openmoji-g1-marginal-corruption-v9-74dc5fd-0ef00ff4-9b9b1699

completed partially-falsified

Hypothesis

v8's detector reached 2.806 precision lift, but a zero-parameter detector that knows only the corpus marginal reaches 2.831 on the same data, so v8 matched a density test rather than learning geometry. Factorized corruption draws replacements uniformly from the full legal vocabulary, so many land on tokens real icons never use. Drawing them from the corpus distribution for the field's role instead removes that shortcut: the marginal detector falls to 0.771 while the local-continuity signal survives at 1.893. Under this leak-free process the model must use the surrounding geometry, and this run measures for the first time how much of that signal it can actually learn.

Why run it

This is the first honest detection benchmark in the project. Every earlier detector number - v6 1.234, v7 1.365, v8 2.806, and the free references at 2.13 and 2.83 - was measured on a process whose corruption is identifiable from a token marginal alone. A trained model that beats the continuity statistic HERE has learned geometry; one that does not has not, and no amount of the earlier numbers would have told us which.

Scope limits

Detection-only, so it reconstructs nothing and held-out accuracy is meaningless here. Single seed, probability 0.35. Not comparable with v8 in absolute terms: the task is strictly harder because the easy fraction of corruptions has been removed, so a lower lift than v8's 2.806 is expected and is not a regression.

Measured behaviour

Held-out loss

Table view

optimizer step	held-out loss
0	1.4055
60	0.6704
120	0.6533
180	0.6493
240	0.6493
300	0.6487
360	0.6469
420	0.6472
480	0.6451
540	0.6458
600	0.6456
660	0.6446
720	0.6449
780	0.6436
840	0.6429
900	0.6427
960	0.6434
1020	0.6412
1080	0.6412
1140	0.6402
1200	0.6393
1260	0.6405
1320	0.6392
1380	0.6395
1440	0.6390
1500	0.6361
1560	0.6369
1620	0.6396
1680	0.6372
1740	0.6352
1800	0.6369
1860	0.6349
1920	0.6350
1980	0.6333
2040	0.6365
2100	0.6362
2160	0.6319
2220	0.6323
2280	0.6329
2340	0.6305
2400	0.6328
2460	0.6325
2520	0.6323
2580	0.6330
2640	0.6334
2700	0.6322
2760	0.6351
2820	0.6339

Held-out token accuracy

aggregatechanged fieldsretained fields

Table view

optimizer step	aggregate	changed fields	retained fields
0	0.2675	0.0021	0.4095
60	0.5955	0.0004	0.9138
120	0.6346	0.0001	0.9740
180	0.6445	0.0000	0.9891
240	0.6491	0.0000	0.9962
300	0.6488	0.0000	0.9957
360	0.6490	0.0000	0.9960
420	0.6476	0.0000	0.9940
480	0.6484	0.0000	0.9952
540	0.6434	0.0000	0.9875
600	0.6455	0.0000	0.9907
660	0.6412	0.0000	0.9841
720	0.6454	0.0001	0.9906
780	0.6466	0.0000	0.9925
840	0.6450	0.0000	0.9899
900	0.6442	0.0000	0.9887
960	0.6433	0.0001	0.9872
1020	0.6414	0.0001	0.9844
1080	0.6397	0.0001	0.9818
1140	0.6366	0.0001	0.9770
1200	0.6366	0.0002	0.9770
1260	0.6290	0.0003	0.9653
1320	0.6256	0.0002	0.9601
1380	0.6270	0.0001	0.9623
1440	0.6207	0.0001	0.9525
1500	0.6247	0.0004	0.9586
1560	0.6190	0.0002	0.9499
1620	0.6121	0.0003	0.9393
1680	0.6014	0.0001	0.9230
1740	0.6126	0.0002	0.9400
1800	0.6063	0.0003	0.9304
1860	0.6051	0.0004	0.9285
1920	0.6028	0.0005	0.9250
1980	0.6126	0.0002	0.9402
2040	0.5842	0.0004	0.8965
2100	0.5880	0.0004	0.9023
2160	0.5969	0.0006	0.9157
2220	0.5966	0.0005	0.9154
2280	0.5975	0.0002	0.9169
2340	0.6026	0.0003	0.9247
2400	0.5913	0.0006	0.9072
2460	0.5887	0.0004	0.9033
2520	0.5776	0.0004	0.8863
2580	0.5773	0.0006	0.8858
2640	0.5777	0.0005	0.8864
2700	0.5853	0.0004	0.8981
2760	0.5733	0.0008	0.8794
2820	0.5701	0.0006	0.8747

Training token accuracy (per batch)

