openmoji-g1-noise-conditioned-v4-6b935b2-5a9eaf44-9b9b1699

completed falsified

Hypothesis

The Gate G denoiser produced output near-independent of its input because it trained at one fixed corruption level and was never told the level, so it had no reason to learn how much to trust `x_t` and settled on a group-conditioned prior. Sampling a corruption level per example across 0.05 to 0.50 and conditioning the model on that level makes the output actually depend on the input: the model stops destroying lightly corrupted inputs, and its predictions at 0.05 and 0.35 stop being nearly the same picture.

Why run it

This is the first Gate G run that targets the diagnosed mechanism rather than a resource. If it works, the earlier recovery numbers become interpretable as denoising for the first time and the gate has a path forward. If it does not, the failure is not about noise-level information and the next candidates are an edit-mask or residual objective, which are larger changes.

Scope limits

Fixed-topology and geometry-only, single seed. Model size is v2's, plus 3,168 parameters of noise projection, because v3 showed capacity is not the constraint. Success here would mean the model reads its input, not that its output is usable; the renders have never yet produced a recognizable icon at any setting.

Measured behaviour

Held-out loss

Table view

optimizer step	held-out loss
0	16.1460
60	10.8750
120	9.4844
180	8.5912
240	8.3136
300	8.1153
360	7.8407
420	7.8147
480	7.6312
540	7.5708
600	7.5198
660	7.5473
720	7.5784
780	7.5604
840	7.4843
900	7.6190
960	7.7106
1020	7.4285
1080	7.4970
1140	7.5370
1200	7.5565
1260	7.6018
1320	7.4705
1380	7.5737
1440	7.6598
1500	7.6000

Held-out token accuracy

aggregatechanged fieldsretained fields

Table view

optimizer step	aggregate	changed fields	retained fields
0	0.0028	0.0031	0.0027
60	0.0704	0.0218	0.0965
120	0.1579	0.0318	0.2256
180	0.2099	0.0342	0.3043
240	0.2314	0.0361	0.3363
300	0.2510	0.0395	0.3646
360	0.2627	0.0395	0.3826
420	0.2670	0.0426	0.3875
480	0.2723	0.0411	0.3964
540	0.2793	0.0429	0.4063
600	0.2843	0.0447	0.4130
660	0.2847	0.0476	0.4120
720	0.2865	0.0460	0.4157
780	0.2900	0.0512	0.4183
840	0.2929	0.0521	0.4222
900	0.2922	0.0537	0.4202
960	0.2962	0.0553	0.4256
1020	0.2980	0.0556	0.4282
1080	0.2981	0.0577	0.4272
1140	0.3010	0.0600	0.4305
1200	0.3015	0.0567	0.4329
1260	0.3006	0.0591	0.4302
1320	0.3024	0.0595	0.4328
1380	0.3034	0.0597	0.4343
1440	0.3026	0.0618	0.4319
1500	0.3064	0.0613	0.4380

Training token accuracy (per batch)

