0.6121
563	0.5974
564	0.6068
565	0.6071
566	0.6077
567	0.5761
568	0.5864
569	0.6110
570	0.5885
571	0.6160
572	0.6261
573	0.6210
574	0.6340
575	0.6049
576	0.6181
577	0.5893
578	0.5929
579	0.6028
580	0.5836
581	0.6164
582	0.5992
583	0.6112
584	0.5997
585	0.6224
586	0.6154
587	0.6056
588	0.6256
589	0.5954
590	0.6124
591	0.6060
592	0.6036
593	0.6224
594	0.6218
595	0.6011
596	0.5865
597	0.5751
598	0.5917
599	0.5926
600	0.6026
601	0.6059
602	0.6019
603	0.6289
604	0.6265
605	0.6387
606	0.6089
607	0.6140
608	0.6073
609	0.5928
610	0.5834
611	0.6103
612	0.6156
613	0.5856
614	0.6488
615	0.6192
616	0.6167
617	0.6061
618	0.6135
619	0.5863
620	0.5829
621	0.5713
622	0.6145
623	0.5789
624	0.6035
625	0.6126
626	0.6164
627	0.6033
628	0.6007
629	0.6077
630	0.6221
631	0.6338
632	0.6112
633	0.6345
634	0.5782
635	0.6248
636	0.6035
637	0.5975
638	0.6056
639	0.6244
640	0.5866
641	0.6189
642	0.5988
643	0.6139
644	0.6019
645	0.6215
646	0.5959
647	0.6066
648	0.6079
649	0.6176
650	0.6197
651	0.6184
652	0.6017
653	0.6434
654	0.6197
655	0.6214
656	0.6129
657	0.5946
658	0.6082
659	0.5767
660	0.5724
661	0.5953
662	0.6136
663	0.5883
664	0.6037
665	0.6111
666	0.6065
667	0.6216
668	0.5822
669	0.5732
670	0.6155
671	0.6085
672	0.6132
673	0.6108
674	0.6643
675	0.6029
676	0.6062
677	0.6011
678	0.6050
679	0.6082
680	0.6035
681	0.5751
682	0.5971
683	0.6140
684	0.6268
685	0.6342
686	0.6134
687	0.6083
688	0.6113
689	0.6436
690	0.6321
691	0.5974
692	0.6167
693	0.6174
694	0.6095
695	0.6111
696	0.6041
697	0.6053
698	0.6054
699	0.6039
700	0.6029
701	0.6138
702	0.6318
703	0.6144
704	0.6425
705	0.6088
706	0.6316
707	0.6119
708	0.5948
709	0.6119
710	0.6073
711	0.5933
712	0.6342
713	0.5955
714	0.6231
715	0.6329
716	0.6144
717	0.6051
718	0.5911
719	0.6262
720	0.6086
721	0.6211
722	0.6238
723	0.5877
724	0.6077
725	0.6118
726	0.6233
727	0.6224
728	0.6107
729	0.6328
730	0.6152
731	0.6107
732	0.6020
733	0.6408
734	0.5831
735	0.6187
736	0.6307
737	0.6022
738	0.6126
739	0.6401
740	0.6089
741	0.6325
742	0.6543
743	0.6064
744	0.6226
745	0.5834
746	0.6049
747	0.6228
748	0.6249
749	0.6022
750	0.5973
751	0.6305
752	0.6139
753	0.6291
754	0.6096
755	0.6202
756	0.6309
757	0.6363
758	0.5979
759	0.6247
760	0.6292
761	0.6634
762	0.6403
763	0.6162
764	0.5941
765	0.6131
766	0.6199
767	0.6299
768	0.6147
769	0.6174
770	0.6201
771	0.6204
772	0.6152
773	0.6261
774	0.6413
775	0.6156
776	0.6084
777	0.6396
778	0.6341
779	0.6412
780	0.6127
781	0.6147
782	0.6111
783	0.6014
784	0.6307
785	0.5992
786	0.6285
787	0.6168
788	0.6432
789	0.6119
790	0.6207
791	0.6190
792	0.6128
793	0.6036
794	0.5991
795	0.6188
796	0.6127
797	0.6187
798	0.6381
799	0.6386
800	0.6482
801	0.6160
802	0.6184
803	0.6284
804	0.6187
805	0.6014
806	0.6305
807	0.6182
808	0.6374
809	0.6193
810	0.6116
811	0.6295
812	0.6232
813	0.6353
814	0.5988
815	0.6306
816	0.6000
817	0.6218
818	0.6136
819	0.6234
820	0.6324
821	0.6470
822	0.6395
823	0.6493
824	0.6141
825	0.6058
826	0.6306
827	0.5757
828	0.5985
829	0.6022
830	0.6185
831	0.6070
832	0.6316
833	0.6230
834	0.6283
835	0.6459
836	0.6016
837	0.5942
838	0.6242
839	0.6001
840	0.6262
841	0.6436
842	0.6420
843	0.6194
844	0.6209
845	0.6405
846	0.6250
847	0.6408
848	0.6380
849	0.5910
850	0.6709
851	0.6200
852	0.6428
853	0.6508
854	0.6296
855	0.6399
856	0.6396
857	0.6521
858	0.6241
859	0.6314
860	0.6034
861	0.6301
862	0.6085
863	0.6350
864	0.6321
865	0.6264
866	0.6401
867	0.6331
868	0.6476
869	0.6262
870	0.6527
871	0.6164
872	0.6593
873	0.6257
874	0.6337
875	0.6312
876	0.6262
877	0.6204
878	0.6000
879	0.6550
880	0.6388