Table view

optimizer step	train accuracy
1	0.2793
2	0.2790
3	0.3806
4	0.4000
5	0.4351
6	0.4741
7	0.5008
8	0.5185
9	0.5284
10	0.4963
11	0.4852
12	0.5032
13	0.4913
14	0.4837
15	0.4645
16	0.4499
17	0.4399
18	0.4599
19	0.4378
20	0.4559
21	0.4292
22	0.4376
23	0.4500
24	0.4659
25	0.4654
26	0.4710
27	0.4789
28	0.4998
29	0.4977
30	0.4929
31	0.5275
32	0.5186
33	0.4926
34	0.5145
35	0.5280
36	0.5219
37	0.5152
38	0.5190
39	0.5241
40	0.5338
41	0.5548
42	0.5574
43	0.5436
44	0.5688
45	0.5438
46	0.5471
47	0.5597
48	0.5418
49	0.5326
50	0.5473
51	0.5526
52	0.5578
53	0.5504
54	0.5623
55	0.5717
56	0.5594
57	0.5836
58	0.5872
59	0.6005
60	0.5905
61	0.5897
62	0.6065
63	0.5987
64	0.6012
65	0.6129
66	0.6041
67	0.5892
68	0.6060
69	0.5996
70	0.5997
71	0.6203
72	0.6118
73	0.6228
74	0.6023
75	0.6211
76	0.6194
77	0.6171
78	0.6233
79	0.5843
80	0.6206
81	0.6156
82	0.6122
83	0.6028
84	0.6101
85	0.6190
86	0.6132
87	0.6062
88	0.6128
89	0.6258
90	0.6257
91	0.6284
92	0.6263
93	0.6145
94	0.6409
95	0.6349
96	0.6240
97	0.6276
98	0.6402
99	0.6346
100	0.6262
101	0.6424
102	0.6309
103	0.6368
104	0.6317
105	0.6471
106	0.6318
107	0.6358
108	0.6294
109	0.6546
110	0.6407
111	0.6407
112	0.6394
113	0.6348
114	0.6441
115	0.6458
116	0.6444
117	0.6344
118	0.6444
119	0.6317
120	0.6538
121	0.6308
122	0.6557
123	0.6318
124	0.6423
125	0.6528
126	0.6444
127	0.6380
128	0.6443
129	0.6406
130	0.6419
131	0.6325
132	0.6308
133	0.6318
134	0.6360
135	0.6186
136	0.6221
137	0.6212
138	0.6440
139	0.6358
140	0.6281
141	0.6399
142	0.6425
143	0.6377
144	0.6342
145	0.6413
146	0.6452
147	0.6469
148	0.6386
149	0.6422
150	0.6308
151	0.6451
152	0.6339
153	0.6376
154	0.6302
155	0.6601
156	0.6402
157	0.6320
158	0.6423
159	0.6365
160	0.6461
161	0.6388
162	0.6414
163	0.6519
164	0.6382
165	0.6343
166	0.6379
167	0.6420
168	0.6386
169	0.6520
170	0.6514
171	0.6499
172	0.6497
173	0.6488
174	0.6399
175	0.6452
176	0.6368
177	0.6391
178	0.6341
179	0.6512
180	0.6372
181	0.6371
182	0.6419
183	0.6422
184	0.6496
185	0.6567
186	0.6498
187	0.6414
188	0.6374
189	0.6398
190	0.6477
191	0.6467
192	0.6405
193	0.6494
194	0.6381
195	0.6497
196	0.6460
197	0.6337
198	0.6343
199	0.6412
200	0.6497
201	0.6437
202	0.6438
203	0.6598
204	0.6525
205	0.6426
206	0.6482
207	0.6513
208	0.6607
209	0.6530
210	0.6530
211	0.6359
212	0.6537
213	0.6472
214	0.6347
215	0.6433
216	0.6314
217	0.6396
218	0.6401
219	0.6520
220	0.6386
221	0.6454
222	0.6429
223	0.6500
224	0.6559
225	0.6602
226	0.6433
227	0.6367
228	0.6372
229	0.6468
230	0.6469
231	0.6294
232	0.6412
233	0.6479
234	0.6445
235	0.6269
236	0.6499
237	0.6375
238	0.6504
239	0.6551
240	0.6533
241	0.6393
242	0.6427
243	0.6581
244	0.6397
245	0.6488
246	0.6482
247	0.6350
248	0.6520
249	0.6401
250	0.6464
251	0.6536
252	0.6528
253	0.6500
254	0.6562
255	0.6498
256	0.6516
257	0.6526
258	0.6442
259	0.6631
260	0.6412
261	0.6343
262	0.6457
263	0.6520
264	0.6427
265	0.6538
266	0.6405
267	0.6536
268	0.6386
269	0.6440
270	0.6469
271	0.6375
272	0.6563
273	0.6397
274	0.6329
275	0.6293
276	0.6443
277	0.6468
278	0.6368
279	0.6540
280	0.6374
281	0.6477
282	0.6547
283	0.6414
284	0.6547
285	0.6495
286	0.6524
287	0.6418
288	0.6413
289	0.6551
290	0.6561
291	0.6357
292	0.6488
293	0.6475
294	0.6521
295	0.6446
296	0.6487
297	0.6506
298	0.6534
299	0.6449
300	0.6506
301	0.6562
302	0.6537
303	0.6419
304	0.6455
305	0.6270
306	0.6390
307	0.6392
308	0.6484
309	0.6474
310	0.6356
311	0.6340
312	0.6578
313	0.6548
314	0.6547
315	0.6407
316	0.6441
317	0.6513
318	0.6471
319	0.6432
320	0.6587
321	0.6462
322	0.6425
323	0.6343
324	0.6409
325	0.6344
326	0.6444
327	0.6463
328	0.6426
329	0.6549
330	0.6389
331	0.6407
332	0.6413
333	0.6495
334	0.6399
335	0.6337
336	0.6572
337	0.6344
338	0.6533
339	0.6480
340	0.6406
341	0.6438
342	0.6422
343	0.6471
344	0.6495
345	0.6506
346	0.6454
347	0.6523
348	0.6333
349	0.6500
350	0.6545
351	0.6526
352	0.6455
353	0.6362
354	0.6485
355	0.6454
356	0.6452
357	0.6504
358	0.6382
359	0.6339
360	0.6467
361	0.6509
362	0.6431
363	0.6551
364	0.6376
365	0.6402
366	0.6530
367	0.6512
368	0.6379
369	0.6384
370	0.6415
371	0.6608
372	0.6409
373	0.6536
374	0.6477
375	0.6482
376	0.6460
377	0.6471
378	0.6381
379	0.6518
380	0.6485
381	0.6404
382	0.6450
383	0.6392
384	0.6396
385	0.6350
386	0.6416
387	0.6431
388	0.6491
389	0.6348
390	0.6420
391	0.6462
392	0.6576
393	0.6306
394	0.6596
395	0.6356
396	0.6550
397	0.6592
398	0.6458
399	0.6498
400	0.6402
401	0.6477
402	0.6497
403	0.6463
404	0.6509
405	0.6481
406	0.6469
407	0.6353
408	0.6542
409	0.6509
410	0.6459
411	0.6373
412	0.6475
413	0.6447
414	0.6398
415	0.6408
416	0.6412
417	0.6457
418	0.6437
419	0.6401
420	0.6485
421	0.6523
422	0.6538
423	0.6425
424	0.6539
425	0.6467
426	0.6467
427	0.6472
428	0.6417
429	0.6422
430	0.6517
431	0.6462
432	0.6582
433	0.6458
434	0.6499
435	0.6500
436	0.6365
437	0.6522
438	0.6533
439	0.6404
440	0.6587
441	0.6477
442	0.6431
443	0.6439
444	0.6455
445	0.6477
446	0.6384
447	0.6453
448	0.6406
449	0.6425
450	0.6259
451	0.6352
452	0.6487
453	0.6513
454	0.6435
455	0.6528
456	0.6583
457	0.6561
458	0.6370
459	0.6356
460	0.6483
461	0.6450
462	0.6442
463	0.6422
464	0.6459
465	0.6449
466	0.6407
467	0.6383
468	0.6514
469	0.6455
470	0.6533
471	0.6546
472	0.6453
473	0.6619
474	0.6313
475	0.6478
476	0.6491
477	0.6440
478	0.6409
479	0.6445
480	0.6556
481	0.6418
482	0.6582
483	0.6552
484	0.6556
485	0.6401
486	0.6581
487	0.6544
488	0.6416
489	0.6537
490	0.6479
491	0.6353
492	0.6395
493	0.6425
494	0.6415
495	0.6427
496	0.6466
497	0.6463
498	0.6398
499	0.6464
500	0.6619
501	0.6539
502	0.6515
503	0.6422
504	0.6525
505	0.6363
506	0.6437
507	0.6413
508	0.6544
509	0.6469
510	0.6380
511	0.6443
512	0.6558
513	0.6502
514	0.6446
515	0.6443
516	0.6531
517	0.6459
518	0.6442
519	0.6441
520	0.6519
521	0.6362
522	0.6545
523	0.6506
524	0.6440
525	0.6494
526	0.6611
527	0.6302
528	0.6556
529	0.6511
530	0.6499
531	0.6405
532	0.6603
533	0.6541
534	0.6562
535	0.6458
536	0.6367
537	0.6468
538	0.6431
539	0.6432
540	0.6349
541	0.6454
542	0.6422
543	0.6339
544	0.6504
545	0.6475
546	0.6317
547	0.6481
548	0.6273
549	0.6530
550	0.6321
551	0.6358
552	0.6410
553	0.6364
554	0.6474
555	0.6365
556	0.6350
557	0.6458
558	0.6314
559	0.6584
560	0.6390
561	0.6486
562	0.6425
563	0.6399
564	0.6459
565	0.6484
566	0.6383
567	0.6409
568	0.6414
569	0.6517
570	0.6379
571	0.6454
572	0.6610
573	0.6542
574	0.6481
575	0.6484
576	0.6526
577	0.6347
578	0.6446
579	0.6496
580	0.6346
581	0.6549
582	0.6426
583	0.6506
584	0.6381
585	0.6413
586	0.6460
587	0.6451
588	0.6579
589	0.6339
590	0.6552
591	0.6480
592	0.6366
593	0.6504
594	0.6474
595	0.6397
596	0.6477
597	0.6372
598	0.6412
599	0.6460
600	0.6525
601	0.6482
602	0.6605
