Neural Network Playground

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260

<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<link rel="icon" href="data:,">
<meta name="viewport" content="width=device-width, initial-scale=1, viewport-fit=cover">
<title>Neural Network Playground — sloprun.dev</title>
<style>
/* ── sloprun design tokens (verbatim from demos/tokens.css) ───────────────── */
:root {
  --bg: #F6F6F3;          /* neutral paper, not cream */
  --surface: #FFFFFF;
  --ink: #1A1D21;
  --muted: #5A6068;
  --line: #E3E2DC;
  --accent: #C05600;      /* phosphor amber, darkened for light bg */
  --accent-soft: #FFF3E6;
  --run: #1A7F37;         /* earned green */
  --run-soft: #E7F4EA;
  --danger: #C93C37;
  --radius: 10px;
  --font-sans: system-ui, -apple-system, "Segoe UI", sans-serif;
  --font-mono: ui-monospace, "SF Mono", "Cascadia Code", Menlo, Consolas, monospace;
}
@media (prefers-color-scheme: dark) {
  :root:not([data-theme="light"]) {
    --bg: #14161A; --surface: #1C1F26; --ink: #E8E6E1; --muted: #9AA0A8;
    --line: #2A2E36; --accent: #FFB454; --accent-soft: #2A2118;
    --run: #3FB950; --run-soft: #16281B; --danger: #F47067;
  }
}
:root[data-theme="dark"] {
  --bg: #14161A; --surface: #1C1F26; --ink: #E8E6E1; --muted: #9AA0A8;
  --line: #2A2E36; --accent: #FFB454; --accent-soft: #2A2118;
  --run: #3FB950; --run-soft: #16281B; --danger: #F47067;
}
body { background: var(--bg); color: var(--ink); font-family: var(--font-sans); }
/* ── end tokens ───────────────────────────────────────────────────────────── */

/* demo-local: the two class colours. Blue vs amber means "which group",
   nothing else. Green stays reserved for "every point is on the right side". */
:root {
  --c-a: #2E6FBF;   /* group A */
  --c-b: #C05600;   /* group B */
  --shadow: 0 1px 2px rgba(26,29,33,.05);
  --scrim: rgba(255,255,255,.88);
}
@media (prefers-color-scheme: dark) {
  :root:not([data-theme="light"]) {
    --c-a: #6FA8FF; --c-b: #FFB454;
    --shadow: 0 1px 2px rgba(0,0,0,.35);
    --scrim: rgba(28,31,38,.88);
  }
}
:root[data-theme="dark"] {
  --c-a: #6FA8FF; --c-b: #FFB454;
  --shadow: 0 1px 2px rgba(0,0,0,.35);
  --scrim: rgba(28,31,38,.88);
}

* { box-sizing: border-box; }
html, body { margin: 0; padding: 0; }
html { -webkit-text-size-adjust: 100%; }
body { min-height: 100%; overflow-x: hidden; }

.wrap {
  max-width: 1080px; margin: 0 auto; padding: 22px 16px 26px;
  display: flex; flex-direction: column; gap: 16px;
}
h1 { margin: 0 0 6px; font-size: clamp(21px, 3.6vw, 29px); letter-spacing: -.021em; line-height: 1.15; }
.lede { margin: 0; max-width: 70ch; color: var(--muted); font-size: 14.5px; line-height: 1.55; }
.lede b { color: var(--ink); font-weight: 600; }

.stage { display: grid; grid-template-columns: minmax(0,1fr) 318px; gap: 16px; align-items: start; }
.stage > .colmain { grid-column: 1; grid-row: 1; }
.stage > .panel   { grid-column: 2; grid-row: 1 / span 2; }
.stage > .netcard { grid-column: 1; grid-row: 2; }
@media (max-width: 900px) {
  .stage { grid-template-columns: minmax(0,1fr); }
  .stage > .colmain, .stage > .panel, .stage > .netcard { grid-column: 1; grid-row: auto; }
}

.field {
  position: relative; overflow: hidden;
  border: 1px solid var(--line); border-radius: var(--radius);
  background: var(--surface); aspect-ratio: 6 / 5; box-shadow: var(--shadow);
}
@media (max-width: 470px) { .field { aspect-ratio: 1 / 1; } }
canvas { display: block; width: 100%; height: 100%; }
#board { touch-action: pan-y; cursor: crosshair; }
canvas:focus { outline: none; }
canvas:focus-visible { outline: 2px solid var(--accent); outline-offset: -3px; }