881	0.6106
882	0.6620
883	0.6315
884	0.6296
885	0.5950
886	0.6404
887	0.6413
888	0.6148
889	0.6240
890	0.6165
891	0.6069
892	0.6293
893	0.6535
894	0.5935
895	0.6432
896	0.6274
897	0.6383
898	0.6292
899	0.6457
900	0.6303
901	0.6522
902	0.6475
903	0.6346
904	0.6486
905	0.6130
906	0.6472
907	0.6331
908	0.6355
909	0.6442
910	0.6513
911	0.6117
912	0.6339
913	0.5900
914	0.6263
915	0.6110
916	0.6319
917	0.6318
918	0.6206
919	0.6433
920	0.6337
921	0.6516
922	0.6347
923	0.6434
924	0.6631
925	0.6145
926	0.6113
927	0.6173
928	0.6516
929	0.6578
930	0.6211
931	0.6152
932	0.6064
933	0.6381
934	0.6276
935	0.6370
936	0.6261
937	0.6080
938	0.6672
939	0.6452
940	0.6571
941	0.6064
942	0.6604
943	0.6413
944	0.6284
945	0.6203
946	0.6536
947	0.6344
948	0.6076
949	0.6567
950	0.6309
951	0.6179
952	0.6319
953	0.6422
954	0.6231
955	0.6339
956	0.6217
957	0.6381
958	0.5911
959	0.6316
960	0.6216
961	0.6088
962	0.6220
963	0.6409
964	0.6271
965	0.6403
966	0.6576
967	0.6517
968	0.6693
969	0.5898
970	0.6471
971	0.6071
972	0.6305
973	0.6181
974	0.6705
975	0.6124
976	0.6423
977	0.6101
978	0.6477
979	0.6445
980	0.6512
981	0.6404
982	0.6211
983	0.6180
984	0.6528
985	0.6389
986	0.6573
987	0.6191
988	0.6342
989	0.6530
990	0.6009
991	0.6391
992	0.6084
993	0.6491
994	0.6144
995	0.6149
996	0.6348
997	0.6356
998	0.6054
999	0.6282
1000	0.6296
1001	0.6118
1002	0.6398
1003	0.6174
1004	0.5968
1005	0.6321
1006	0.6187
1007	0.6431
1008	0.6253
1009	0.6577
1010	0.6548
1011	0.6019
1012	0.6266
1013	0.6383
1014	0.6282
1015	0.6434
1016	0.6129
1017	0.6140
1018	0.6536
1019	0.6303
1020	0.6708
1021	0.6469
1022	0.6342
1023	0.6403
1024	0.6674
1025	0.6837
1026	0.6232
1027	0.6258
1028	0.6575
1029	0.6287
1030	0.6368
1031	0.6424
1032	0.6464
1033	0.6369
1034	0.6481
1035	0.6285
1036	0.6460
1037	0.6357
1038	0.6499
1039	0.6212
1040	0.6493
1041	0.6464
1042	0.6467
1043	0.6397
1044	0.6363
1045	0.6265
1046	0.6231
1047	0.6586
1048	0.6171
1049	0.6529
1050	0.6461
1051	0.6259
1052	0.6286
1053	0.6187
1054	0.6336
1055	0.6412
1056	0.6199
1057	0.6305
1058	0.6029
1059	0.6239
1060	0.6354
1061	0.6497
1062	0.6486
1063	0.6251
1064	0.6439
1065	0.6247
1066	0.6249
1067	0.6344
1068	0.6393
1069	0.6250
1070	0.6518
1071	0.6331
1072	0.6048
1073	0.6205
1074	0.6610
1075	0.6415
1076	0.6501
1077	0.6699
1078	0.6310
1079	0.6909
1080	0.6159
1081	0.6151
1082	0.6435
1083	0.6360
1084	0.6214
1085	0.6357
1086	0.6532
1087	0.6327
1088	0.6802
1089	0.6352
1090	0.6540
1091	0.6532
1092	0.6441
1093	0.6071
1094	0.6508
1095	0.6476
1096	0.6323
1097	0.6915
1098	0.6248
1099	0.6063
1100	0.6190
1101	0.6555
1102	0.6248
1103	0.6406
1104	0.6282
1105	0.6391
1106	0.6375
1107	0.6213
1108	0.6414
1109	0.6538
1110	0.6310
1111	0.6381
1112	0.6257
1113	0.6558
1114	0.6620
1115	0.6587
1116	0.6405
1117	0.6633
1118	0.6293
1119	0.6454
1120	0.6393
1121	0.6394
1122	0.6217
1123	0.6275
1124	0.6281
1125	0.6324
1126	0.6167
1127	0.6388
1128	0.6274
1129	0.6262
1130	0.6470
1131	0.6177
1132	0.6267
1133	0.6379
1134	0.6635
1135	0.6611
1136	0.6569
1137	0.6321
1138	0.6018
1139	0.6414
1140	0.6287
1141	0.6584
1142	0.6339
1143	0.6463
1144	0.6557
1145	0.6154
1146	0.6625
1147	0.6400
1148	0.6423
1149	0.6100
1150	0.6370
1151	0.6326
1152	0.6449
1153	0.6245
1154	0.6421
1155	0.6284
1156	0.6562
1157	0.6442
1158	0.6271
1159	0.6304
1160	0.6053
1161	0.6348
1162	0.6251
1163	0.6098
1164	0.6377
1165	0.6338
1166	0.6145
1167	0.6410
1168	0.6225
1169	0.6353
1170	0.6681
1171	0.6231
1172	0.6023
1173	0.6641
1174	0.5991
1175	0.6333
1176	0.6413
1177	0.6606
1178	0.6187
1179	0.6063
1180	0.6321
1181	0.6444