603	0.6471
604	0.6493
605	0.6615
606	0.6456
607	0.6575
608	0.6432
609	0.6349
610	0.6292
611	0.6358
612	0.6466
613	0.6449
614	0.6620
615	0.6467
616	0.6484
617	0.6384
618	0.6427
619	0.6449
620	0.6447
621	0.6381
622	0.6432
623	0.6302
624	0.6463
625	0.6509
626	0.6486
627	0.6478
628	0.6408
629	0.6487
630	0.6623
631	0.6559
632	0.6432
633	0.6496
634	0.6445
635	0.6442
636	0.6446
637	0.6443
638	0.6409
639	0.6464
640	0.6344
641	0.6477
642	0.6398
643	0.6526
644	0.6387
645	0.6442
646	0.6367
647	0.6443
648	0.6463
649	0.6445
650	0.6521
651	0.6491
652	0.6556
653	0.6623
654	0.6460
655	0.6467
656	0.6424
657	0.6462
658	0.6492
659	0.6359
660	0.6352
661	0.6413
662	0.6440
663	0.6375
664	0.6484
665	0.6479
666	0.6450
667	0.6447
668	0.6335
669	0.6323
670	0.6548
671	0.6458
672	0.6437
673	0.6466
674	0.6553
675	0.6386
676	0.6490
677	0.6421
678	0.6360
679	0.6486
680	0.6445
681	0.6355
682	0.6411
683	0.6410
684	0.6537
685	0.6476
686	0.6414
687	0.6491
688	0.6389
689	0.6436
690	0.6440
691	0.6420
692	0.6572
693	0.6406
694	0.6452
695	0.6482
696	0.6392
697	0.6371
698	0.6475
699	0.6374
700	0.6333
701	0.6454
702	0.6487
703	0.6381
704	0.6519
705	0.6393
706	0.6529
707	0.6435
708	0.6304
709	0.6536
710	0.6420
711	0.6491
712	0.6410
713	0.6468
714	0.6492
715	0.6559
716	0.6561
717	0.6529
718	0.6392
719	0.6443
720	0.6465
721	0.6479
722	0.6464
723	0.6405
724	0.6410
725	0.6419
726	0.6526
727	0.6388
728	0.6484
729	0.6466
730	0.6473
731	0.6413
732	0.6448
733	0.6533
734	0.6298
735	0.6446
736	0.6538
737	0.6456
738	0.6446
739	0.6508
740	0.6435
741	0.6492
742	0.6483
743	0.6428
744	0.6460
745	0.6279
746	0.6494
747	0.6527
748	0.6501
749	0.6273
750	0.6361
751	0.6498
752	0.6463
753	0.6325
754	0.6428
755	0.6422
756	0.6501
757	0.6471
758	0.6330
759	0.6531
760	0.6395
761	0.6643
762	0.6465
763	0.6517
764	0.6427
765	0.6530
766	0.6554
767	0.6591
768	0.6443
769	0.6394
770	0.6366
771	0.6418
772	0.6341
773	0.6516
774	0.6554
775	0.6368
776	0.6361
777	0.6620
778	0.6560
779	0.6526
780	0.6378
781	0.6430
782	0.6441
783	0.6454
784	0.6558
785	0.6382
786	0.6409
787	0.6543
788	0.6587
789	0.6379
790	0.6399
791	0.6482
792	0.6396
793	0.6211
794	0.6435
795	0.6448
796	0.6467
797	0.6490
798	0.6427
799	0.6569
800	0.6435
801	0.6353
802	0.6425
803	0.6466
804	0.6445
805	0.6330
806	0.6344
807	0.6423
808	0.6472
809	0.6412
810	0.6544
811	0.6395
812	0.6419
813	0.6397
814	0.6393
815	0.6387
816	0.6343
817	0.6449
818	0.6460
819	0.6442
820	0.6447
821	0.6565
822	0.6449
823	0.6536
824	0.6415
825	0.6332
826	0.6524
827	0.6274
828	0.6327
829	0.6420
830	0.6497
831	0.6464
832	0.6500
833	0.6431
834	0.6403
835	0.6425
836	0.6300
837	0.6349
838	0.6374
839	0.6492
840	0.6538
841	0.6512
842	0.6460
843	0.6382
844	0.6536
845	0.6432
846	0.6475
847	0.6492
848	0.6609
849	0.6375
850	0.6588
851	0.6339
852	0.6474
853	0.6420
854	0.6519
855	0.6406
856	0.6546
857	0.6344
858	0.6374
859	0.6465
860	0.6397
861	0.6329
862	0.6298
863	0.6508
864	0.6456
865	0.6464
866	0.6378
867	0.6455
868	0.6484
869	0.6436
870	0.6446
871	0.6458
872	0.6520
873	0.6402
874	0.6393
875	0.6444
876	0.6376
877	0.6398
878	0.6424
879	0.6439
880	0.6404
881	0.6394
882	0.6472
883	0.6523
884	0.6489
885	0.6340
886	0.6472
887	0.6439
888	0.6473
889	0.6374
890	0.6356
891	0.6400
892	0.6489
893	0.6497
894	0.6291
895	0.6496
896	0.6481
897	0.6451
898	0.6507
899	0.6514
900	0.6521
901	0.6460
902	0.6619
903	0.6303
904	0.6488
905	0.6447
906	0.6462
907	0.6446
908	0.6529
909	0.6363
910	0.6432
911	0.6370
912	0.6417
913	0.6329
914	0.6382
915	0.6476
916	0.6337
917	0.6547
918	0.6447
919	0.6551
920	0.6291
921	0.6523
922	0.6430
923	0.6413
924	0.6505
925	0.6362
926	0.6285
927	0.6287
928	0.6475
929	0.6498
930	0.6328
931	0.6499
932	0.6461
933	0.6325
934	0.6437
935	0.6471
936	0.6414
937	0.6482
938	0.6608
939	0.6353
940	0.6446
941	0.6451
942	0.6495
943	0.6445
944	0.6428
945	0.6430
946	0.6451
947	0.6359
948	0.6484
949	0.6470
950	0.6514
951	0.6400
952	0.6373
953	0.6392
954	0.6454
955	0.6450
956	0.6353
957	0.6423
958	0.6338
959	0.6453
960	0.6464
961	0.6316
962	0.6487
963	0.6503
964	0.6497
965	0.6444
966	0.6382
967	0.6485
968	0.6485
969	0.6414
970	0.6369
971	0.6274
972	0.6399
973	0.6359
974	0.6483
975	0.6386
976	0.6385
977	0.6380
978	0.6475
979	0.6393
980	0.6408
981	0.6481
982	0.6461
983	0.6285
984	0.6560
985	0.6433
986	0.6512
987	0.6392
988	0.6303
989	0.6505
990	0.6394
991	0.6418
992	0.6439
993	0.6580
994	0.6382
995	0.6530
996	0.6453
997	0.6469
998	0.6279
999	0.6459
1000	0.6487
1001	0.6337
1002	0.6422
1003	0.6267
1004	0.6413
1005	0.6351
1006	0.6376
1007	0.6507
1008	0.6444
1009	0.6438
1010	0.6417
1011	0.6253
1012	0.6421
1013	0.6450
1014	0.6392
1015	0.6365
1016	0.6396
1017	0.6411
1018	0.6431
1019	0.6311
1020	0.6530
1021	0.6365
1022	0.6423
1023	0.6484
1024	0.6487
1025	0.6500
1026	0.6458
1027	0.6367
1028	0.6515
1029	0.6408
1030	0.6386
1031	0.6390
1032	0.6329
1033	0.6411
1034	0.6408
1035	0.6448
1036	0.6419
1037	0.6378
1038	0.6536
1039	0.6471
1040	0.6365
1041	0.6567
1042	0.6407
1043	0.6353
1044	0.6458
1045	0.6399
1046	0.6403
1047	0.6411
1048	0.6433
1049	0.6430
1050	0.6446
1051	0.6388
1052	0.6341
1053	0.6298
1054	0.6276
1055	0.6388
1056	0.6385
1057	0.6281
1058	0.6342
1059	0.6393
1060	0.6405
1061	0.6490
1062	0.6431
1063	0.6384
1064	0.6433
1065	0.6410
1066	0.6249
1067	0.6437
1068	0.6410
1069	0.6313
1070	0.6286
1071	0.6353
1072	0.6234
1073	0.6572
1074	0.6400
1075	0.6557
1076	0.6528
1077	0.6393
1078	0.6455
1079	0.6572
1080	0.6411
1081	0.6327
1082	0.6441
1083	0.6468
1084	0.6354
1085	0.6483
1086	0.6528
1087	0.6455
1088	0.6533
1089	0.6457
1090	0.6487
1091	0.6385
1092	0.6304
1093	0.6349
1094	0.6492
1095	0.6347
1096	0.6207
1097	0.6546
1098	0.6421
1099	0.6437
1100	0.6493
1101	0.6376
1102	0.6335
1103	0.6465
1104	0.6328
1105	0.6389
1106	0.6432
1107	0.6368
1108	0.6478
1109	0.6434
1110	0.6385
1111	0.6370
1112	0.6358
1113	0.6523
1114	0.6398
1115	0.6473
1116	0.6457
1117	0.6473
1118	0.6476
1119	0.6529
1120	0.6403
1121	0.6455
1122	0.6408
1123	0.6355
1124	0.6318
1125	0.6408
1126	0.6338
1127	0.6431
1128	0.6279
1129	0.6381
1130	0.6463
1131	0.6318
1132	0.6494
1133	0.6245
1134	0.6462
1135	0.6285
1136	0.6425
1137	0.6387
1138	0.6261
1139	0.6318
1140	0.6451
1141	0.6399
1142	0.6308
1143	0.6315
1144	0.6367
1145	0.6284
1146	0.6411
1147	0.6327
1148	0.6281
1149	0.6386
1150	0.6391
1151	0.6499
1152	0.6465
1153	0.6398
1154	0.6381
1155	0.6305
1156	0.6444
1157	0.6388
1158	0.6368
1159	0.6349
1160	0.6228
1161	0.6337
1162	0.6422
1163	0.6345
1164	0.6385
1165	0.6367
1166	0.6442
1167	0.6464
1168	0.6493
1169	0.6388
1170	0.6543
1171	0.6440
1172	0.6239
1173	0.6409
1174	0.6219
1175	0.6227
1176	0.6383
1177	0.6257
1178	0.6246
1179	0.6261
1180	0.6319
1181	0.6323