Table view

optimizer step	train accuracy
1	0.0029
2	0.0044
3	0.0035
4	0.0037
5	0.0036
6	0.0045
7	0.0040
8	0.0046
9	0.0060
10	0.0065
11	0.0065
12	0.0063
13	0.0076
14	0.0083
15	0.0095
16	0.0143
17	0.0111
18	0.0167
19	0.0179
20	0.0179
21	0.0196
22	0.0163
23	0.0265
24	0.0199
25	0.0259
26	0.0388
27	0.0244
28	0.0329
29	0.0428
30	0.0324
31	0.0352
32	0.0384
33	0.0269
34	0.0610
35	0.0471
36	0.0479
37	0.0238
38	0.0504
39	0.0415
40	0.0378
41	0.0702
42	0.0623
43	0.0295
44	0.0666
45	0.0591
46	0.0368
47	0.0573
48	0.0719
49	0.0654
50	0.0584
51	0.0840
52	0.0739
53	0.0591
54	0.0634
55	0.1127
56	0.0523
57	0.0986
58	0.1021
59	0.1196
60	0.1020
61	0.0822
62	0.0933
63	0.1392
64	0.1038
65	0.1248
66	0.0895
67	0.0699
68	0.1544
69	0.1052
70	0.1154
71	0.1612
72	0.1390
73	0.1349
74	0.1248
75	0.1030
76	0.1185
77	0.1113
78	0.1531
79	0.1175
80	0.1194
81	0.1335
82	0.1475
83	0.1699
84	0.1272
85	0.1935
86	0.1589
87	0.1449
88	0.1468
89	0.1796
90	0.1971
91	0.2131
92	0.1719
93	0.1265
94	0.1275
95	0.1547
96	0.1901
97	0.1677
98	0.1325
99	0.1403
100	0.2099
101	0.2220
102	0.1999
103	0.1696
104	0.2259
105	0.2027
106	0.1819
107	0.1739
108	0.2335
109	0.2451
110	0.1802
111	0.2277
112	0.1797
113	0.1473
114	0.1926
115	0.1739
116	0.1726
117	0.2135
118	0.1924
119	0.2492
120	0.1873
121	0.1681
122	0.2241
123	0.2312
124	0.1362
125	0.2676
126	0.1705
127	0.2153
128	0.2920
129	0.2525
130	0.3645
131	0.1510
132	0.2674
133	0.2234
134	0.2114
135	0.2060
136	0.3079
137	0.2092
138	0.2881
139	0.2073
140	0.2606
141	0.2484
142	0.2287
143	0.2649
144	0.2101
145	0.2192
146	0.2294
147	0.2883
148	0.2872
149	0.2332
150	0.2951
151	0.2519
152	0.2362
153	0.2720
154	0.2161
155	0.2736
156	0.2137
157	0.2268
158	0.2805
159	0.2687
160	0.1992
161	0.2211
162	0.2850
163	0.2538
164	0.2995
165	0.3099
166	0.2077
167	0.2438
168	0.3496
169	0.1858
170	0.2960
171	0.3172
172	0.3382
173	0.2189
174	0.2348
175	0.2816
176	0.2766
177	0.3333
178	0.2604
179	0.2424
180	0.4096
181	0.3132
182	0.3129
183	0.3333
184	0.2951
185	0.3009
186	0.3012
187	0.4189
188	0.2988
189	0.2623
190	0.3066
191	0.3190
192	0.3040
193	0.3196
194	0.3023
195	0.3776
196	0.3143
197	0.3275
198	0.3774
199	0.2935
200	0.3287
201	0.2309
202	0.4056
203	0.3137
204	0.3613
205	0.3201
206	0.3053
207	0.2774
208	0.2521
209	0.3487
210	0.3505
211	0.2341
212	0.4244
213	0.2401
214	0.2785
215	0.2833
216	0.3613
217	0.2943
218	0.2048
219	0.3544
220	0.2843
221	0.2817
222	0.2688
223	0.3189
224	0.3038
225	0.2948
226	0.3041
227	0.3546
228	0.3703
229	0.2923
230	0.3666
231	0.2997
232	0.2837
233	0.3147
234	0.3382
235	0.2237
236	0.4065
237	0.3020
238	0.3767
239	0.4429
240	0.3181
241	0.4151
242	0.3217
243	0.3054
244	0.3082
245	0.2835
246	0.3249
247	0.3429
248	0.2820
249	0.2608
250	0.4549
251	0.3730
252	0.3428
253	0.4006
254	0.3180
255	0.3074
256	0.3221
257	0.3785
258	0.3323
259	0.4774
260	0.3828
261	0.2633
262	0.2829
263	0.3314
264	0.2843
265	0.3305
266	0.3206
267	0.3496
268	0.2806
269	0.3788
270	0.3929
271	0.3031
272	0.3941
273	0.3105
274	0.3507
275	0.3230
276	0.3971
277	0.4410
278	0.3220
279	0.4323
280	0.2757
281	0.2724
282	0.3395
283	0.3953
284	0.2946
285	0.3652
286	0.3076
287	0.3936
288	0.2947
289	0.3629
290	0.4297
291	0.3028
292	0.3599
293	0.3029
294	0.3244
295	0.3240
296	0.3928
297	0.4306
298	0.3725
299	0.3502
300	0.3458
301	0.3773
302	0.2907
303	0.4122
304	0.3684
305	0.3622
306	0.3132
307	0.3261
308	0.3484
309	0.3502
310	0.4871
311	0.3223
312	0.3268
313	0.3184
314	0.3833