1182	0.6587
1183	0.6185
1184	0.6059
1185	0.6681
1186	0.6353
1187	0.6807
1188	0.6815
1189	0.6401
1190	0.6384
1191	0.6044
1192	0.6854
1193	0.6533
1194	0.6149
1195	0.6331
1196	0.6635
1197	0.6400
1198	0.6703
1199	0.6500
1200	0.6297
1201	0.6792
1202	0.6185
1203	0.6522
1204	0.6515
1205	0.6627
1206	0.6238
1207	0.6703
1208	0.6498
1209	0.6727
1210	0.6074
1211	0.6595
1212	0.6251
1213	0.6305
1214	0.6679
1215	0.6591
1216	0.6139
1217	0.6634
1218	0.6469
1219	0.6395
1220	0.6167
1221	0.6500
1222	0.6663
1223	0.6162
1224	0.6488
1225	0.6378
1226	0.5895
1227	0.6458
1228	0.6719
1229	0.5924
1230	0.6406
1231	0.6225
1232	0.6521
1233	0.6368
1234	0.6639
1235	0.6242
1236	0.6688
1237	0.6406
1238	0.6782
1239	0.6636
1240	0.5962
1241	0.6752
1242	0.6380
1243	0.6626
1244	0.6591
1245	0.6572
1246	0.6503
1247	0.6395
1248	0.6238
1249	0.6547
1250	0.6102
1251	0.6538
1252	0.5981
1253	0.6295
1254	0.6291
1255	0.6391
1256	0.6493
1257	0.6376
1258	0.6397
1259	0.6554
1260	0.6213
1261	0.6304
1262	0.6436
1263	0.6527
1264	0.6682
1265	0.6676
1266	0.5982
1267	0.6022
1268	0.6319
1269	0.6505
1270	0.6475
1271	0.6219
1272	0.6114
1273	0.6422
1274	0.6704
1275	0.6423
1276	0.6437
1277	0.6470
1278	0.6538
1279	0.6262
1280	0.6237
1281	0.6492
1282	0.6603
1283	0.6115
1284	0.6843
1285	0.6089
1286	0.6172
1287	0.6177
1288	0.6568
1289	0.6148
1290	0.6370
1291	0.6307
1292	0.6565
1293	0.5972
1294	0.6261
1295	0.6702
1296	0.6233
1297	0.6275
1298	0.6305
1299	0.6175
1300	0.6265
1301	0.6623
1302	0.6368
1303	0.6852
1304	0.6059
1305	0.6509
1306	0.6334
1307	0.6354
1308	0.6416
1309	0.6835
1310	0.6304
1311	0.6642
1312	0.6210
1313	0.6586
1314	0.6524
1315	0.6745
1316	0.6530
1317	0.6239
1318	0.6439
1319	0.6288
1320	0.6474
1321	0.6370
1322	0.6359
1323	0.6667
1324	0.6359
1325	0.6335
1326	0.6583
1327	0.6320
1328	0.6358
1329	0.6166
1330	0.6165
1331	0.6325
1332	0.6331
1333	0.6270
1334	0.6242
1335	0.6465
1336	0.6169
1337	0.6511
1338	0.6607
1339	0.6119
1340	0.6616
1341	0.6448
1342	0.6458
1343	0.6273
1344	0.6625
1345	0.6588
1346	0.6175
1347	0.6182
1348	0.6458
1349	0.6269
1350	0.6698
1351	0.6300
1352	0.6171
1353	0.6907
1354	0.6511
1355	0.6597
1356	0.6844
1357	0.6340
1358	0.6603
1359	0.6416
1360	0.7106
1361	0.6172
1362	0.6216
1363	0.6523
1364	0.6397
1365	0.6731
1366	0.6717
1367	0.6295
1368	0.6556
1369	0.6726
1370	0.6225
1371	0.6603
1372	0.6609
1373	0.6411
1374	0.6189
1375	0.6871
1376	0.6545
1377	0.6621
1378	0.6444
1379	0.6337
1380	0.6390
1381	0.6019
1382	0.6744
1383	0.6376
1384	0.6326
1385	0.6517
1386	0.6236
1387	0.6296
1388	0.6328
1389	0.6427
1390	0.6572
1391	0.6204
1392	0.6528
1393	0.6183
1394	0.6434
1395	0.6448
1396	0.6552
1397	0.6218
1398	0.6426
1399	0.6256
1400	0.6513
1401	0.6382
1402	0.6537
1403	0.6601
1404	0.6468
1405	0.6602
1406	0.6652
1407	0.6586
1408	0.6036
1409	0.6968
1410	0.6359
1411	0.6451
1412	0.6762
1413	0.6508
1414	0.6800
1415	0.6537
1416	0.6404
1417	0.6373
1418	0.6379
1419	0.6340
1420	0.6365
1421	0.6449
1422	0.6109
1423	0.6875
1424	0.6294
1425	0.6450
1426	0.6670
1427	0.6556
1428	0.6137
1429	0.6345
1430	0.6485
1431	0.6454
1432	0.6988
1433	0.6297
1434	0.6025
1435	0.6126
1436	0.6487
1437	0.6479
1438	0.6478
1439	0.6329
1440	0.6465
1441	0.6235
1442	0.6378
1443	0.6674
1444	0.6512
1445	0.6667
1446	0.6441
1447	0.6070
1448	0.6598
1449	0.6658
1450	0.6702
1451	0.6246
1452	0.6703
1453	0.6304
1454	0.6333
1455	0.6244
1456	0.6500
1457	0.6228
1458	0.6329
1459	0.6311
1460	0.6567
1461	0.6390
1462	0.6374
1463	0.6575
1464	0.6207
1465	0.6442
1466	0.6355
1467	0.6223
1468	0.6523
1469	0.6536
1470	0.6806
1471	0.6507
1472	0.6259
1473	0.6549