1182	0.6410
1183	0.6305
1184	0.6439
1185	0.6414
1186	0.6474
1187	0.6513
1188	0.6429
1189	0.6447
1190	0.6250
1191	0.6294
1192	0.6374
1193	0.6437
1194	0.6438
1195	0.6374
1196	0.6494
1197	0.6457
1198	0.6454
1199	0.6319
1200	0.6416
1201	0.6417
1202	0.6303
1203	0.6369
1204	0.6528
1205	0.6346
1206	0.6416
1207	0.6394
1208	0.6531
1209	0.6472
1210	0.6320
1211	0.6509
1212	0.6344
1213	0.6411
1214	0.6375
1215	0.6498
1216	0.6392
1217	0.6450
1218	0.6475
1219	0.6438
1220	0.6369
1221	0.6528
1222	0.6437
1223	0.6408
1224	0.6423
1225	0.6371
1226	0.6364
1227	0.6524
1228	0.6460
1229	0.6276
1230	0.6422
1231	0.6265
1232	0.6415
1233	0.6343
1234	0.6431
1235	0.6363
1236	0.6408
1237	0.6414
1238	0.6404
1239	0.6439
1240	0.6364
1241	0.6391
1242	0.6378
1243	0.6475
1244	0.6277
1245	0.6416
1246	0.6306
1247	0.6291
1248	0.6390
1249	0.6469
1250	0.6327
1251	0.6423
1252	0.6184
1253	0.6400
1254	0.6277
1255	0.6266
1256	0.6338
1257	0.6316
1258	0.6282
1259	0.6217
1260	0.6272
1261	0.6223
1262	0.6257
1263	0.6304
1264	0.6348
1265	0.6408
1266	0.6372
1267	0.6362
1268	0.6374
1269	0.6422
1270	0.6435
1271	0.6239
1272	0.6409
1273	0.6245
1274	0.6427
1275	0.6291
1276	0.6472
1277	0.6240
1278	0.6325
1279	0.6343
1280	0.6336
1281	0.6355
1282	0.6359
1283	0.6188
1284	0.6458
1285	0.6255
1286	0.6288
1287	0.6159
1288	0.6390
1289	0.6248
1290	0.6368
1291	0.6329
1292	0.6290
1293	0.6180
1294	0.6311
1295	0.6426
1296	0.6345
1297	0.6463
1298	0.6254
1299	0.6310
1300	0.6250
1301	0.6339
1302	0.6251
1303	0.6249
1304	0.6271
1305	0.6211
1306	0.6454
1307	0.6307
1308	0.6468
1309	0.6382
1310	0.6390
1311	0.6334
1312	0.6294
1313	0.6429
1314	0.6418
1315	0.6464
1316	0.6364
1317	0.6417
1318	0.6404
1319	0.6350
1320	0.6324
1321	0.6377
1322	0.6450
1323	0.6311
1324	0.6258
1325	0.6311
1326	0.6293
1327	0.6324
1328	0.6301
1329	0.6240
1330	0.6442
1331	0.6278
1332	0.6310
1333	0.6366
1334	0.6341
1335	0.6440
1336	0.6308
1337	0.6381
1338	0.6467
1339	0.6403
1340	0.6469
1341	0.6398
1342	0.6442
1343	0.6310
1344	0.6341
1345	0.6296
1346	0.6290
1347	0.6276
1348	0.6387
1349	0.6273
1350	0.6375
1351	0.6506
1352	0.6374
1353	0.6485
1354	0.6423
1355	0.6247
1356	0.6386
1357	0.6353
1358	0.6303
1359	0.6291
1360	0.6370
1361	0.6273
1362	0.6285
1363	0.6453
1364	0.6397
1365	0.6483
1366	0.6467
1367	0.6333
1368	0.6251
1369	0.6511
1370	0.6370
1371	0.6365
1372	0.6283
1373	0.6229
1374	0.6369
1375	0.6415
1376	0.6340
1377	0.6323
1378	0.6185
1379	0.6303
1380	0.6263
1381	0.6248
1382	0.6365
1383	0.6325
1384	0.6190
1385	0.6185
1386	0.6283
1387	0.6185
1388	0.6176
1389	0.6232
1390	0.6326
1391	0.6380
1392	0.6258
1393	0.6334
1394	0.6451
1395	0.6357
1396	0.6391
1397	0.6211
1398	0.6339
1399	0.6132
1400	0.6354
1401	0.6248
1402	0.6262
1403	0.6192
1404	0.6343
1405	0.6376
1406	0.6331
1407	0.6418
1408	0.6369
1409	0.6430
1410	0.6429
1411	0.6362
1412	0.6181
1413	0.6412
1414	0.6131
1415	0.6344
1416	0.6318
1417	0.6298
1418	0.6298
1419	0.6300
1420	0.6264
1421	0.6303
1422	0.6365
1423	0.6195
1424	0.6275
1425	0.6234
1426	0.6268
1427	0.6337
1428	0.6311
1429	0.6145
1430	0.6188
1431	0.6293
1432	0.6330
1433	0.6187
1434	0.6345
1435	0.6316
1436	0.6259
1437	0.6382
1438	0.6218
1439	0.6270
1440	0.6184
1441	0.6248
1442	0.6182
1443	0.6372
1444	0.6321
1445	0.6231
1446	0.6267
1447	0.6194
1448	0.6443
1449	0.6267
1450	0.6326
1451	0.6286
1452	0.6180
1453	0.6390
1454	0.6138
1455	0.6326
1456	0.6323
1457	0.6308
1458	0.6316
1459	0.6199
1460	0.6229
1461	0.6399
1462	0.6354
1463	0.6381
1464	0.6334
1465	0.6330
1466	0.6272
1467	0.6233
1468	0.6241
1469	0.6341
1470	0.6281
1471	0.6085
1472	0.6382
1473	0.6352
1474	0.6205
1475	0.6399
1476	0.6261
1477	0.6366
1478	0.6224
1479	0.6352
1480	0.6273
1481	0.6282
1482	0.6260
1483	0.6232
1484	0.6325
1485	0.6456
1486	0.6233
1487	0.6192
1488	0.6154
1489	0.6269
1490	0.6189
1491	0.6165
1492	0.6129
1493	0.6285
1494	0.6404
1495	0.6303
1496	0.6245
1497	0.6265
1498	0.6440
1499	0.6339
1500	0.6242
1501	0.6292
1502	0.6333
1503	0.6286
1504	0.6293
1505	0.6270
1506	0.6347
1507	0.6394
1508	0.6294
1509	0.6447
1510	0.6422
1511	0.6384
1512	0.6205
1513	0.6345
1514	0.6144
1515	0.6328
1516	0.6432
1517	0.6284
1518	0.6292
1519	0.6109
1520	0.6344
1521	0.6275
1522	0.6322
1523	0.6249
1524	0.6306
1525	0.6403
1526	0.6371
1527	0.6291
1528	0.6439
1529	0.6320
1530	0.6321
1531	0.6364
1532	0.6328
1533	0.6354
1534	0.6196
1535	0.6306
1536	0.6194
1537	0.6280
1538	0.6194
1539	0.6113
1540	0.6140
1541	0.6186
1542	0.6163
1543	0.6280
1544	0.6126
1545	0.6397
1546	0.6166
1547	0.6152
1548	0.6097
1549	0.6220
1550	0.6257
1551	0.6150
1552	0.6349
1553	0.6205
1554	0.6329
1555	0.6352
1556	0.6328
1557	0.6311
1558	0.6256
1559	0.6270
1560	0.6242
1561	0.6281
1562	0.6254
1563	0.6211
1564	0.6146
1565	0.6207
1566	0.6152
1567	0.6301
1568	0.6359
1569	0.6334
1570	0.6313
1571	0.6279
1572	0.6311
1573	0.6348
1574	0.6287
1575	0.6112
1576	0.6138
1577	0.6052
1578	0.6048
1579	0.6175
1580	0.6152
1581	0.6230
1582	0.6389
1583	0.6310
1584	0.6285
1585	0.6415
1586	0.6284
1587	0.6358
1588	0.6363
1589	0.6280
1590	0.6219
1591	0.6187
1592	0.6125
1593	0.6250
1594	0.6096
1595	0.6271
1596	0.6141
1597	0.6126
1598	0.6238
1599	0.6254
1600	0.6300
1601	0.6366
1602	0.6276
1603	0.6377
1604	0.6350
1605	0.6304
1606	0.6220
1607	0.6331
1608	0.6267
1609	0.6230
1610	0.6332
1611	0.6125
1612	0.6287
1613	0.6219
1614	0.6127
1615	0.6335
1616	0.6364
1617	0.6277
1618	0.6153
1619	0.6263
1620	0.6462
1621	0.6234
1622	0.6106
1623	0.6289
1624	0.6313
1625	0.6233
1626	0.6276
1627	0.6225
1628	0.6311
1629	0.6388
1630	0.6138
1631	0.6314
1632	0.6372
1633	0.6260
1634	0.6362
1635	0.6336
1636	0.6133
1637	0.6205
1638	0.6041
1639	0.6095
1640	0.6267
1641	0.6253
1642	0.6177
1643	0.6250
1644	0.6245
1645	0.6171
1646	0.6253
1647	0.6338
1648	0.6143
1649	0.6207
1650	0.6217
1651	0.6231
1652	0.6261
1653	0.6329
1654	0.6240
1655	0.6381
1656	0.6455
1657	0.6367
1658	0.6239
1659	0.6288
1660	0.6373
1661	0.6164
1662	0.6173
1663	0.6104
1664	0.6270
1665	0.6142
1666	0.6272
1667	0.6195
1668	0.6271
1669	0.6197
1670	0.6316
1671	0.6222
1672	0.6088
1673	0.6245
1674	0.6369
1675	0.6221
1676	0.6103
1677	0.6261
1678	0.6218
1679	0.6097
1680	0.6064
1681	0.6089
1682	0.5994
1683	0.6145
1684	0.6110
1685	0.6206
1686	0.6274
1687	0.6250
1688	0.6168
1689	0.6240
1690	0.6284
1691	0.6152
1692	0.6260
1693	0.6142
1694	0.6231
1695	0.6130
1696	0.6146
1697	0.6219
1698	0.6335
1699	0.6225
1700	0.6133
1701	0.6423
1702	0.6191
1703	0.6354
1704	0.6211
1705	0.6271
1706	0.6193
1707	0.6185
1708	0.6192
1709	0.6190
1710	0.6284
1711	0.6355
1712	0.6234
1713	0.6148
1714	0.6032
1715	0.5952
1716	0.6272
1717	0.6302
1718	0.6323
1719	0.6124
1720	0.6149
1721	0.6183
1722	0.6156
1723	0.6246
1724	0.6137
1725	0.6249
1726	0.6370
1727	0.6355
1728	0.6331