.chip {
  position: absolute; left: 10px; bottom: 10px; pointer-events: none;
  font-family: var(--font-mono); font-size: 11px; letter-spacing: .01em;
  color: var(--muted); background: var(--scrim);
  border: 1px solid var(--line); border-radius: 999px; padding: 5px 10px;
  max-width: calc(100% - 20px); white-space: nowrap; overflow: hidden; text-overflow: ellipsis;
}
.chip b { color: var(--ink); font-weight: 600; }

.tag {
  position: absolute; right: 10px; top: 10px; pointer-events: none;
  font-family: var(--font-mono); font-size: 10px; letter-spacing: .07em; text-transform: uppercase;
  color: var(--muted); background: var(--scrim);
  border: 1px solid var(--line); border-radius: 999px; padding: 4px 9px;
}

.keys { display: flex; flex-wrap: wrap; gap: 7px 14px; align-items: center; padding: 9px 2px 0; }
.key { display: inline-flex; align-items: center; gap: 6px; font-family: var(--font-mono); font-size: 11px; color: var(--muted); letter-spacing: .01em; }
.key .dot { width: 11px; height: 11px; border-radius: 50%; flex: none; }
.key .dot.hollow { background: none !important; border: 1.6px solid var(--muted); }

.panel { display: flex; flex-direction: column; gap: 12px; }
.card { background: var(--surface); border: 1px solid var(--line); border-radius: var(--radius); padding: 12px 13px; box-shadow: var(--shadow); }
.card h2 {
  margin: 0 0 10px; font-family: var(--font-mono); font-size: 10.5px;
  letter-spacing: .1em; text-transform: uppercase; color: var(--muted); font-weight: 600;
  display: flex; justify-content: space-between; align-items: baseline; gap: 8px;
}
.card h2 .n { letter-spacing: .04em; text-transform: none; color: var(--muted); font-weight: 400; font-size: 10.5px; }

.stats { display: grid; grid-template-columns: repeat(3, minmax(0,1fr)); gap: 8px; }
.stat { display: flex; flex-direction: column; gap: 1px; min-width: 0; }
.stat .k { font-family: var(--font-mono); font-size: 9.5px; letter-spacing: .07em; text-transform: uppercase; color: var(--muted); }
.stat .v {
  font-family: var(--font-mono); font-size: 17px; letter-spacing: -.025em;
  font-variant-numeric: tabular-nums; color: var(--ink); line-height: 1.2;
}
.stat .v.ok { color: var(--run); }
.stat .v.bad { color: var(--danger); }

.sparkbox { margin-top: 10px; border: 1px solid var(--line); border-radius: 7px; overflow: hidden; background: var(--bg); }
#spark { height: 46px; }
.sparkfoot {
  display: flex; justify-content: space-between; gap: 8px;
  font-family: var(--font-mono); font-size: 9.5px; letter-spacing: .05em;
  color: var(--muted); margin-top: 4px;
}

.pillrow { display: flex; align-items: center; gap: 8px; margin-top: 11px; flex-wrap: wrap; }
.pill {
  font-family: var(--font-mono); font-size: 9.5px; letter-spacing: .07em; text-transform: uppercase;
  border: 1px solid var(--line); border-radius: 999px; padding: 3px 9px; color: var(--muted); white-space: nowrap;
}
.pill.go   { color: var(--accent); background: var(--accent-soft); border-color: transparent; }
.pill.done { color: var(--run); background: var(--run-soft); border-color: transparent; }
.pill.bad  { color: var(--danger); border-color: currentColor; }
.verdict { margin: 9px 0 0; font-size: 12.5px; line-height: 1.5; color: var(--muted); min-height: 3em; }
.verdict b { color: var(--ink); font-weight: 600; }

.grid3 { display: grid; grid-template-columns: repeat(3, minmax(0,1fr)); gap: 7px; }
#dataBtns button { font-size: 12px; padding: 7px 5px; }
.grid2 { display: grid; grid-template-columns: repeat(2, minmax(0,1fr)); gap: 7px; }
.seg button[aria-pressed="true"] { background: var(--accent-soft); color: var(--accent); border-color: var(--accent); font-weight: 600; }
.note { margin: 9px 0 0; font-size: 12px; line-height: 1.45; color: var(--muted); }
.sub { margin: 12px 0 5px; font-family: var(--font-mono); font-size: 9.5px; letter-spacing: .07em; text-transform: uppercase; color: var(--muted); }
.sub:first-of-type { margin-top: 2px; }