315	0.3961
316	0.3408
317	0.3071
318	0.4674
319	0.3350
320	0.3713
321	0.4050
322	0.2934
323	0.3718
324	0.2884
325	0.3243
326	0.3850
327	0.3565
328	0.3238
329	0.3405
330	0.3537
331	0.3392
332	0.4093
333	0.3897
334	0.2697
335	0.4632
336	0.3088
337	0.3755
338	0.3515
339	0.5373
340	0.2912
341	0.3052
342	0.3473
343	0.4446
344	0.3796
345	0.3522
346	0.3445
347	0.3378
348	0.3520
349	0.4534
350	0.4433
351	0.4165
352	0.4046
353	0.3560
354	0.5182
355	0.4584
356	0.3429
357	0.3910
358	0.4114
359	0.3120
360	0.4672
361	0.3740
362	0.3010
363	0.4129
364	0.4386
365	0.4019
366	0.4415
367	0.4047
368	0.3116
369	0.4677
370	0.4252
371	0.4256
372	0.3136
373	0.4676
374	0.3527
375	0.3368
376	0.3530
377	0.4717
378	0.3289
379	0.3714
380	0.4422
381	0.3702
382	0.3365
383	0.3756
384	0.4276
385	0.3329
386	0.3983
387	0.3375
388	0.3421
389	0.3846
390	0.4584
391	0.3195
392	0.3701
393	0.3674
394	0.4041
395	0.4317
396	0.4160
397	0.3794
398	0.4707
399	0.3176
400	0.4456
401	0.4222
402	0.3003
403	0.3971
404	0.4023
405	0.3982
406	0.4418
407	0.4117
408	0.3820
409	0.4514
410	0.3465
411	0.3789
412	0.3873
413	0.4479
414	0.4054
415	0.3700
416	0.3879
417	0.4007
418	0.4374
419	0.3843
420	0.4503
421	0.4630
422	0.4263
423	0.3378
424	0.3908
425	0.4095
426	0.5073
427	0.4328
428	0.3439
429	0.2833
430	0.3936
431	0.4076
432	0.4042
433	0.3068
434	0.3557
435	0.4422
436	0.3659
437	0.4177
438	0.3859
439	0.4168
440	0.4882
441	0.4166
442	0.3837
443	0.4184
444	0.4139
445	0.3870
446	0.4627
447	0.3603
448	0.3403
449	0.4304
450	0.4165
451	0.3335
452	0.3740
453	0.3606
454	0.3925
455	0.4096
456	0.3442
457	0.4534
458	0.4475
459	0.3115
460	0.4379
461	0.3184
462	0.3180
463	0.4234
464	0.4624
465	0.4272
466	0.3530
467	0.3678
468	0.4263
469	0.4151
470	0.4119
471	0.4989
472	0.3695
473	0.4552
474	0.3756
475	0.4101
476	0.4664
477	0.5086
478	0.4684
479	0.3218
480	0.3859
481	0.3824
482	0.4116
483	0.4432
484	0.4018
485	0.3966
486	0.4427
487	0.3772
488	0.4155
489	0.3706
490	0.3575
491	0.4223
492	0.3274
493	0.4094
494	0.3752
495	0.4036
496	0.3401
497	0.4409
498	0.4144
499	0.4548
500	0.4193
501	0.3372
502	0.3566
503	0.5299
504	0.3534
505	0.3869
506	0.4782
507	0.4211
508	0.3311
509	0.3396
510	0.4573
511	0.3930
512	0.5195
513	0.3887
514	0.3655
515	0.5255
516	0.3817
517	0.4412
518	0.4806
519	0.4753
520	0.3901
521	0.4302
522	0.5188
523	0.4700
524	0.4304
525	0.4011
526	0.4590
527	0.4807
528	0.4830
529	0.3550
530	0.4390
531	0.4827
532	0.3826
533	0.4344
534	0.4454
535	0.4700
536	0.3139
537	0.4546
538	0.4586
539	0.4104
540	0.4599
541	0.4255
542	0.3783
543	0.3297
544	0.5160
545	0.4170
546	0.3765
547	0.4873
548	0.3006
549	0.5106
550	0.2986
551	0.4302
552	0.4099
553	0.2841
554	0.4833
555	0.3792
556	0.4047
557	0.4789
558	0.3922
559	0.3527
560	0.3901
561	0.4359
562	0.4387
563	0.3842
564	0.4789
565	0.4469
566	0.3980
567	0.4275
568	0.4593
569	0.4339
570	0.3347
571	0.4601
572	0.4282
573	0.4099
574	0.5331
575	0.4030
576	0.5127
577	0.4355
578	0.3353
579	0.4556
580	0.3722
581	0.4650
582	0.4355
583	0.4081
584	0.3867
585	0.5247
586	0.3912
587	0.4916
588	0.5280
589	0.4256
590	0.4274
591	0.4275
592	0.4323
593	0.4563
594	0.5283
595	0.4735
596	0.2893
597	0.3628
598	0.4768
599	0.4160
600	0.4600
601	0.4020
602	0.3210
603	0.4734
604	0.4512
605	0.4455
606	0.4519
607	0.4534
608	0.4834
609	0.4304
610	0.3990
611	0.5216
612	0.5124
613	0.3471
614	0.5426
615	0.4281
616	0.3577
617	0.4535
618	0.4403
619	0.4031
620	0.3712
621	0.3344
622	0.4682