1729	0.6295
1730	0.6329
1731	0.6307
1732	0.6311
1733	0.6182
1734	0.6186
1735	0.6325
1736	0.6307
1737	0.6188
1738	0.6272
1739	0.6247
1740	0.6170
1741	0.6211
1742	0.6395
1743	0.6224
1744	0.6072
1745	0.6268
1746	0.6246
1747	0.6150
1748	0.6148
1749	0.6159
1750	0.6180
1751	0.6362
1752	0.6137
1753	0.6102
1754	0.6063
1755	0.6232
1756	0.6209
1757	0.6310
1758	0.6220
1759	0.6190
1760	0.6248
1761	0.6274
1762	0.6240
1763	0.6196
1764	0.6263
1765	0.6182
1766	0.6166
1767	0.6230
1768	0.6315
1769	0.6122
1770	0.6269
1771	0.6067
1772	0.6093
1773	0.6259
1774	0.6224
1775	0.6352
1776	0.6266
1777	0.6111
1778	0.6189
1779	0.6123
1780	0.6266
1781	0.6241
1782	0.6186
1783	0.6085
1784	0.6105
1785	0.6014
1786	0.6295
1787	0.6342
1788	0.6327
1789	0.6337
1790	0.6192
1791	0.6294
1792	0.6214
1793	0.6299
1794	0.6165
1795	0.6179
1796	0.6313
1797	0.6305
1798	0.6311
1799	0.6360
1800	0.6239
1801	0.6330
1802	0.6250
1803	0.6270
1804	0.6186
1805	0.6285
1806	0.6039
1807	0.6340
1808	0.6314
1809	0.6069
1810	0.6305
1811	0.6236
1812	0.6198
1813	0.6029
1814	0.6169
1815	0.6212
1816	0.6214
1817	0.6118
1818	0.6165
1819	0.6004
1820	0.6236
1821	0.6285
1822	0.6107
1823	0.6278
1824	0.6258
1825	0.6239
1826	0.6046
1827	0.5988
1828	0.6118
1829	0.6111
1830	0.5964
1831	0.6061
1832	0.6045
1833	0.6178
1834	0.6186
1835	0.6239
1836	0.6339
1837	0.6189
1838	0.6275
1839	0.6229
1840	0.6223
1841	0.6277
1842	0.6303
1843	0.6229
1844	0.6175
1845	0.6182
1846	0.6250
1847	0.6141
1848	0.6210
1849	0.6021
1850	0.6216
1851	0.6113
1852	0.6249
1853	0.6122
1854	0.6236
1855	0.6068
1856	0.6199
1857	0.6132
1858	0.5974
1859	0.6196
1860	0.6133
1861	0.6114
1862	0.6054
1863	0.6241
1864	0.6244
1865	0.6293
1866	0.6156
1867	0.6127
1868	0.6214
1869	0.6117
1870	0.6239
1871	0.6158
1872	0.6087
1873	0.6137
1874	0.5999
1875	0.6119
1876	0.6250
1877	0.6126
1878	0.6167
1879	0.6009
1880	0.6255
1881	0.6216
1882	0.6024
1883	0.6025
1884	0.6332
1885	0.6055
1886	0.6098
1887	0.6110
1888	0.6221
1889	0.6189
1890	0.6215
1891	0.6289
1892	0.6307
1893	0.6206
1894	0.6259
1895	0.6139
1896	0.6276
1897	0.6004
1898	0.6214
1899	0.6233
1900	0.6061
1901	0.5980
1902	0.6129
1903	0.6077
1904	0.6198
1905	0.6127
1906	0.6134
1907	0.6224
1908	0.6163
1909	0.6233
1910	0.6243
1911	0.6092
1912	0.6085
1913	0.6184
1914	0.6121
1915	0.6157
1916	0.6111
1917	0.6209
1918	0.6213
1919	0.6088
1920	0.6234
1921	0.6208
1922	0.6258
1923	0.6156
1924	0.6257
1925	0.6185
1926	0.6279
1927	0.6200
1928	0.6190
1929	0.6152
1930	0.6127
1931	0.6255
1932	0.6135
1933	0.6184
1934	0.6210
1935	0.6127
1936	0.6208
1937	0.6256
1938	0.6260
1939	0.6078
1940	0.6276
1941	0.6277
1942	0.6345
1943	0.5981
1944	0.6277
1945	0.6436
1946	0.6151
1947	0.6158
1948	0.6277
1949	0.6251
1950	0.6138
1951	0.6239
1952	0.6126
1953	0.6138
1954	0.6256
1955	0.6131
1956	0.6285
1957	0.6232
1958	0.6338
1959	0.6219
1960	0.6208
1961	0.5935
1962	0.5959
1963	0.6184
1964	0.6174
1965	0.6163
1966	0.6395
1967	0.6234
1968	0.6372
1969	0.6283
1970	0.6231
1971	0.5932
1972	0.6152
1973	0.5985
1974	0.5926
1975	0.6105
1976	0.6001
1977	0.5983
1978	0.6218
1979	0.6129
1980	0.6252
1981	0.6237
1982	0.6349
1983	0.6207
1984	0.6181
1985	0.6104
1986	0.6035
1987	0.6230
1988	0.6139
1989	0.6182
1990	0.6124
1991	0.6300
1992	0.6224
1993	0.6095
1994	0.6058
1995	0.6083
1996	0.5967
1997	0.6115
1998	0.5930
1999	0.6072
2000	0.5987
2001	0.6105
2002	0.6326
2003	0.6183
2004	0.6242
2005	0.6240
2006	0.6209
2007	0.6269
2008	0.6191
2009	0.6387
2010	0.6272
2011	0.6122
2012	0.6159
2013	0.6120
2014	0.6157
2015	0.6006
2016	0.6142
2017	0.5873
2018	0.6137
2019	0.6322
2020	0.6106
2021	0.6320
2022	0.6375
2023	0.6256
2024	0.6284
2025	0.6332
2026	0.6068
2027	0.6220
2028	0.5948
2029	0.6295
2030	0.6075
2031	0.6393
2032	0.6243
2033	0.6332
2034	0.6090
2035	0.6115
2036	0.6278
2037	0.6001
2038	0.6170
2039	0.5969
2040	0.5969
2041	0.6115
2042	0.6027
2043	0.6162
2044	0.6235
2045	0.6234
2046	0.6437
2047	0.6378
2048	0.6240
2049	0.6271
2050	0.6167
2051	0.6106
2052	0.6148
2053	0.6132
2054	0.6052
2055	0.6127
2056	0.6056
2057	0.6137
2058	0.6227
2059	0.6149
2060	0.6318
2061	0.6327
2062	0.6234
2063	0.6227
2064	0.6125
2065	0.6158
2066	0.6112
2067	0.6125
2068	0.6113
2069	0.6082
2070	0.6331
2071	0.6187
2072	0.6121
2073	0.6133
2074	0.6131
2075	0.6193
2076	0.6066
2077	0.6290
2078	0.6443
2079	0.6125
2080	0.6216
2081	0.6238
2082	0.6187
2083	0.6121
2084	0.6140
2085	0.6016
2086	0.6275
2087	0.6176
2088	0.6046
2089	0.6105
2090	0.6284
2091	0.6172
2092	0.6330
2093	0.6082
2094	0.6160
2095	0.6078
2096	0.6131
2097	0.5958
2098	0.5983
2099	0.5996
2100	0.6087
2101	0.5976
2102	0.6052
2103	0.6029
2104	0.6228
2105	0.5990
2106	0.6127
2107	0.6085
2108	0.6122
2109	0.5930
2110	0.6114
2111	0.6073
2112	0.6038
2113	0.6099
2114	0.6210
2115	0.6106
2116	0.6231
2117	0.5987
2118	0.6106
2119	0.5974
2120	0.5992
2121	0.6013
2122	0.6109
2123	0.6101
2124	0.6107
2125	0.6153
2126	0.6118
2127	0.5922
2128	0.6180
2129	0.5998
2130	0.5960
2131	0.6176
2132	0.6079
2133	0.6150
2134	0.6002
2135	0.5960
2136	0.6181
2137	0.6046
2138	0.6165
2139	0.6034
2140	0.6254
2141	0.6193
2142	0.6317
2143	0.6231
2144	0.5991
2145	0.6252
2146	0.6227
2147	0.6078
2148	0.6038
2149	0.5959
2150	0.6162
2151	0.5969
2152	0.6006
2153	0.6046
2154	0.6205
2155	0.6277
2156	0.6334
2157	0.6281
2158	0.6130
2159	0.6303
2160	0.6370
2161	0.6062
2162	0.5966
2163	0.6169
2164	0.6137
2165	0.6107
2166	0.6215
2167	0.6046
2168	0.6339
2169	0.6211
2170	0.6280
2171	0.6263
2172	0.6271
2173	0.6212
2174	0.6036
2175	0.6090
2176	0.5995
2177	0.6175
2178	0.6100
2179	0.6173
2180	0.6151
2181	0.6249
2182	0.6108
2183	0.6087
2184	0.6205
2185	0.6024
2186	0.5941
2187	0.6000
2188	0.6036
2189	0.6088
2190	0.6095
2191	0.6127
2192	0.6261
2193	0.6161
2194	0.6075
2195	0.6197
2196	0.6034
2197	0.6089
2198	0.5934
2199	0.6120
2200	0.6089
2201	0.6135
2202	0.6205
2203	0.6251
2204	0.6150
2205	0.6114
2206	0.6285
2207	0.5998
2208	0.5931
2209	0.5981
2210	0.6023
2211	0.6126
2212	0.6153
2213	0.6161
2214	0.6244
2215	0.6136
2216	0.6337
2217	0.6224
2218	0.6162
2219	0.6459
2220	0.6138
2221	0.6158
2222	0.6100
2223	0.6110
2224	0.6151
2225	0.6018
2226	0.5994
2227	0.6119
2228	0.6185
2229	0.6163
2230	0.6177
2231	0.6152
2232	0.6120
2233	0.6296
2234	0.6162
2235	0.5940
2236	0.6065
2237	0.6106
2238	0.6049
2239	0.6157
2240	0.6070
2241	0.6064
2242	0.6128
2243	0.6052
2244	0.6063
2245	0.6060
2246	0.6086
2247	0.6122
2248	0.6050
2249	0.5949
2250	0.5968
2251	0.5971
2252	0.6029
2253	0.6041
2254	0.6194
2255	0.6252
2256	0.6127
2257	0.6188
2258	0.6171
2259	0.6220
2260	0.6087
2261	0.6157
2262	0.6065
2263	0.6155
2264	0.5920
2265	0.6030
2266	0.5976
2267	0.6018
2268	0.6020
2269	0.6145
2270	0.5894
2271	0.5877
2272	0.6240
2273	0.6050
2274	0.6132