.ctl-top { display: flex; justify-content: space-between; align-items: baseline; gap: 8px; margin-bottom: 3px; }
.ctl-top label { font-size: 13px; }
.ctl-top .val { font-family: var(--font-mono); font-size: 12px; color: var(--ink); font-variant-numeric: tabular-nums; }
.ticks { display: flex; justify-content: space-between; font-family: var(--font-mono); font-size: 9.5px; color: var(--muted); letter-spacing: .04em; margin-top: 1px; }

input[type=range] {
  -webkit-appearance: none; appearance: none; display: block;
  width: 100%; height: 20px; margin: 0; background: transparent; cursor: pointer;
}
input[type=range]::-webkit-slider-runnable-track { height: 4px; border-radius: 2px; background: var(--line); }
input[type=range]::-webkit-slider-thumb {
  -webkit-appearance: none; width: 15px; height: 15px; margin-top: -5.5px;
  border-radius: 50%; background: var(--accent); border: 2px solid var(--surface); box-shadow: 0 0 0 1px var(--line);
}
input[type=range]::-moz-range-track { height: 4px; border-radius: 2px; background: var(--line); }
input[type=range]::-moz-range-thumb {
  width: 13px; height: 13px; border-radius: 50%; background: var(--accent);
  border: 2px solid var(--surface); box-shadow: 0 0 0 1px var(--line);
}
input[type=range]:focus-visible { outline: 2px solid var(--accent); outline-offset: 2px; border-radius: 6px; }

button {
  font: inherit; font-size: 13px; font-family: var(--font-sans);
  color: var(--ink); background: var(--surface);
  border: 1px solid var(--line); border-radius: 8px;
  padding: 7px 10px; cursor: pointer; line-height: 1.2; min-width: 0;
  transition: border-color .12s ease, color .12s ease, background .12s ease;
}
button:hover:not(:disabled) { border-color: var(--accent); color: var(--accent); }
button:focus-visible { outline: 2px solid var(--accent); outline-offset: 2px; }
button:disabled { opacity: .42; cursor: not-allowed; }
button.primary { background: var(--accent); color: var(--surface); border-color: var(--accent); font-weight: 600; }
button.primary:hover:not(:disabled) { color: var(--surface); filter: brightness(1.07); }
.row { display: flex; flex-wrap: wrap; gap: 7px; margin-top: 11px; }
.row > button { flex: 1 1 auto; }

.netcard .netwrap { border: 1px solid var(--line); border-radius: 7px; background: var(--bg); overflow: hidden; }
#net { display: block; width: 100%; }
.mathrow { display: flex; flex-wrap: wrap; gap: 6px 16px; margin-top: 10px; }
.mathrow div { font-family: var(--font-mono); font-size: 10.5px; line-height: 1.5; color: var(--muted); letter-spacing: -.01em; }
.mathrow div b { color: var(--ink); font-weight: 600; }

footer { font-family: var(--font-mono); font-size: 11px; color: var(--muted); letter-spacing: .05em; padding-top: 2px; }
.sr-only { position: absolute; width: 1px; height: 1px; padding: 0; margin: -1px; overflow: hidden; clip: rect(0 0 0 0); white-space: nowrap; border: 0; }

@media (max-width: 540px) { .netcard h2 .n { display: none; } }

@media (prefers-reduced-motion: reduce) {
  *, *::before, *::after { transition-duration: .001ms !important; animation-duration: .001ms !important; }
}
</style>
</head>
<body>
<div class="wrap">

  <header>
    <h1>Neural Network Playground</h1>
    <p class="lede">
      A small neural network, written from scratch right here, learning to tell two groups of dots apart.
      It starts knowing nothing: the coloured background is its <b>current guess for every point on the
      map</b>, and it bends that guess a little after every pass through the data. The two spirals are the
      hard one — watch it find the wrap.
    </p>
  </header>