1474	0.6376
1475	0.6351
1476	0.6702
1477	0.6473
1478	0.6525
1479	0.6542
1480	0.6398
1481	0.6559
1482	0.6519
1483	0.6907
1484	0.6282
1485	0.6515
1486	0.6262
1487	0.6500
1488	0.6440
1489	0.6455
1490	0.6167
1491	0.6900
1492	0.6305
1493	0.6259
1494	0.6529
1495	0.6213
1496	0.6454
1497	0.6174
1498	0.6321
1499	0.6562
1500	0.6348
1501	0.6252
1502	0.6592
1503	0.6144
1504	0.6387
1505	0.6568
1506	0.6307
1507	0.6259
1508	0.6654
1509	0.6166
1510	0.6611
1511	0.6464
1512	0.6837
1513	0.6255
1514	0.6058
1515	0.6360
1516	0.6803
1517	0.6539
1518	0.6632
1519	0.5947
1520	0.6563
1521	0.6634
1522	0.6775
1523	0.6736
1524	0.6578
1525	0.6522
1526	0.6525
1527	0.6938
1528	0.7052
1529	0.6336
1530	0.6423
1531	0.6732
1532	0.6330
1533	0.6750
1534	0.6609
1535	0.6399
1536	0.6706
1537	0.6532
1538	0.6543
1539	0.6536
1540	0.6463
1541	0.6257
1542	0.6915
1543	0.6455
1544	0.6556
1545	0.6447
1546	0.6423
1547	0.6126
1548	0.6495
1549	0.6736
1550	0.6547
1551	0.6029
1552	0.6760
1553	0.6666
1554	0.6302
1555	0.6261
1556	0.6412
1557	0.6710
1558	0.6180
1559	0.6480
1560	0.6326
1561	0.5989
1562	0.6438
1563	0.6624
1564	0.6007
1565	0.6477
1566	0.6349
1567	0.6736
1568	0.6511
1569	0.6508
1570	0.6602
1571	0.6721
1572	0.6384
1573	0.6774
1574	0.6603
1575	0.5876
1576	0.6612
1577	0.6349
1578	0.6462
1579	0.6709
1580	0.6694
1581	0.6642
1582	0.6871
1583	0.6339
1584	0.6465
1585	0.6300
1586	0.6564
1587	0.6033
1588	0.6379
1589	0.6384
1590	0.6470
1591	0.6604
1592	0.6536
1593	0.6741
1594	0.6758
1595	0.6553
1596	0.6317
1597	0.6050
1598	0.6842
1599	0.6640
1600	0.6887
1601	0.6230
1602	0.6052
1603	0.6370
1604	0.6675
1605	0.6549
1606	0.6160
1607	0.6357
1608	0.6571
1609	0.6655
1610	0.6657
1611	0.6153
1612	0.6795
1613	0.6546
1614	0.6222
1615	0.6293
1616	0.6845
1617	0.6721
1618	0.6356
1619	0.6670
1620	0.6555
1621	0.6233
1622	0.6370
1623	0.6522
1624	0.6395
1625	0.6422
1626	0.6409
1627	0.6666
1628	0.6132
1629	0.6413
1630	0.6680
1631	0.6340
1632	0.6299
1633	0.6495
1634	0.6343
1635	0.6383
1636	0.6859
1637	0.6452
1638	0.6851
1639	0.5888
1640	0.6619
1641	0.6443
1642	0.6493
1643	0.6298
1644	0.6819
1645	0.6150
1646	0.6924
1647	0.6338
1648	0.6576
1649	0.6401
1650	0.6729
1651	0.6650
1652	0.6216
1653	0.6512
1654	0.6440
1655	0.6662
1656	0.6537
1657	0.6478
1658	0.6610
1659	0.6756
1660	0.6537
1661	0.6595
1662	0.6220
1663	0.6306
1664	0.6432
1665	0.6059
1666	0.6458
1667	0.6357
1668	0.6203
1669	0.5936
1670	0.6592
1671	0.6299
1672	0.6428
1673	0.6659
1674	0.6329
1675	0.6235
1676	0.6575
1677	0.6143
1678	0.6567
1679	0.6746
1680	0.6684
1681	0.6091
1682	0.6116
1683	0.6575
1684	0.6151
1685	0.6655
1686	0.6322
1687	0.6005
1688	0.7009
1689	0.6365
1690	0.6902
1691	0.6817
1692	0.6431
1693	0.6581
1694	0.6551
1695	0.7011
1696	0.6457
1697	0.6350
1698	0.6393
1699	0.6742
1700	0.6619
1701	0.6679
1702	0.6426
1703	0.6782
1704	0.6629
1705	0.6492
1706	0.6691
1707	0.6695
1708	0.6561
1709	0.6115
1710	0.6882
1711	0.6542
1712	0.6683
1713	0.6521
1714	0.6369
1715	0.6410
1716	0.6108
1717	0.6797
1718	0.6709
1719	0.6090
1720	0.6888
1721	0.6282
1722	0.6273
1723	0.6278
1724	0.6671
1725	0.6574
1726	0.6158
1727	0.6718
1728	0.6283
1729	0.6379
1730	0.6516
1731	0.6704
1732	0.6454
1733	0.6273
1734	0.6512
1735	0.6684
1736	0.6586
1737	0.6552
1738	0.6560
1739	0.6587
1740	0.6535
1741	0.6571
1742	0.6766
1743	0.6096
1744	0.6972
1745	0.6255
1746	0.6603
1747	0.6923
1748	0.6340
1749	0.6737
1750	0.6517
1751	0.6376
1752	0.6530
1753	0.6226
1754	0.6353
1755	0.6456
1756	0.6303
1757	0.6281
1758	0.6785
1759	0.6545
1760	0.6635
1761	0.6898
1762	0.6445
1763	0.6455
1764	0.6552
1765	0.6401