2275	0.6119
2276	0.6013
2277	0.6124
2278	0.6075
2279	0.6161
2280	0.6191
2281	0.6239
2282	0.6006
2283	0.6155
2284	0.6238
2285	0.6182
2286	0.6116
2287	0.5988
2288	0.6092
2289	0.6012
2290	0.6116
2291	0.6224
2292	0.6212
2293	0.6316
2294	0.6300
2295	0.6262
2296	0.5969
2297	0.6065
2298	0.6154
2299	0.6188
2300	0.6155
2301	0.6238
2302	0.6130
2303	0.6253
2304	0.6124
2305	0.6205
2306	0.6009
2307	0.6150
2308	0.6085
2309	0.6130
2310	0.6133
2311	0.5937
2312	0.6024
2313	0.6050
2314	0.6138
2315	0.6233
2316	0.6174
2317	0.6151
2318	0.6197
2319	0.6048
2320	0.6073
2321	0.6003
2322	0.6145
2323	0.6201
2324	0.6072
2325	0.6023
2326	0.6317
2327	0.6289
2328	0.6234
2329	0.6154
2330	0.6068
2331	0.6065
2332	0.6270
2333	0.6038
2334	0.6080
2335	0.5919
2336	0.6142
2337	0.5999
2338	0.6103
2339	0.6234
2340	0.6075
2341	0.6390
2342	0.6100
2343	0.6194
2344	0.6374
2345	0.6298
2346	0.6008
2347	0.6175
2348	0.5985
2349	0.6099
2350	0.6087
2351	0.6147
2352	0.5982
2353	0.6042
2354	0.6055
2355	0.6146
2356	0.6102
2357	0.6143
2358	0.6157
2359	0.6215
2360	0.6184
2361	0.6086
2362	0.6117
2363	0.6203
2364	0.6037
2365	0.6055
2366	0.6231
2367	0.6163
2368	0.6279
2369	0.6054
2370	0.6162
2371	0.6060
2372	0.5940
2373	0.6076
2374	0.6046
2375	0.6118
2376	0.6008
2377	0.6117
2378	0.6160
2379	0.6424
2380	0.6320
2381	0.6150
2382	0.6106
2383	0.6189
2384	0.6135
2385	0.6117
2386	0.5996
2387	0.5986
2388	0.6178
2389	0.6128
2390	0.6261
2391	0.6011
2392	0.6122
2393	0.6156
2394	0.6208
2395	0.6125
2396	0.6121
2397	0.6075
2398	0.6100
2399	0.6096
2400	0.6344
2401	0.6062
2402	0.6192
2403	0.6087
2404	0.6122
2405	0.6114
2406	0.6103
2407	0.6090
2408	0.6163
2409	0.6078
2410	0.6083
2411	0.6109
2412	0.5959
2413	0.6062
2414	0.6024
2415	0.6041
2416	0.5915
2417	0.5915
2418	0.5991
2419	0.5950
2420	0.6121
2421	0.6120
2422	0.6161
2423	0.6135
2424	0.6111
2425	0.6138
2426	0.6003
2427	0.5992
2428	0.5802
2429	0.6086
2430	0.5980
2431	0.6145
2432	0.6091
2433	0.6138
2434	0.6064
2435	0.6105
2436	0.6160
2437	0.6116
2438	0.6082
2439	0.6126
2440	0.6021
2441	0.6054
2442	0.6030
2443	0.6023
2444	0.5964
2445	0.6029
2446	0.6182
2447	0.6214
2448	0.6150
2449	0.6400
2450	0.5968
2451	0.6211
2452	0.6164
2453	0.6142
2454	0.6129
2455	0.5893
2456	0.5918
2457	0.6039
2458	0.6014
2459	0.6100
2460	0.6224
2461	0.6290
2462	0.6135
2463	0.6097
2464	0.6108
2465	0.6052
2466	0.6181
2467	0.6163
2468	0.6093
2469	0.6071
2470	0.6147
2471	0.6193
2472	0.6065
2473	0.6035
2474	0.6029
2475	0.6061
2476	0.5906
2477	0.5983
2478	0.6116
2479	0.5929
2480	0.6022
2481	0.6036
2482	0.6018
2483	0.5984
2484	0.6087
2485	0.6275
2486	0.6042
2487	0.6125
2488	0.6097
2489	0.6034
2490	0.6369
2491	0.6177
2492	0.6077
2493	0.6107
2494	0.6239
2495	0.6179
2496	0.6040
2497	0.6022
2498	0.5940
2499	0.6046
2500	0.6187
2501	0.6144
2502	0.5982
2503	0.6270
2504	0.6068
2505	0.6145
2506	0.6283
2507	0.6123
2508	0.6099
2509	0.6036
2510	0.5989
2511	0.6071
2512	0.6311
2513	0.6182
2514	0.6214
2515	0.6214
2516	0.6370
2517	0.6217
2518	0.6082
2519	0.6282
2520	0.6044
2521	0.5947
2522	0.6255
2523	0.6006
2524	0.6234
2525	0.6026
2526	0.5960
2527	0.6085
2528	0.5929
2529	0.5927
2530	0.6172
2531	0.6050
2532	0.6024
2533	0.5970
2534	0.6087
2535	0.6081
2536	0.5960
2537	0.6164
2538	0.6155
2539	0.6063
2540	0.6023
2541	0.5941
2542	0.5969
2543	0.5907
2544	0.5896
2545	0.5980
2546	0.5962
2547	0.6101
2548	0.6022
2549	0.6057
2550	0.6099
2551	0.6160
2552	0.6170
2553	0.6105
2554	0.6175
2555	0.6019
2556	0.6079
2557	0.5993
2558	0.6038
2559	0.5988
2560	0.6132
2561	0.6121
2562	0.6052
2563	0.6098
2564	0.5969
2565	0.5972
2566	0.6183
2567	0.6063
2568	0.6100
2569	0.6102
2570	0.5819
2571	0.6094
2572	0.6042
2573	0.6038
2574	0.5957
2575	0.6052
2576	0.5944
2577	0.6146
2578	0.6087
2579	0.5947
2580	0.6017
2581	0.6189
2582	0.6147
2583	0.6018
2584	0.5912
2585	0.5896
2586	0.6065
2587	0.6057
2588	0.6196
2589	0.6242
2590	0.6028
2591	0.6142
2592	0.6121
2593	0.6006
2594	0.5860
2595	0.6083
2596	0.6015
2597	0.6136
2598	0.6131
2599	0.6019
2600	0.6285
2601	0.6300
2602	0.6120
2603	0.6121
2604	0.6128
2605	0.6000
2606	0.5925
2607	0.6249
2608	0.6231
2609	0.6158
2610	0.6177
2611	0.6206
2612	0.6071
2613	0.6153
2614	0.6012
2615	0.6036
2616	0.5948
2617	0.6152
2618	0.6039
2619	0.6234
2620	0.6241
2621	0.6189
2622	0.6251
2623	0.6244
2624	0.6240
2625	0.5892
2626	0.6171
2627	0.6138
2628	0.6146
2629	0.5948
2630	0.6119
2631	0.6006
2632	0.6070
2633	0.6221
2634	0.6223
2635	0.6304
2636	0.6042
2637	0.6116
2638	0.6072
2639	0.6101
2640	0.5953
2641	0.6000
2642	0.6152
2643	0.6027
2644	0.6121
2645	0.6024
2646	0.6034
2647	0.5932
2648	0.6182
2649	0.5914
2650	0.6226
2651	0.6092
2652	0.6097
2653	0.6263
2654	0.6015
2655	0.6197
2656	0.6080
2657	0.6082
2658	0.6030
2659	0.6142
2660	0.5930
2661	0.6058
2662	0.6013
2663	0.5959
2664	0.5956
2665	0.6078
2666	0.6060
2667	0.6123
2668	0.6135
2669	0.6050
2670	0.5988
2671	0.6212
2672	0.5934
2673	0.5935
2674	0.6065
2675	0.6015
2676	0.6105
2677	0.6059
2678	0.6003
2679	0.6179
2680	0.6188
2681	0.6069
2682	0.6066
2683	0.6109
2684	0.6063
2685	0.5996
2686	0.6130
2687	0.5902
2688	0.6008
2689	0.6152
2690	0.6231
2691	0.6208
2692	0.6205
2693	0.6177
2694	0.5987
2695	0.6026
2696	0.5982
2697	0.5802
2698	0.5997
2699	0.6082
2700	0.6004
2701	0.6035
2702	0.6047
2703	0.6102
2704	0.6051
2705	0.6185
2706	0.6026
2707	0.6080
2708	0.6228
2709	0.6124
2710	0.6165
2711	0.5970
2712	0.5996
2713	0.5919
2714	0.6047
2715	0.5948
2716	0.6011
2717	0.5928
2718	0.6219
2719	0.6073
2720	0.6158
2721	0.6144
2722	0.6167
2723	0.6102
2724	0.6269
2725	0.6174
2726	0.5962
2727	0.5970
2728	0.6034
2729	0.5941
2730	0.5923
2731	0.5878
2732	0.5943
2733	0.6088
2734	0.6173
2735	0.6169
2736	0.5962
2737	0.6070
2738	0.5831
2739	0.6060
2740	0.6011
2741	0.6025
2742	0.6106
2743	0.6077
2744	0.6112
2745	0.6174
2746	0.6036
2747	0.6183
2748	0.6118
2749	0.6070
2750	0.6057
2751	0.6002
2752	0.6075
2753	0.6051
2754	0.6000
2755	0.6055
2756	0.6159
2757	0.6272
2758	0.6133
2759	0.5990
2760	0.6181
2761	0.6148
2762	0.6105
2763	0.6041
2764	0.6133
2765	0.5987
2766	0.6159
2767	0.5859
2768	0.6203
2769	0.5965
2770	0.5982
2771	0.6089
2772	0.6045
2773	0.6080
2774	0.6123
2775	0.6174
2776	0.5942
2777	0.6079
2778	0.6109
2779	0.6061
2780	0.6303
2781	0.6106
2782	0.6036
2783	0.6129
2784	0.6229
2785	0.6025
2786	0.6104
2787	0.6114
2788	0.6216
2789	0.6044
2790	0.5979
2791	0.6249
2792	0.6073
2793	0.6151
2794	0.6302
2795	0.6164
2796	0.6129
2797	0.6111
2798	0.6026
2799	0.6021
2800	0.5997
2801	0.5930
2802	0.6070
2803	0.6160
2804	0.5957
2805	0.6173
2806	0.6087
2807	0.6047
2808	0.5827
2809	0.5987
2810	0.6067
2811	0.5994
2812	0.6056
2813	0.6142
2814	0.6157
2815	0.5938
2816	0.6039
2817	0.6109
2818	0.5980
2819	0.5987
2820	0.6014