  <main class="stage">

    <div class="colmain">
      <div class="field" id="field">
        <canvas id="board" aria-label="Map of the training points with the network's current decision boundary drawn behind them."
          >A scatter of dots in two groups, with the network's coloured guess drawn behind them.</canvas>
        <div class="tag" id="tag" aria-hidden="true">two spirals</div>
        <div class="chip" id="chip" aria-hidden="true"></div>
      </div>
      <div class="keys">
        <span class="key"><span class="dot" style="background:var(--c-a)"></span>group A</span>
        <span class="key"><span class="dot" style="background:var(--c-b)"></span>group B</span>
        <span class="key"><span class="dot hollow"></span>still on the wrong side</span>
        <span class="key">background = what the network would guess there</span>
      </div>
    </div>

    <div class="panel">

      <section class="card" aria-label="Training readout">
        <h2>Training <span class="n" id="paramsV">0 weights</span></h2>
        <div class="stats">
          <div class="stat"><span class="k">epoch</span><span class="v" id="epochV">0</span></div>
          <div class="stat"><span class="k">loss</span><span class="v" id="lossV">0.693</span></div>
          <div class="stat"><span class="k">correct</span><span class="v" id="accV">—</span></div>
        </div>
        <div class="sparkbox"><canvas id="spark" aria-hidden="true"></canvas></div>
        <div class="sparkfoot" aria-hidden="true"><span>LOSS PER EPOCH</span><span id="rateV">—</span></div>
        <div class="pillrow"><span class="pill" id="pill">ready</span></div>
        <p class="verdict" id="verdict"></p>
      </section>

      <section class="card" aria-label="Controls">
        <h2>Run</h2>
        <div class="row" style="margin-top:0">
          <button type="button" class="primary" id="playBtn">Pause</button>
          <button type="button" id="stepBtn">Step 200</button>
          <button type="button" id="resetBtn">Reset</button>
        </div>
        <div class="sub" style="margin-top:13px">Learning rate</div>
        <div class="ctl-top">
          <label for="lr">How big a nudge per pass</label><span class="val" id="lrV">0.030</span>
        </div>
        <input type="range" id="lr" min="0" max="1000" step="1" value="596" aria-describedby="lrHelp">
        <div class="ticks" aria-hidden="true"><span>0.001</span><span>0.017</span><span>0.300</span></div>
        <p class="note" id="lrHelp">Too small and it crawls. Too large and it thrashes — the loss curve goes ragged and never settles.</p>
      </section>

      <section class="card" aria-label="Data">
        <h2>Data</h2>
        <div class="grid3 seg" id="dataBtns" role="group" aria-label="Choose a dataset"></div>
        <p class="note" id="dataNote"></p>
      </section>

      <section class="card" aria-label="Network shape">
        <h2>Shape <span class="n" id="shapeV">2-8-8-1</span></h2>
        <div class="sub">Hidden layers</div>
        <div class="grid3 seg" id="layerBtns" role="group" aria-label="Number of hidden layers"></div>
        <div class="ctl-top" style="margin-top:13px">
          <label for="neurons">Neurons per layer</label><span class="val" id="neuronsV">8</span>
        </div>
        <input type="range" id="neurons" min="2" max="8" step="1" value="8">
        <div class="ticks" aria-hidden="true"><span>2</span><span>5</span><span>8</span></div>
        <div class="sub">Activation</div>
        <div class="grid2 seg" id="actBtns" role="group" aria-label="Activation function"></div>
        <p class="note">Changing the shape builds a fresh network and starts over from random weights.</p>
      </section>

    </div>

    <section class="card netcard" aria-label="Inside the network">
      <h2>Inside the network <span class="n">each tile is what that unit sees, over the same map</span></h2>
      <div class="netwrap"><canvas id="net" aria-hidden="true"></canvas></div>
      <p class="sr-only" id="netAlt"></p>
      <div class="mathrow">
        <div><b>forward</b> &nbsp;a = act(W·x + b), then p = sigmoid(...)</div>
        <div><b>loss</b> &nbsp;−[y·log p + (1−y)·log(1−p)]</div>
        <div><b>backward</b> &nbsp;chain rule, layer by layer</div>
        <div><b>update</b> &nbsp;Adam, mini-batches of 32</div>
      </div>
    </section>

  </main>

  <footer>demo · sloprun.dev</footer>
</div>

<div class="sr-only" role="status" aria-live="polite" id="live"></div>

<script>
(function () {
  "use strict";