1766	0.6751
1767	0.6934
1768	0.6642
1769	0.5936
1770	0.6171
1771	0.6535
1772	0.6323
1773	0.6630
1774	0.6234
1775	0.6525
1776	0.6834
1777	0.6560
1778	0.6752
1779	0.6464
1780	0.6590
1781	0.6661
1782	0.6402
1783	0.6239
1784	0.6941
1785	0.6515
1786	0.6385
1787	0.6880
1788	0.6305
1789	0.6395
1790	0.6382
1791	0.6590
1792	0.6469
1793	0.6472
1794	0.6443
1795	0.6949
1796	0.6370
1797	0.6686
1798	0.6552
1799	0.6381
1800	0.6437
1801	0.6293
1802	0.6575
1803	0.6394
1804	0.6724
1805	0.6829
1806	0.6456
1807	0.6133
1808	0.6954
1809	0.6628
1810	0.6404
1811	0.6583
1812	0.6852
1813	0.6555
1814	0.6503
1815	0.6410
1816	0.6692
1817	0.6459
1818	0.6934
1819	0.6253
1820	0.6369
1821	0.6468
1822	0.6619
1823	0.6502
1824	0.6499
1825	0.6375
1826	0.6753
1827	0.6387
1828	0.6411
1829	0.6471
1830	0.6026
1831	0.6368
1832	0.6152
1833	0.6228
1834	0.6518
1835	0.6504
1836	0.6396
1837	0.6367
1838	0.6515
1839	0.6479
1840	0.6813
1841	0.6532
1842	0.6206
1843	0.6731
1844	0.6217
1845	0.6502
1846	0.6278
1847	0.7029
1848	0.6478
1849	0.6041
1850	0.6462
1851	0.6618
1852	0.6604
1853	0.6491
1854	0.6256
1855	0.6469
1856	0.6736
1857	0.6800
1858	0.6860
1859	0.6924
1860	0.6424
1861	0.6600
1862	0.6935
1863	0.6981
1864	0.6203
1865	0.6428
1866	0.6705
1867	0.6349
1868	0.6610
1869	0.6720
1870	0.6590
1871	0.6783
1872	0.6541
1873	0.6603
1874	0.6318
1875	0.6543
1876	0.6670
1877	0.6751
1878	0.6287
1879	0.7058
1880	0.6402
1881	0.6553
1882	0.6276
1883	0.6585
1884	0.6501
1885	0.6911
1886	0.6024
1887	0.6646
1888	0.6808
1889	0.6237
1890	0.6173
1891	0.6543
1892	0.6824
1893	0.6524
1894	0.6590
1895	0.6436
1896	0.6115
1897	0.6197
1898	0.6526
1899	0.6496
1900	0.6468
1901	0.6186
1902	0.6729
1903	0.6433
1904	0.6630
1905	0.6445
1906	0.6830
1907	0.6221
1908	0.6863
1909	0.6628
1910	0.6438
1911	0.6419
1912	0.6703
1913	0.6676
1914	0.6640
1915	0.7178
1916	0.6573
1917	0.6804
1918	0.6228
1919	0.6233
1920	0.6545
1921	0.6544
1922	0.6641
1923	0.6268
1924	0.6537
1925	0.6535
1926	0.6880
1927	0.6412
1928	0.6771
1929	0.6770
1930	0.6446
1931	0.6522
1932	0.6220
1933	0.6674
1934	0.6836
1935	0.6973
1936	0.6176
1937	0.6062
1938	0.6539
1939	0.6499
1940	0.6724
1941	0.6584
1942	0.6447
1943	0.6498
1944	0.6727
1945	0.6761
1946	0.6560
1947	0.6850
1948	0.6637
1949	0.6611
1950	0.6350
1951	0.6741
1952	0.6811
1953	0.6655
1954	0.6593
1955	0.6483
1956	0.6363
1957	0.6574
1958	0.6400
1959	0.6502
1960	0.6454
1961	0.6563
1962	0.6330
1963	0.6506
1964	0.6495
1965	0.6613
1966	0.6520
1967	0.6211
1968	0.6679
1969	0.6350
1970	0.6169
1971	0.6547
1972	0.6614
1973	0.6945
1974	0.6054
1975	0.6541
1976	0.6517
1977	0.6337
1978	0.6416
1979	0.6861
1980	0.6573
1981	0.6696
1982	0.6626
1983	0.6589
1984	0.6590
1985	0.6627
1986	0.6741
1987	0.6252
1988	0.6567
1989	0.6483
1990	0.6610
1991	0.6539
1992	0.6486
1993	0.6645
1994	0.6487
1995	0.6515
1996	0.6643
1997	0.6267
1998	0.6073
1999	0.6560
2000	0.6186
2001	0.6438
2002	0.6423
2003	0.6401
2004	0.6085
2005	0.6646
2006	0.6312
2007	0.6669
2008	0.6674
2009	0.6387
2010	0.6217
2011	0.6913
2012	0.6248
2013	0.6554
2014	0.6701
2015	0.6787
2016	0.6144
2017	0.6181
2018	0.6798
2019	0.6486
2020	0.6837
2021	0.6356
2022	0.6217
2023	0.7272
2024	0.6519
2025	0.6856
2026	0.7022
2027	0.6643
2028	0.6526
2029	0.6686
2030	0.7213
2031	0.6827
2032	0.6527
2033	0.6345
2034	0.6684
2035	0.6711
2036	0.6669
2037	0.6474
2038	0.6722
2039	0.6907
2040	0.6483
2041	0.6842
2042	0.6537
2043	0.6802
2044	0.6253
2045	0.6990
2046	0.6621
2047	0.6857
2048	0.6609
2049	0.6726
2050	0.6672
2051	0.6035
2052	0.6967
2053	0.6778
2054	0.6168
2055	0.7052
2056	0.6352
2057	0.6381