623	0.4009
624	0.3642
625	0.4364
626	0.3771
627	0.3967
628	0.3828
629	0.3676
630	0.3709
631	0.5039
632	0.4812
633	0.5200
634	0.3154
635	0.4824
636	0.4379
637	0.3965
638	0.4907
639	0.5256
640	0.4071
641	0.4764
642	0.4060
643	0.4176
644	0.4572
645	0.5441
646	0.4281
647	0.3792
648	0.4087
649	0.4503
650	0.4346
651	0.4427
652	0.3323
653	0.5217
654	0.4440
655	0.4101
656	0.5165
657	0.3361
658	0.4475
659	0.3503
660	0.3846
661	0.4370
662	0.4264
663	0.3478
664	0.3700
665	0.4101
666	0.4079
667	0.4659
668	0.3958
669	0.3397
670	0.4730
671	0.4117
672	0.4952
673	0.4145
674	0.5039
675	0.4234
676	0.3505
677	0.3459
678	0.4649
679	0.4894
680	0.4614
681	0.3504
682	0.4185
683	0.4233
684	0.4894
685	0.5115
686	0.4450
687	0.4147
688	0.4804
689	0.5467
690	0.4735
691	0.4283
692	0.4773
693	0.4678
694	0.4459
695	0.4423
696	0.4554
697	0.4439
698	0.4520
699	0.4734
700	0.4867
701	0.4611
702	0.4706
703	0.4081
704	0.5062
705	0.4194
706	0.4940
707	0.4480
708	0.4665
709	0.4359
710	0.4170
711	0.3532
712	0.5345
713	0.3252
714	0.4901
715	0.4803
716	0.3431
717	0.4144
718	0.4031
719	0.4806
720	0.4273
721	0.4011
722	0.4558
723	0.3579
724	0.4323
725	0.4478
726	0.4545
727	0.4514
728	0.4420
729	0.4605
730	0.4512
731	0.4721
732	0.4570
733	0.4949
734	0.3447
735	0.5194
736	0.4646
737	0.4342
738	0.4601
739	0.4791
740	0.4439
741	0.4823
742	0.5111
743	0.4570
744	0.4847
745	0.3553
746	0.4054
747	0.4527
748	0.4826
749	0.3897
750	0.4101
751	0.4247
752	0.4606
753	0.5133
754	0.4032
755	0.5169
756	0.5200
757	0.4471
758	0.3904
759	0.4553
760	0.4214
761	0.5064
762	0.5568
763	0.3991
764	0.3316
765	0.4526
766	0.4215
767	0.4893
768	0.4059
769	0.3903
770	0.4290
771	0.5002
772	0.3944
773	0.4587
774	0.4462
775	0.4526
776	0.4679
777	0.4294
778	0.4389
779	0.5147
780	0.4791
781	0.4587
782	0.4443
783	0.3696
784	0.4464
785	0.3448
786	0.4368
787	0.4222
788	0.4659
789	0.4375
790	0.4705
791	0.3759
792	0.4569
793	0.4640
794	0.4059
795	0.3977
796	0.3658
797	0.3720
798	0.4669
799	0.4415
800	0.4768
801	0.4070
802	0.4034
803	0.4793
804	0.4639
805	0.4015
806	0.4904
807	0.4363
808	0.4501
809	0.4419
810	0.3943
811	0.4937
812	0.4608
813	0.5358
814	0.3579
815	0.4783
816	0.4058
817	0.4213
818	0.4762
819	0.4869
820	0.4022
821	0.4169
822	0.4393
823	0.4686
824	0.3944
825	0.3662
826	0.4955
827	0.4098
828	0.4672
829	0.4449
830	0.4352
831	0.3875
832	0.4226
833	0.3394
834	0.4573
835	0.5714
836	0.3796
837	0.3589
838	0.5581
839	0.3508
840	0.4993
841	0.5427
842	0.4599
843	0.3685
844	0.3357
845	0.4668
846	0.4359
847	0.5249
848	0.4027
849	0.3745
850	0.4435
851	0.4559
852	0.5411
853	0.5393
854	0.4547
855	0.4264
856	0.4814
857	0.5601
858	0.4066
859	0.4399
860	0.3976
861	0.5323
862	0.4316
863	0.5158
864	0.4360
865	0.4437
866	0.5523
867	0.4600
868	0.5170
869	0.4416
870	0.5379
871	0.3765
872	0.5465
873	0.5100
874	0.5036
875	0.4662
876	0.5189
877	0.4052
878	0.3822
879	0.4711
880	0.5425
881	0.3624
882	0.5538
883	0.4150
884	0.4515
885	0.4018
886	0.4621
887	0.5146
888	0.4044
889	0.4957
890	0.3813
891	0.4218
892	0.4215
893	0.5782
894	0.3375
895	0.4348
896	0.4846
897	0.5206
898	0.4222
899	0.5568
900	0.3870
901	0.5047
902	0.4498
903	0.4854
904	0.4778
905	0.3219
906	0.5058
907	0.4065
908	0.4771
909	0.4847
910	0.4601
911	0.4329
912	0.5142
913	0.3879
914	0.4598
915	0.3777
916	0.4845
917	0.4164
918	0.3658
919	0.4445
920	0.4488
921	0.4079
922	0.4519
923	0.5244
924	0.5315
925	0.4104
926	0.4660
927	0.4922
928	0.4723
929	0.5434
930	0.4955