  /* ═══ theme plumbing ════════════════════════════════════════════════════ */
  var root = document.documentElement;
  var mqDark = window.matchMedia ? matchMedia("(prefers-color-scheme: dark)") : null;
  var mqRM   = window.matchMedia ? matchMedia("(prefers-reduced-motion: reduce)") : null;
  var reduced = !!(mqRM && mqRM.matches);
  var coarse  = !!(window.matchMedia && matchMedia("(hover: none)").matches);
  var PROBE_HINT = coarse ? "drag across the map to read the network's guess"
                          : "hover the map to read the network's guess";

  window.addEventListener("message", function (e) {
    var d = e.data;
    if (d && d.type === "sloprun:theme" && (d.theme === "light" || d.theme === "dark")) {
      root.setAttribute("data-theme", d.theme);
      readTheme(); drawAll();
    }
  });
  if (mqDark) {
    var onScheme = function () { readTheme(); drawAll(); };
    if (mqDark.addEventListener) mqDark.addEventListener("change", onScheme);
    else if (mqDark.addListener) mqDark.addListener(onScheme);
  }

  /* ═══ colour helpers ════════════════════════════════════════════════════ */
  var COL = {};
  function hexToRgb(h) {
    h = String(h).trim();
    if (h.charAt(0) === "#") h = h.slice(1);
    if (h.length === 3) h = h.charAt(0)+h.charAt(0)+h.charAt(1)+h.charAt(1)+h.charAt(2)+h.charAt(2);
    var n = parseInt(h, 16);
    if (isNaN(n)) return null;
    return [(n >> 16) & 255, (n >> 8) & 255, n & 255];
  }
  function tok(name, fb) { return hexToRgb(getComputedStyle(root).getPropertyValue(name)) || fb; }
  function css(c) { return "rgb(" + c[0] + "," + c[1] + "," + c[2] + ")"; }
  function rgba(c, a) { return "rgba(" + c[0] + "," + c[1] + "," + c[2] + "," + a + ")"; }
  function mix(a, b, t) {
    return [a[0]+(b[0]-a[0])*t, a[1]+(b[1]-a[1])*t, a[2]+(b[2]-a[2])*t];
  }
  function readTheme() {
    COL.surface = tok("--surface", [255,255,255]);
    COL.bg      = tok("--bg",      [246,246,243]);
    COL.ink     = tok("--ink",     [26,29,33]);
    COL.muted   = tok("--muted",   [90,96,104]);
    COL.line    = tok("--line",    [227,226,220]);
    COL.accent  = tok("--accent",  [192,86,0]);
    COL.run     = tok("--run",     [26,127,55]);
    COL.a       = tok("--c-a",     [46,111,191]);
    COL.b       = tok("--c-b",     [192,86,0]);
    COL.dark    = (COL.bg[0] + COL.bg[1] + COL.bg[2]) / 3 < 128;
    COL.fieldA  = COL.dark ? 0.34 : 0.40;   /* the guess field sits behind the dots */
  }
  readTheme();

  /* ═══ deterministic randomness ══════════════════════════════════════════ */
  function mulberry32(a) {
    return function () {
      a |= 0; a = a + 0x6D2B79F5 | 0;
      var t = Math.imul(a ^ a >>> 15, 1 | a);
      t = t + Math.imul(t ^ t >>> 7, 61 | t) ^ t;
      return ((t ^ t >>> 14) >>> 0) / 4294967296;
    };
  }
  function gauss(rnd) {
    var u = 0, v = 0;
    while (u === 0) u = rnd();
    while (v === 0) v = rnd();
    return Math.sqrt(-2 * Math.log(u)) * Math.cos(2 * Math.PI * v);
  }

  /* ═══ datasets ══════════════════════════════════════════════════════════
     Every point lives in roughly [-1, 1] on both axes. Labels are 0 / 1. */
  var DATASETS = [
    { id: "spiral",  name: "two spirals",
      note: "Two arms wound around each other. A straight line can never split these — the network has to learn a curve that spirals too.",
      brief: "Two arms that wrap around each other." },
    { id: "circles", name: "rings",
      note: "A blob inside a ring. Easy for even a tiny network: one closed curve does it, and the loss drops in a couple of hundred passes.",
      brief: "A blob inside a ring." },
    { id: "xor",     name: "four blobs",
      note: "Diagonal pairs belong together. Needs at least one hidden layer — this is the classic problem a single layer of weights cannot solve.",
      brief: "Diagonal pairs belong together." }
  ];