2058	0.6375
2059	0.6565
2060	0.6961
2061	0.6161
2062	0.6787
2063	0.6275
2064	0.6262
2065	0.6571
2066	0.6576
2067	0.6141
2068	0.6550
2069	0.6504
2070	0.7011
2071	0.6282
2072	0.6698
2073	0.6542
2074	0.6782
2075	0.6646
2076	0.6798
2077	0.6835
2078	0.6313
2079	0.7170
2080	0.6373
2081	0.6625
2082	0.6979
2083	0.6736
2084	0.6629
2085	0.6728
2086	0.6392
2087	0.6632
2088	0.6228
2089	0.6728
2090	0.6453
2091	0.6499
2092	0.6330
2093	0.6770
2094	0.6589
2095	0.6514
2096	0.6875
2097	0.6494
2098	0.6360
2099	0.6531
2100	0.6470
2101	0.6606
2102	0.6922
2103	0.6679
2104	0.6070
2105	0.5988
2106	0.6732
2107	0.6453
2108	0.6638
2109	0.6248
2110	0.6069
2111	0.6739
2112	0.6861
2113	0.6771
2114	0.6529
2115	0.6673
2116	0.6771
2117	0.6305
2118	0.6494
2119	0.6871
2120	0.6709
2121	0.6136
2122	0.7126
2123	0.6291
2124	0.6256
2125	0.6574
2126	0.6650
2127	0.6331
2128	0.6352
2129	0.6383
2130	0.6828
2131	0.6333
2132	0.6523
2133	0.6437
2134	0.6455
2135	0.6176
2136	0.6541
2137	0.6411
2138	0.6337
2139	0.6811
2140	0.6722
2141	0.6903
2142	0.6020
2143	0.6932
2144	0.6518
2145	0.6403
2146	0.6539
2147	0.7131
2148	0.6290
2149	0.6821
2150	0.6313
2151	0.6686
2152	0.6785
2153	0.7004
2154	0.6540
2155	0.6379
2156	0.6482
2157	0.6797
2158	0.6459
2159	0.6763
2160	0.6485
2161	0.6976
2162	0.6525
2163	0.6433
2164	0.6693
2165	0.6247
2166	0.6635
2167	0.6230
2168	0.6300
2169	0.6587
2170	0.6538
2171	0.6348
2172	0.6404
2173	0.6530
2174	0.6358
2175	0.6534
2176	0.6656
2177	0.6257
2178	0.6794
2179	0.6405
2180	0.6437
2181	0.6253
2182	0.6947
2183	0.6657
2184	0.6180
2185	0.6450
2186	0.6543
2187	0.6573
2188	0.6709
2189	0.6301
2190	0.6295
2191	0.6912
2192	0.6826
2193	0.7041
2194	0.6879
2195	0.6579
2196	0.6551
2197	0.7177
2198	0.7177
2199	0.6324
2200	0.6482
2201	0.6812
2202	0.6504
2203	0.6693
2204	0.6569
2205	0.6659
2206	0.6833
2207	0.6737
2208	0.6485
2209	0.6757
2210	0.6625
2211	0.6664
2212	0.6656
2213	0.6663
2214	0.6870
2215	0.6477
2216	0.6777
2217	0.6505
2218	0.6591
2219	0.6651
2220	0.7137
2221	0.6254
2222	0.6843
2223	0.6670
2224	0.6291
2225	0.6561
2226	0.6282
2227	0.6843
2228	0.6520
2229	0.6562
2230	0.6589
2231	0.6131
2232	0.6497
2233	0.6782
2234	0.6811
2235	0.6509
2236	0.6420
2237	0.6662
2238	0.6595
2239	0.6767
2240	0.6466
2241	0.6897
2242	0.6109
2243	0.6784
2244	0.6684
2245	0.6387
2246	0.6408
2247	0.6857
2248	0.6446
2249	0.6776
2250	0.7026
2251	0.6449
2252	0.7114
2253	0.6342
2254	0.6510
2255	0.6574
2256	0.6456
2257	0.6605
2258	0.6436
2259	0.6629
2260	0.6596
2261	0.7018
2262	0.6386
2263	0.6820
2264	0.6736
2265	0.6577
2266	0.6291
2267	0.6413
2268	0.6725
2269	0.6731
2270	0.7046
2271	0.6264
2272	0.6142
2273	0.6434
2274	0.6795
2275	0.6519
2276	0.6506
2277	0.6514
2278	0.6674
2279	0.6848
2280	0.6785
2281	0.6735
2282	0.6793
2283	0.6650
2284	0.6530
2285	0.6444
2286	0.6781
2287	0.6731
2288	0.6782
2289	0.6459
2290	0.6780
2291	0.6493
2292	0.6730
2293	0.6362
2294	0.6702
2295	0.6531
2296	0.6566
2297	0.6478
2298	0.6605
2299	0.6487
2300	0.6542
2301	0.6723
2302	0.6291
2303	0.6626
2304	0.6328
2305	0.6239
2306	0.6638
2307	0.6731
2308	0.7022
2309	0.6575
2310	0.6453
2311	0.6399
2312	0.6456
2313	0.6373
2314	0.7027
2315	0.6698
2316	0.6806
2317	0.6724
2318	0.6347
2319	0.6830
2320	0.6683
2321	0.6886
2322	0.6115
2323	0.6636
2324	0.6260
2325	0.6798
2326	0.6532
2327	0.6698
2328	0.6864
2329	0.6702
2330	0.6643
2331	0.6528
2332	0.6542
2333	0.6291
2334	0.6647
2335	0.6184
2336	0.6421
2337	0.6555
2338	0.6385
2339	0.6156
2340	0.6535
2341	0.6410
2342	0.6492
2343	0.6999
2344	0.6600
2345	0.6343
2346	0.6968
2347	0.6319
2348	0.6567