931	0.3362
932	0.3764
933	0.4558
934	0.4461
935	0.4514
936	0.3887
937	0.3222
938	0.5481
939	0.5187
940	0.5604
941	0.3419
942	0.5425
943	0.4893
944	0.4276
945	0.4592
946	0.4997
947	0.5063
948	0.3630
949	0.6162
950	0.3596
951	0.3796
952	0.4446
953	0.4774
954	0.4038
955	0.4295
956	0.4170
957	0.4643
958	0.4482
959	0.4736
960	0.4542
961	0.4381
962	0.4004
963	0.4963
964	0.3629
965	0.4291
966	0.5475
967	0.4830
968	0.5475
969	0.3492
970	0.5398
971	0.3624
972	0.4320
973	0.4654
974	0.5421
975	0.3809
976	0.4948
977	0.3636
978	0.4805
979	0.4961
980	0.4663
981	0.4846
982	0.4005
983	0.4435
984	0.4374
985	0.4629
986	0.5199
987	0.3918
988	0.4873
989	0.4926
990	0.3489
991	0.5340
992	0.3791
993	0.4176
994	0.3926
995	0.4034
996	0.4638
997	0.4323
998	0.4050
999	0.4403
1000	0.4210
1001	0.4444
1002	0.5897
1003	0.4928
1004	0.3608
1005	0.5128
1006	0.4417
1007	0.4465
1008	0.4026
1009	0.4971
1010	0.5111
1011	0.3577
1012	0.4295
1013	0.4541
1014	0.5121
1015	0.4728
1016	0.4217
1017	0.4174
1018	0.5151
1019	0.4519
1020	0.5478
1021	0.5371
1022	0.4670
1023	0.4583
1024	0.4829
1025	0.4662
1026	0.3802
1027	0.5009
1028	0.5024
1029	0.4055
1030	0.4775
1031	0.4994
1032	0.4948
1033	0.4770
1034	0.5101
1035	0.4511
1036	0.5013
1037	0.4941
1038	0.4604
1039	0.4091
1040	0.5518
1041	0.4770
1042	0.4585
1043	0.5219
1044	0.4521
1045	0.4675
1046	0.4704
1047	0.5420
1048	0.4577
1049	0.5425
1050	0.5071
1051	0.3828
1052	0.5194
1053	0.4084
1054	0.4738
1055	0.4452
1056	0.4019
1057	0.4623
1058	0.3685
1059	0.4351
1060	0.4651
1061	0.4680
1062	0.4222
1063	0.3716
1064	0.4669
1065	0.4953
1066	0.5303
1067	0.5180
1068	0.4814
1069	0.4582
1070	0.5190
1071	0.4128
1072	0.4817
1073	0.3919
1074	0.5225
1075	0.4652
1076	0.5102
1077	0.5597
1078	0.3603
1079	0.6722
1080	0.3878
1081	0.4534
1082	0.4579
1083	0.4735
1084	0.4376
1085	0.4625
1086	0.4066
1087	0.4711
1088	0.5528
1089	0.4394
1090	0.4745
1091	0.5218
1092	0.5240
1093	0.3927
1094	0.5026
1095	0.4855
1096	0.5024
1097	0.6366
1098	0.4130
1099	0.3204
1100	0.4742
1101	0.5829
1102	0.4076
1103	0.4193
1104	0.4432
1105	0.4555
1106	0.4832
1107	0.4600
1108	0.4774
1109	0.4982
1110	0.4847
1111	0.5063
1112	0.4114
1113	0.5006
1114	0.5353
1115	0.4832
1116	0.4519
1117	0.5051
1118	0.4219
1119	0.4371
1120	0.5014
1121	0.4591
1122	0.4104
1123	0.4555
1124	0.4232
1125	0.4306
1126	0.3968
1127	0.4318
1128	0.5278
1129	0.4338
1130	0.3839
1131	0.4220
1132	0.3750
1133	0.4190
1134	0.4791
1135	0.5213
1136	0.4724
1137	0.4351
1138	0.4466
1139	0.4866
1140	0.4127
1141	0.6647
1142	0.4270
1143	0.5122
1144	0.4559
1145	0.3593
1146	0.5535
1147	0.5266
1148	0.5128
1149	0.3899
1150	0.4583
1151	0.4391
1152	0.4511
1153	0.4813
1154	0.5253
1155	0.4348
1156	0.4890
1157	0.4751
1158	0.4289
1159	0.4223
1160	0.4082
1161	0.5520
1162	0.3904
1163	0.4567
1164	0.4870
1165	0.4177
1166	0.3865
1167	0.4498
1168	0.3531
1169	0.4622
1170	0.5352
1171	0.4301
1172	0.3856
1173	0.5480
1174	0.4427
1175	0.4622
1176	0.5006
1177	0.5574
1178	0.3640
1179	0.4202
1180	0.4094
1181	0.4347
1182	0.5028
1183	0.3758
1184	0.3977
1185	0.4804
1186	0.4446
1187	0.5480
1188	0.5760
1189	0.4932
1190	0.5115
1191	0.4123
1192	0.6115
1193	0.5669
1194	0.4007
1195	0.4659
1196	0.4885
1197	0.4063
1198	0.5240
1199	0.4976
1200	0.3992
1201	0.5434
1202	0.5263
1203	0.5368
1204	0.4369
1205	0.5339
1206	0.3805
1207	0.5378
1208	0.5445
1209	0.5659
1210	0.4110
1211	0.5351
1212	0.3839
1213	0.4064
1214	0.4866
1215	0.5411
1216	0.4349
1217	0.5000
1218	0.4277