Measured result

Verbatim from the run's summary.json.

bucket

bucket-p32-t128

bucket_icons

3359

checkpoint_bytes

4715161

checkpoint_round_trip

true

checkpoint_sha256

cff49ce8e8f2988d…

config_sha256

0ef00ff401f00797…

corruption

factorized_marginal_respecting_geometry

cuda_version

13.4

deterministic_algorithms

true

device

cuda

eval_every

final_train

loss

0.6135861277580261

step

2820

train_token_accuracy

0.6014218009478673

group_vocabulary_size

locked_path_exact

true

metrics_sha256

9276f6ee9ed29e00…

model_parameters

579872

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

2820

evals_without_improvement

min_delta

0.0

objective

held_out_loss

patience_evals

selected_held_out_loss

0.6305077075958252

selected_step

2340

stopped_early

true

steps

6300

study_version

openmoji-g1-marginal-corruption-v9

subgroup_vocabulary_size

118

torch_version

2.14.0a0+4fdf77b940.nv26.08

validation

accuracy

0.6025794841031794

changed_accuracy

0.0002869234631662004

changed_total

13941

loss

0.6305077075958252

retained_accuracy

0.9246940576207465

retained_total

26067

validation_final_step

accuracy

0.5701359728054389

changed_accuracy

0.0005738469263324008

changed_total

13941

loss

0.6339353919029236

retained_accuracy

0.8747458472398051

retained_total

26067

validation_trace_sha256

69b9d3422e328aad…

validation_untrained

accuracy

0.2675214957008598

changed_accuracy

0.002080195107954953

changed_total

13941

loss

1.4054925441741943

retained_accuracy

0.4094832546898377

retained_total

26067

Written result

v8 reached 2.806 precision lift, which looked like a breakthrough until a judge pointed out that a detector knowing only the corpus token marginal reaches 2.831 on the same data with zero parameters and zero context. Factorized corruption draws replacements uniformly from the full legal vocabulary, so a large share land on tokens real icons essentially never use, and spotting them needs no geometry at all.

This run removes that shortcut. Replacements are drawn from the corpus distribution for the field's own role — keyed by segment kind as well as slot, since slot 0 is an endpoint for a LINE and a control handle for a CUBIC — so every corrupted token is in-distribution by construction. The marginal is estimated from train programs only. Nothing else changes from v8.