2349	0.6845
2350	0.6913
2351	0.6172
2352	0.6229
2353	0.6464
2354	0.6711
2355	0.6697
2356	0.6411
2357	0.6370
2358	0.6922
2359	0.6624
2360	0.7320
2361	0.7086
2362	0.6646
2363	0.6724
2364	0.6543
2365	0.7328
2366	0.6735
2367	0.6431
2368	0.6364
2369	0.6807
2370	0.6447
2371	0.6799
2372	0.6554
2373	0.6715
2374	0.7060
2375	0.6526
2376	0.6853
2377	0.6622
2378	0.6980
2379	0.6470
2380	0.7162
2381	0.6647
2382	0.6734
2383	0.6507
2384	0.6890
2385	0.6439
2386	0.6395
2387	0.6851
2388	0.6861
2389	0.6227
2390	0.6978
2391	0.6401
2392	0.6381
2393	0.6123
2394	0.6775
2395	0.6919
2396	0.6200
2397	0.6672
2398	0.6523
2399	0.6208
2400	0.6909
2401	0.7019
2402	0.5936
2403	0.6583
2404	0.6458
2405	0.6977
2406	0.6662
2407	0.6834
2408	0.6323
2409	0.7057
2410	0.6578
2411	0.6871
2412	0.6763
2413	0.6111
2414	0.6949
2415	0.6604
2416	0.6581
2417	0.6931
2418	0.6761
2419	0.6727
2420	0.6697
2421	0.6341
2422	0.6660
2423	0.6216
2424	0.6897
2425	0.6605
2426	0.6176
2427	0.6422
2428	0.6559
2429	0.6737
2430	0.6744
2431	0.6880
2432	0.6912
2433	0.6469
2434	0.6528
2435	0.6536
2436	0.6845
2437	0.7060
2438	0.6850
2439	0.6053
2440	0.6131
2441	0.6560
2442	0.6625
2443	0.6614
2444	0.6404
2445	0.5982
2446	0.6891
2447	0.7089
2448	0.6947
2449	0.6572
2450	0.6792
2451	0.6901
2452	0.6447
2453	0.6605
2454	0.7051
2455	0.6795
2456	0.6142
2457	0.7086
2458	0.6184
2459	0.6393
2460	0.6546
2461	0.6719
2462	0.6244
2463	0.6546
2464	0.6453
2465	0.6903
2466	0.6328
2467	0.6628
2468	0.6724
2469	0.6398
2470	0.6171
2471	0.6659
2472	0.6283
2473	0.6363
2474	0.7057
2475	0.6760
2476	0.7002
2477	0.6031
2478	0.6942
2479	0.6138
2480	0.6526
2481	0.6495
2482	0.7112
2483	0.6132
2484	0.6883
2485	0.6351
2486	0.6720
2487	0.6893
2488	0.7023
2489	0.6596
2490	0.6554
2491	0.6548
2492	0.6793
2493	0.6535
2494	0.6744
2495	0.6523
2496	0.7037
2497	0.6689
2498	0.6443
2499	0.7003
2500	0.6517
2501	0.6695
2502	0.6352
2503	0.6335
2504	0.6591
2505	0.6577
2506	0.6497
2507	0.6386
2508	0.6600
2509	0.6314
2510	0.6699
2511	0.6572
2512	0.6365
2513	0.6858
2514	0.6538
2515	0.6571
2516	0.6564
2517	0.7132
2518	0.6932
2519	0.6417
2520	0.6518
2521	0.6599
2522	0.6768
2523	0.6803
2524	0.6384
2525	0.6227
2526	0.7036
2527	0.6704
2528	0.6949
2529	0.7021
2530	0.6703
2531	0.6768
2532	0.7093
2533	0.7142
2534	0.6396
2535	0.6474
2536	0.6749
2537	0.6644
2538	0.6733
2539	0.6890
2540	0.6408
2541	0.6939
2542	0.6902
2543	0.6458
2544	0.6850
2545	0.6778
2546	0.6617
2547	0.6245
2548	0.6920
2549	0.6564
2550	0.6649
2551	0.6822
2552	0.6638
2553	0.6633
2554	0.6315
2555	0.7130
2556	0.6473
2557	0.6763
2558	0.6967
2559	0.6256
2560	0.6588
2561	0.6457
2562	0.6689
2563	0.6782
2564	0.6429
2565	0.6528
2566	0.6362
2567	0.6460
2568	0.6826
2569	0.6693
2570	0.6294
2571	0.6575
2572	0.6524
2573	0.6754
2574	0.6725
2575	0.6437
2576	0.6717
2577	0.6578
2578	0.6735
2579	0.6686
2580	0.6729
2581	0.6340
2582	0.7143
2583	0.6386
2584	0.6833
2585	0.7058
2586	0.6497
2587	0.7224
2588	0.6500
2589	0.6630
2590	0.6515
2591	0.6617
2592	0.6582
2593	0.6550
2594	0.6417
2595	0.6544
2596	0.7142
2597	0.6658
2598	0.6693
2599	0.6934
2600	0.6946
2601	0.6430
2602	0.6661
2603	0.6882
2604	0.6906
2605	0.7213
2606	0.6289
2607	0.6177
2608	0.6380
2609	0.6913
2610	0.6431
2611	0.6776
2612	0.6455
2613	0.6818
2614	0.6619
2615	0.6506
2616	0.6829
2617	0.6757
2618	0.6882
2619	0.6751
2620	0.6404
2621	0.6691
2622	0.7080
2623	0.6987
2624	0.6300
2625	0.7022
2626	0.6433
2627	0.6570
2628	0.6460
2629	0.6866
2630	0.6373
2631	0.6786
2632	0.6453
2633	0.6749
2634	0.6504
2635	0.6742
2636	0.6758
2637	0.6391
2638	0.6607
2639	0.6506
2640	0.6099