1219	0.4090
1220	0.3929
1221	0.5074
1222	0.5365
1223	0.3689
1224	0.4605
1225	0.4660
1226	0.4245
1227	0.4700
1228	0.5340
1229	0.3370
1230	0.4672
1231	0.4407
1232	0.5789
1233	0.4420
1234	0.5471
1235	0.4497
1236	0.5376
1237	0.4253
1238	0.5326
1239	0.5755
1240	0.3482
1241	0.4975
1242	0.4232
1243	0.5562
1244	0.5697
1245	0.5882
1246	0.4764
1247	0.5053
1248	0.4399
1249	0.4627
1250	0.3921
1251	0.5177
1252	0.4478
1253	0.4243
1254	0.5031
1255	0.5164
1256	0.4668
1257	0.4627
1258	0.4909
1259	0.5888
1260	0.4991
1261	0.4541
1262	0.4601
1263	0.5459
1264	0.5404
1265	0.5081
1266	0.3957
1267	0.4087
1268	0.4156
1269	0.5071
1270	0.4427
1271	0.4706
1272	0.4020
1273	0.4934
1274	0.5356
1275	0.4977
1276	0.4706
1277	0.6032
1278	0.5111
1279	0.4623
1280	0.3860
1281	0.5590
1282	0.5702
1283	0.3732
1284	0.6358
1285	0.3964
1286	0.4063
1287	0.4555
1288	0.5076
1289	0.4290
1290	0.4399
1291	0.4205
1292	0.5739
1293	0.4540
1294	0.4103
1295	0.5581
1296	0.4269
1297	0.4066
1298	0.4811
1299	0.3951
1300	0.4082
1301	0.5167
1302	0.4936
1303	0.6438
1304	0.3310
1305	0.5215
1306	0.4671
1307	0.4380
1308	0.4205
1309	0.5411
1310	0.4481
1311	0.5014
1312	0.4171
1313	0.4985
1314	0.4804
1315	0.5565
1316	0.4873
1317	0.3913
1318	0.4522
1319	0.4452
1320	0.4961
1321	0.5204
1322	0.4093
1323	0.5368
1324	0.4769
1325	0.4487
1326	0.4573
1327	0.4620
1328	0.4452
1329	0.4521
1330	0.4088
1331	0.4276
1332	0.4573
1333	0.4125
1334	0.4028
1335	0.4786
1336	0.3734
1337	0.5121
1338	0.4889
1339	0.3813
1340	0.5207
1341	0.4526
1342	0.4682
1343	0.4498
1344	0.5342
1345	0.5579
1346	0.4059
1347	0.4000
1348	0.4921
1349	0.5223
1350	0.5230
1351	0.3962
1352	0.4136
1353	0.5992
1354	0.4535
1355	0.5311
1356	0.5353
1357	0.4201
1358	0.5088
1359	0.4419
1360	0.5896
1361	0.4099
1362	0.4855
1363	0.4875
1364	0.5343
1365	0.4990
1366	0.5173
1367	0.3925
1368	0.5060
1369	0.5002
1370	0.5440
1371	0.5324
1372	0.4864
1373	0.5057
1374	0.4048
1375	0.5224
1376	0.4601
1377	0.4438
1378	0.5367
1379	0.4819
1380	0.4632
1381	0.3926
1382	0.5150
1383	0.5510
1384	0.5375
1385	0.4979
1386	0.3998
1387	0.4954
1388	0.4074
1389	0.5040
1390	0.5399
1391	0.4043
1392	0.4955
1393	0.3813
1394	0.4600
1395	0.4872
1396	0.4759
1397	0.4423
1398	0.4769
1399	0.4848
1400	0.5398
1401	0.4863
1402	0.5045
1403	0.5243
1404	0.4756
1405	0.4629
1406	0.4806
1407	0.4983
1408	0.4015
1409	0.5797
1410	0.4270
1411	0.4879
1412	0.5104
1413	0.4808
1414	0.5626
1415	0.4752
1416	0.5146
1417	0.5003
1418	0.4773
1419	0.4690
1420	0.5216
1421	0.4504
1422	0.4327
1423	0.5213
1424	0.4928
1425	0.4920
1426	0.5643
1427	0.5562
1428	0.4594
1429	0.5237
1430	0.4378
1431	0.4630
1432	0.6028
1433	0.4414
1434	0.4062
1435	0.4225
1436	0.4244
1437	0.4940
1438	0.4796
1439	0.4740
1440	0.4920
1441	0.4451
1442	0.5674
1443	0.5500
1444	0.5071
1445	0.5396
1446	0.5244
1447	0.3806
1448	0.4542
1449	0.5351
1450	0.5263
1451	0.4149
1452	0.6183
1453	0.4326
1454	0.4919
1455	0.3986
1456	0.5495
1457	0.4811
1458	0.4561
1459	0.4008
1460	0.5478
1461	0.4455
1462	0.4565
1463	0.5201
1464	0.5255
1465	0.3973
1466	0.4650
1467	0.3754
1468	0.5047
1469	0.5009
1470	0.5970
1471	0.5053
1472	0.3854
1473	0.5156
1474	0.4901
1475	0.3973
1476	0.5794
1477	0.4830
1478	0.5060
1479	0.4463
1480	0.4131
1481	0.5517
1482	0.5492
1483	0.5247
1484	0.4117
1485	0.4758
1486	0.4198
1487	0.5492
1488	0.4959
1489	0.4785
1490	0.3674
1491	0.5684
1492	0.4950
1493	0.4835
1494	0.5071
1495	0.3599
1496	0.5560
1497	0.3860
1498	0.4667
1499	0.5297
1500	0.4608