Measured on 128 held-out icons before the run, at the matched 8.573% flag rate:

free detector	under factorized	under marginal
marginal density (0 params, no context)	2.840	0.771
local continuity (0 params, neighbours)	2.126	1.893

The density shortcut is gone — below chance, because corrupted tokens are now typical — while the genuine geometric signal survives.

Result

criterion	threshold	observed	outcome
beats_the_free_geometric_detector	> 1.893	1.506	falsified
not_explained_by_density	> 1.20	1.506	pass
structural_safety	locked-path exact, round-trips	both true	pass
reproducibility	identical rerun	identical	pass

Selected step 2,340 of 2,820 run, early stopped — against v8's full 6,300 with no stop. The harder task converges sooner.

What 1.506 means

It is the first honest detection number this project has produced, and it says two things at once.

The model does learn geometry. At 1.506 it is far above the 0.771 a density detector manages here, and above the 1.20 bar set to rule that out. Whatever it is using, it is not the marginal, because the marginal no longer helps.

It still loses to a zero-parameter statistic. The local-continuity heuristic reaches 1.893 on the identical corrupted inputs. A 579,872-parameter transformer, trained under sound optimisation for 2,340 steps, extracts less of the geometric signal than the mean absolute distance to the same slot in adjacent segments.

And v8's 2.806 is now properly read: mostly the leak. The honest ranking under leak-free corruption is continuity 1.893, v9 1.506, density 0.771.

What this settles and what it opens

Settled: the corruption process was leaking, v8's apparent success was largely that leak, and the fix is real and cheap. Every detector number measured under factorized corruption — v6's 1.234, v7's 1.365, v8's 2.806, and the 2.13 and 2.83 free references — describes a task that is partly a density test. This run's 1.506 replaces them as the reference point.

Opened, and exactly as this run's predeclared falsification meaning anticipated: the model cannot beat a trivial geometric statistic even with the shortcut removed and the optimisation fixed, so the remaining candidates are representational. The diagnostic panel measured two, independently:

coordinate values enter as identities, not magnitudes. The trained input embedding is statistically indistinguishable from random initialisation — I confirmed Spearman −0.136 between embedding distance and bin distance, adjacent-bin distance 0.989 of the all-pairs mean — while the output head, where the corpus marginal supplies a direct gradient, does show order at +0.143 and 0.882. The continuity statistic works by subtracting decoded view-unit values, which is precisely the operation a table of 418 unordered categories cannot support.
six fields are summed into one 96-dim slot. About 16 dimensions per field, before attention runs.

The representation lens measured the first in a controlled simulation: with the current encoding a probe on the neighbour-difference target scores AUC 0.5024, chance; replacing the categorical lookup with a numeric code of the same value gives 0.9981. That is the next run, and it is a single factor.

Scope

Detection-only, single seed, probability 0.35. Not comparable with v8 in absolute terms: the task is strictly harder because the easy fraction of corruptions was removed, so 1.506 against 2.806 is not a regression.

State transitions

planned2026-09-21T00:40:00Z
completed2026-09-21T01:00:00Zleak_closed_the_model_learns_geometry_but_still_loses_to_a_zero_parameter_continuity_statistic

Run record

Verbatim from runs/openmoji-g1-marginal-corruption-v9-74dc5fd-0ef00ff4-9b9b1699/run.yaml, the record committed before launch.

schema_version

run_id

openmoji-g1-marginal-corruption-v9-74dc5fd-0ef00ff4-9b9b1699

state

completed

parent_run

openmoji-g1-sound-optimisation-v8-949e0e5-4feee442-9b9b1699

hypothesis

v8's detector reached 2.806 precision lift, but a zero-parameter detector that knows only the corpus marginal reaches 2.831 on the same data, so v8 matched a density test rather than learning geometry. Factorized corruption draws replacements uniformly from the full legal vocabulary, so many land on tokens real icons never use. Drawing them from the corpus distribution for the field's role instead removes that shortcut: the marginal detector falls to 0.771 while the local-continuity signal survives at 1.893. Under this leak-free process the model must use the surrounding geometry, and this run measures for the first time how much of that signal it can actually learn.

expected_information_gain

scope_limits

changed_factors

primary

training.corruption_process factorized -> marginal.

parameters

unchanged at 579,872

held_fixed

v8 exactly otherwise: pool_loss_over_fields, mask_padding, detection_only, slot binding, edit-mask heads, the full 2,681-icon split, the 128-icon validation draw and its corruption seeds, seed 3101, probability 0.35, batch 16, lr 0.001, eval_every 60, the 6,300-step cap, held-out-loss selection with patience 8.

marginal_estimated_from

train programs only, so no held-out information reaches the corruption applied to held-out icons.

code

git_commit

74dc5fd

execution_mode

native-local

config

path

configs/learning/openmoji-g1-marginal-corruption-v9.yaml

sha256

0ef00ff401f00797…

dataset

hybrid_sha256

9b9b1699677a6f97…

bucket

bucket-p32-t128

train_samples

2681

validation_samples

128

free_references_under_each_process

note

Measured before this run, on 128 held-out icons, at the matched 8.573% flag rate. The marginal column is what this run must be judged against.

factorized_uniform

marginal_density_lift

2.84

local_continuity_lift

2.126

marginal_respecting

marginal_density_lift

0.771

local_continuity_lift

1.893

trained_under_factorized

1.234

1.365

2.806

predeclared_criteria

note

Measured by scripts/detector_report.py at the matched 8.573% flag rate.

evaluated_at

the checkpoint selected by held-out loss

primary

beats_the_free_geometric_detector

statement

Precision lift exceeds 1.893, the local-continuity statistic measured under this same corruption process. This is the first criterion in the project that a model can only pass by using geometry, because the density shortcut is gone.

not_explained_by_density

statement

Precision lift exceeds 1.20. The marginal-density detector scores 0.771 here, so anything at or below about 1.0 would mean the model learned nothing this process did not hand it.

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

detector recall and precision at the model's own threshold and at the base flag rate, the selected step and whether early stopping or the cap ended the run, held-out field-pooled loss, not comparable across corruption processes

falsification_meaning

If the model cannot beat a zero-parameter continuity statistic once the density shortcut is removed, then v8's apparent success was entirely the leak, and the conclusion is that this model class does not learn the geometric relation even under sound optimisation. That would make the representational interventions - per-field tokenisation, or ordinal coordinate features - the next runs, and both were measured and argued for by the diagnostic panel.

outputs

local_metadata

runs/openmoji-g1-marginal-corruption-v9-74dc5fd-0ef00ff4-9b9b1699

report_root

reports/learning/openmoji-g1-marginal-corruption-v9

durable_artifacts

/home/dev/.cache/openmoji-g1-marginal-corruption-v9-74dc5fd-0ef00ff4-9b9b1699

planned_at

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

completed_at

2026-09-21 01:00:00+00:00

result

device

cuda

model_parameters

579872

selected_step

2340

completed_steps

2820

ended_by

early_stopping

identical_rerun

true

detector_at_matched_flag_rate

precision

0.5248

lift_over_base_rate

1.506

recall

0.1291

detector_at_base_flag_rate

precision

0.4295

lift_over_base_rate

1.232

predeclared_outcome

overall

partially-falsified

beats_the_free_geometric_detector

passed

false

observed

1.506

threshold

1.893

not_explained_by_density

passed

true

observed

1.506

threshold

1.2

structural_safety

passed

true

reproducibility

passed

true

conclusion

The first honest detection number in the project, and it says two things. The model does learn geometry: 1.506 is far above the 0.771 a density detector manages under this process, so it is not using the marginal, which no longer helps. But it still loses to a zero-parameter local-continuity statistic at 1.893 on the identical corrupted inputs, under sound optimisation and 2,340 steps.

reinterprets_v8

v8's 2.806 was mostly the leak. The honest ranking under leak-free corruption is continuity 1.893, v9 1.506, density 0.771. Every detector number measured under factorized corruption - v6 1.234, v7 1.365, v8 2.806, and the 2.13 and 2.83 free references - describes a task that is partly a density test.

opens

The predeclared falsification meaning applies: the model cannot beat a trivial geometric statistic with the shortcut removed and the optimisation fixed, so the remaining candidates are representational. The measured one is that coordinate values enter as identities rather than magnitudes - the trained input embedding is statistically indistinguishable from random init at Spearman -0.136 and adjacent-bin distance 0.989 of the all-pairs mean, while the output head, which gets a direct gradient from the corpus marginal, does show order at +0.143 and 0.882.