Measured result

Verbatim from the run's summary.json.

bucket

bucket-p32-t128

bucket_icons

3359

checkpoint_bytes

6463494

checkpoint_round_trip

true

checkpoint_sha256

584565f5646ef154…

config_sha256

1687c4625ce85296…

corruption

factorized_marginal_respecting_geometry

cuda_version

13.4

decision_threshold

0.7842316797885812

deterministic_algorithms

true

device

cuda

eval_every

final_train

loss

0.9516589045524597

step

2640

train_token_accuracy

0.6099060014461316

group_vocabulary_size

locked_path_exact

true

metrics_sha256

445e5c763f340a77…

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 2675 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

2640

evals_without_improvement

min_delta

0.0

objective

held_out_loss

patience_evals

selected_held_out_loss

1.0326451063156128

selected_step

2160

stopped_early

true

steps

6300

study_version

openmoji-g1-derived-threshold-v13

subgroup_vocabulary_size

118

torch_version

2.14.0a0+4fdf77b940.nv26.08

train_value_accuracy

0.27513338949733224

validation

accuracy

0.651869626074785

changed_accuracy

0.04935083566458647

changed_total

13941

loss

1.0326451063156128

retained_accuracy

0.9741051904707101

retained_total

26067

validation_final_step

accuracy

0.5990801839632074

changed_accuracy

0.08184491786815867

changed_total

13941

loss

1.0418919324874878

retained_accuracy

0.8757049142594084

retained_total

26067

validation_trace_sha256

1e3e2c5c755065cf…

validation_untrained

accuracy

0.3340831833633273

changed_accuracy

0.0012911555842479018

changed_total

13941

loss

1.7456328868865967

retained_accuracy

0.5120650631066099

retained_total

26067

State transitions

planned2026-09-21T03:10:00Z
completed2026-09-21T03:20:00Zbreak_even_threshold_beats_identity_with_no_calibration_but_trained_on_q_is_inflated

Run record

Verbatim from runs/openmoji-g1-derived-threshold-v13-7c73fab-1687c462-9b9b1699/run.yaml, the record committed before launch.

schema_version

run_id

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

state

completed

parent_run

openmoji-g1-balanced-objective-v12-9bf3fe9-499c2436-9b9b1699

hypothesis

v12 beat the identity baseline, 0.6567 against 0.6502, but only with a decode threshold fitted on a disjoint half of the held-out icons; at its own argmax decode it lost, 0.6086 against 0.6515. That threshold is derivable rather than arbitrary: keeping a retained field is always right and changing one is right only if the value head re-predicts the same token, so changing pays exactly above p = 1/(1+q) in the value head's accuracy q on corrupted fields. Estimating q on TRAIN icons and decoding at that threshold should beat identity with no calibration on held-out data at all.

expected_information_gain

scope_limits

changed_factors

primary

training.derive_decision_threshold false -> true. After training, the value head's accuracy on corrupted TRAIN fields is measured under a corruption draw disjoint from the held-out one, the decode threshold is set to 1/(1+q), and the held-out evaluation is repeated with it.

parameters

unchanged at 525,152

held_fixed

v12 exactly otherwise, including the entropy-balanced value loss weight.

no_training_change

This alters decoding and reporting only. The trained weights are identical to v12's; the checkpoint should match.

code

git_commit

7c73fab

execution_mode

native-local

config

path

configs/learning/openmoji-g1-derived-threshold-v13.yaml

sha256

1687c4625ce85296…

dataset

hybrid_sha256

9b9b1699677a6f97…

bucket

bucket-p32-t128

train_samples

2681

validation_samples

128

baselines

identity_on_the_full_held_out_draw

0.6515

v12_own_argmax_decode

0.6086

v12_with_a_threshold_fitted_on_held_out_icons

calibration_half

0.6568

evaluation_half

0.6567

identity_there

0.6502

v12_value_head_accuracy_on_corrupted_fields

0.1233

predeclared_criteria

evaluated_at

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

primary

beats_identity_without_calibration

statement

Held-out aggregate token accuracy at the DERIVED threshold exceeds the identity policy's, computed as retained_total / (changed_total + retained_total) from this run's own totals. No threshold is fitted on held-out data anywhere in this run.

improves_on_argmax

statement

Aggregate accuracy exceeds v12's own argmax decode, 0.6086.

threshold_is_in_the_expected_region

statement

The derived threshold lies between 0.85 and 0.95. The held-out value accuracy was 0.1233, giving 0.890; a train-split estimate should land near it, and a wildly different value would mean the estimator is measuring something else.

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 train-split value accuracy, the derived threshold, and the resulting flag rate, changed and retained accuracy, which will move in opposite directions by design

falsification_meaning

If the derived threshold does not beat identity while a fitted one does, the break-even argument is too crude - most likely because changing a retained field is not quite never right, or because the value head's accuracy on train icons does not transfer. Either would be worth knowing, and the fallback is to derive the threshold from a held-out-of-train split rather than from train itself.

outputs

local_metadata

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

report_root

reports/learning/openmoji-g1-derived-threshold-v13

durable_artifacts

/home/dev/.cache/openmoji-g1-derived-threshold-v13-7c73fab-1687c462-9b9b1699

planned_at

2026-09-21 03:10:00+00:00

completed_at

2026-09-21 03:20:00+00:00

result

device

cuda

train_value_accuracy

0.2751

derived_threshold

0.7842

held_out

aggregate

0.6519

changed

0.0494

retained

0.9741

identity_baseline

0.6515

margin

0.0003

checkpoint_identical_to_v12

true

identical_rerun

value

true

caveat

Its in-place idempotency rerun failed closed at a later commit, because v14 added `calibration_samples` and `calibration_withheld_from_training` to the summary schema after v13 had run, and the artifact writer refuses to replace a differing file. Re-running to a fresh report root at the current commit reproduces every other field exactly, so the run is deterministic and only the schema moved. The failure is recorded rather than papered over.

predeclared_outcome

overall

partially-falsified

beats_identity_without_calibration

passed

true

observed

0.6519

threshold

0.6515

improves_on_argmax

passed

true

observed

0.6519

threshold

0.6086

threshold_is_in_the_expected_region

passed

false

observed

0.7842

expected

0.85, 0.95

structural_safety

passed

true

reproducibility

passed

true

see

the caveat above

conclusion

The break-even threshold works - this beats identity with nothing fitted on held-out data - but by only 0.0003, and the threshold criterion is falsified for an informative reason. The value head scores 0.2751 on icons the model trained on against 0.1233 held out, so estimating q there gives 0.784 where the held-out accuracy implies 0.890, and the model edits more than pays. The fix is the one this run's falsification meaning named: estimate q on icons withheld from training.