Measured result

Verbatim from the run's summary.json.

bucket

bucket-p32-t128

bucket_icons

3359

checkpoint_bytes

7118998

checkpoint_round_trip

true

checkpoint_sha256

e5d553187e00d4f2…

config_sha256

5a9eaf449658d1e3…

corruption

factorized_role_uniform_geometry

cuda_version

13.4

deterministic_algorithms

true

device

cuda

eval_every

final_train

loss

3.3776731491088867

step

1500

train_token_accuracy

0.4608145106091718

group_vocabulary_size

locked_path_exact

true

metrics_sha256

c1e9d7c1b88ab2db…

model_parameters

580720

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

1500

evals_without_improvement

min_delta

0.0

objective

held_out_loss

patience_evals

selected_held_out_loss

7.4284515380859375

selected_step

1020

stopped_early

true

steps

6300

study_version

openmoji-g1-noise-conditioned-v4

subgroup_vocabulary_size

118

torch_version

2.14.0a0+4fdf77b940.nv26.08

validation

accuracy

0.29801539692061585

changed_accuracy

0.055587351552439546

changed_total

13978

loss

7.4284515380859375

retained_accuracy

0.4281982328082981

retained_total

26030

validation_final_step

accuracy

0.3064137172565487

changed_accuracy

0.06131063099155816

changed_total

13978

loss

7.59999942779541

retained_accuracy

0.438033038801383

retained_total

26030

validation_trace_sha256

b9bee5be6578b97a…

validation_untrained

accuracy

0.0027994401119776045

changed_accuracy

0.0030762626985262557

changed_total

13978

loss

16.14603042602539

retained_accuracy

0.002650787552823665

retained_total

26030

Written result

The corruption sweep found the Gate G denoiser's output near-independent of its input. The obvious informational explanation was that it trained at one fixed corruption level and was never told the level, so it had no reason to learn how much to trust x_t. This run supplies both halves of that fix: a corruption level sampled per example over 0.05 to 0.50, and 16 sinusoidal conditioning features telling the model the level.

Model size is v2's plus 3,168 parameters of noise projection. Selection and the headline held-out numbers stay at corruption 0.35, exactly the task v2 and v3 were measured on. This is a deliberate two-factor change — sampling without telling the model would leave it unable to use the level, telling it without varying it would leave it a constant — and is recorded as such rather than presented as one factor.

Result

Early stopping ended the run at step 1,500; the selected checkpoint is step 1,020. A second complete invocation returned an identical JSON result.

criterion	threshold	observed	outcome
stops_destroying_light_corruption	recovery at 0.10 above 0	-1.0385	falsified
output_depends_on_input	Pearson r below 0.70	0.8619	falsified
light_corruption_damage_collapses	recovery at 0.05 above -0.5	-4.5349	falsified
does_not_wreck_the_trained_task	held-out loss below 8.0	7.4285	pass
structural_safety	locked-path exact, round-trips	both true	pass
reproducibility	identical rerun	identical	pass

Against v2 on the identical 32-icon sweep draw:

corruption p	v2 recovery	v4 recovery
0.05	-4.6517	-4.5349
0.10	-1.2024	-1.0385
0.20	-0.1315	-0.0910
0.35	+0.1800	+0.2083
Pearson r, 0.05 vs 0.35	0.9122	0.8619

Every number moves in the predicted direction and none moves enough to matter. The model still destroys lightly corrupted inputs and still emits nearly the same picture whatever it is given.

Overall: falsified. Noise-level information was not the missing piece.

What it ruled out, and what that forced

This eliminated the informational explanation, which is what made it worth running: the model was not failing for want of knowing how corrupted its input was. That pushed the search to the encoder, where the actual defect turned out to be — see openmoji-g1-slot-bound-v5-f502df1-d8af55ea-9b9b1699. The encoder summed six coordinate lookups from one shared embedding table into a single vector per segment, so a segment was an unordered bag of its values and the model provably could not tell which coordinate held which. It was never ignoring x_t; it was reading a scrambled copy.

This run's negative result is what made that search necessary, and its 0.4282 retained accuracy — barely above v2's 0.4128 despite the extra conditioning — is consistent with a model whose input was scrambled before the conditioning could help.

State transitions

planned2026-09-20T19:45:00Z
completed2026-09-20T20:05:00Znoise_level_conditioning_is_not_the_missing_piece

Run record

Verbatim from runs/openmoji-g1-noise-conditioned-v4-6b935b2-5a9eaf44-9b9b1699/run.yaml, the record committed before launch.

schema_version

run_id

openmoji-g1-noise-conditioned-v4-6b935b2-5a9eaf44-9b9b1699

state

completed

parent_run

openmoji-g1-train-v2-datascale-a50b2e0-c474c94d-9b9b1699

diagnosis_run

openmoji-g1-corruption-sweep-71a080a-4levels-9b9b1699

hypothesis

The Gate G denoiser produced output near-independent of its input because it trained at one fixed corruption level and was never told the level, so it had no reason to learn how much to trust `x_t` and settled on a group-conditioned prior. Sampling a corruption level per example across 0.05 to 0.50 and conditioning the model on that level makes the output actually depend on the input: the model stops destroying lightly corrupted inputs, and its predictions at 0.05 and 0.35 stop being nearly the same picture.

expected_information_gain

scope_limits

changed_factors

primary

Two inseparable halves of one mechanism: corruption level sampled per example over 0.05 to 0.50 instead of fixed at 0.35, and 16 sinusoidal noise-level conditioning features supplied to the model. Sampling without telling the model would leave it unable to use the level; telling it without varying the level would leave it a constant. They are deliberately not separated, and this is recorded as a two-factor change rather than presented as one.

held_fixed

the full 2,681-icon split, the 128-icon validation draw and its corruption seeds, seed 3101, batch size 16, learning rate 0.001, eval_every 60, the 6,300-step cap, the held-out-loss selection policy with patience 8, the codec, and the bucket.

model_parameters

577552

580720

evaluation

Held-out evaluation and checkpoint selection stay at corruption 0.35, exactly the task v2 and v3 were measured on.

code

git_commit

6b935b2d25f8639504d42e45d8dd8964f4671fb0

execution_mode

native-local

config

path

configs/learning/openmoji-g1-noise-conditioned-v4.yaml

sha256

5a9eaf449658d1e3…

dataset

hybrid_sha256

9b9b1699677a6f97…

bucket

bucket-p32-t128

train_samples

2681

validation_samples

128

baseline

run

openmoji-g1-train-v2-datascale-a50b2e0-c474c94d-9b9b1699

held_out_loss_at_0.35

7.3332648277282715

held_out_changed_accuracy_at_0.35

0.057304335384175134

mean_recovery_fraction_72px

0.05

-4.6517

0.10

-1.2024

0.20

-0.1315

0.35

0.18

per_icon_pearson_r_p005_vs_p035

0.9122

execution

seed_set

3101

deterministic_algorithms

true

cublas_workspace_config

:4096:8

optimizer_step_cap

6300

early_stopping

objective

held_out_loss

patience_evals

min_delta

0.0

resource_cap

max_steps

20000

max_storage_gb

network

none

predeclared_criteria

note

Declared before launch. No preflight. The baseline figures above are v2's, measured on the same 32-icon sweep draw and the same 128-icon validation draw.

evaluated_at

the checkpoint selected by held-out loss at corruption 0.35

primary

stops_destroying_light_corruption

statement

Mean per-icon recovery fraction at corruption 0.10 and 72 px is above zero, against v2's -1.2024. Above zero means the model improves a lightly corrupted input rather than damaging it, which is a qualitative change of sign.

output_depends_on_input

statement

The per-icon Pearson correlation between `x_hat_0` render error at corruption 0.05 and at 0.35 falls below 0.70, against v2's 0.9122. This is the direct measure of the diagnosed mechanism.

light_corruption_damage_collapses

statement

Mean recovery fraction at corruption 0.05 is above -0.5, against v2's -4.6517.

does_not_wreck_the_trained_task

statement

Held-out loss at the selected checkpoint, evaluated at 0.35, is below 8.0. v2 reached 7.3333 while training only at 0.35, so some loss at that single level is expected from spreading training across a range; a large regression would mean the range was too wide rather than that conditioning failed.

structural_safety

statement

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

reproducibility

statement

A second complete invocation returns an identical JSON result.

standing_gate_not_a_prediction

retained_preservation

statement

Held-out retained-token accuracy at 0.35 is at least 0.90.

note

Carried from v1, v2 and v3 at 0.3256, 0.4128 and 0.4323. This run targets the mechanism behind that number, so movement here is the most informative single figure even though the bar is not predicted to be met.

falsification_meaning

If the correlation stays high and the model still destroys lightly corrupted inputs, noise-level information was not the missing piece. The next candidates are then structural rather than informational: predict an edit mask or a residual against `x_t` so that copying is the default, and separately condition on which fields were corrupted as a diagnostic upper bound to separate identification failure from prediction failure.

reported_not_gated

the full four-level recovery sweep, for direct comparison with v2's, held-out aggregate, changed and retained accuracy at 0.35, the selected step, whether the cap or early stopping ended the run, wall time and peak memory

outputs

local_metadata

runs/openmoji-g1-noise-conditioned-v4-6b935b2-5a9eaf44-9b9b1699

report_root

reports/learning/openmoji-g1-noise-conditioned-v4

durable_artifacts

/home/dev/.cache/openmoji-g1-noise-conditioned-v4-6b935b2-5a9eaf44-9b9b1699

planned_at

2026-09-20 19:45:00+00:00

completed_at

2026-09-20 20:05:00+00:00

result

device

cuda

model_parameters

580720

selected_step

1020

completed_steps

1500

ended_by

early_stopping

held_out_loss

7.4284515380859375

held_out_aggregate_accuracy

0.29801539692061585

held_out_changed_accuracy

0.055587351552439546

held_out_retained_accuracy

0.4281982328082981

checkpoint_sha256

e5d553187e00d4f2…

identical_rerun

true

mean_recovery_fraction_72px

0.05

-4.5349

0.10

-1.0385

0.20

-0.091

0.35

0.2083

per_icon_pearson_r_p005_vs_p035

0.8619

predeclared_outcome

overall

falsified

stops_destroying_light_corruption

passed

false

observed

-1.0385

threshold

0.0

output_depends_on_input

passed

false

observed

0.8619

threshold

0.7

light_corruption_damage_collapses

passed

false

observed

-4.5349

threshold

-0.5

does_not_wreck_the_trained_task

passed

true

observed

7.4284515380859375

threshold

8.0

structural_safety

passed

true

reproducibility

passed

true

conclusion

Noise-level information was not the missing piece. Every figure moved in the predicted direction and none moved enough to matter. Eliminating the informational explanation is what forced the search into the encoder, where a provable defect was found: the six coordinate lookups were summed from one shared table, so the model could not tell which coordinate held which value.