  function makeData(kind, seed) {
    var rnd = mulberry32(seed), X = [], Y = [], i, c, n;
    if (kind === "spiral") {
      n = 120;
      for (c = 0; c < 2; c++) for (i = 0; i < n; i++) {
        var r = 0.06 + 0.94 * (i / n);
        var t = 1.75 * 2 * Math.PI * (i / n) + c * Math.PI;
        X.push([r * Math.cos(t) + gauss(rnd) * 0.035, r * Math.sin(t) + gauss(rnd) * 0.035]);
        Y.push(c);
      }
    } else if (kind === "circles") {
      n = 120;
      for (c = 0; c < 2; c++) for (i = 0; i < n; i++) {
        var rr = c === 0 ? 0.45 * Math.sqrt(rnd()) : 0.72 + 0.28 * Math.sqrt(rnd());
        var th = rnd() * 2 * Math.PI;
        X.push([rr * Math.cos(th) + gauss(rnd) * 0.03, rr * Math.sin(th) + gauss(rnd) * 0.03]);
        Y.push(c);
      }
    } else {
      n = 60;
      var cs = [[0.5, 0.5, 0], [-0.5, -0.5, 0], [0.5, -0.5, 1], [-0.5, 0.5, 1]];
      for (var b = 0; b < 4; b++) for (i = 0; i < n; i++) {
        X.push([cs[b][0] + gauss(rnd) * 0.16, cs[b][1] + gauss(rnd) * 0.16]);
        Y.push(cs[b][2]);
      }
    }
    return { X: X, Y: Y, n: X.length };
  }

  /* ═══ the network ═══════════════════════════════════════════════════════
     Plain dense layers. Hidden units use tanh or relu, the single output
     unit uses a sigmoid, and the loss is binary cross-entropy. Gradients are
     computed by hand below — nothing is imported, nothing is fetched. */
  function makeNet(sizes, act, seed) {
    var rnd = mulberry32(seed), L = [], l;
    for (l = 1; l < sizes.length; l++) {
      var fi = sizes[l - 1], fo = sizes[l], last = (l === sizes.length - 1);
      /* He init for relu hidden units, Xavier-ish otherwise */
      var scale = (act === "relu" && !last) ? Math.sqrt(2 / fi) : Math.sqrt(1 / fi);
      var W = new Float64Array(fi * fo);
      for (var k = 0; k < W.length; k++) W[k] = gauss(rnd) * scale;
      L.push({
        fi: fi, fo: fo, last: last,
        W: W, B: new Float64Array(fo),
        mW: new Float64Array(fi * fo), vW: new Float64Array(fi * fo),
        mB: new Float64Array(fo), vB: new Float64Array(fo),
        gW: new Float64Array(fi * fo), gB: new Float64Array(fo),
        a: new Float64Array(fo), z: new Float64Array(fo), d: new Float64Array(fo)
      });
    }
    var params = 0;
    for (l = 0; l < L.length; l++) params += L[l].W.length + L[l].B.length;
    return { L: L, act: act, sizes: sizes.slice(), t: 0, params: params };
  }

  var probe2 = [0, 0];
  function forward(net, x) {
    var L = net.L, inp = x, l, i, j;
    for (l = 0; l < L.length; l++) {
      var la = L[l];
      for (j = 0; j < la.fo; j++) {
        var s = la.B[j], off = j * la.fi;
        for (i = 0; i < la.fi; i++) s += la.W[off + i] * inp[i];
        la.z[j] = s;
        la.a[j] = la.last ? 1 / (1 + Math.exp(-s))
                          : (net.act === "relu" ? (s > 0 ? s : 0) : Math.tanh(s));
      }
      inp = la.a;
    }
    return L[L.length - 1].a[0];
  }
  function predictAt(net, x, y) { probe2[0] = x; probe2[1] = y; return forward(net, probe2); }

  var B1 = 0.9, B2 = 0.999, EPS = 1e-8, BATCH = 32;

  function trainEpoch(net, data, lr, order, rnd) {
    var N = data.n, L = net.L, i, l, j, ii, q;
    for (i = N - 1; i > 0; i--) {
      var k = (rnd() * (i + 1)) | 0, tmp = order[i]; order[i] = order[k]; order[k] = tmp;
    }
    var total = 0;
    for (var st = 0; st < N; st += BATCH) {
      var end = st + BATCH; if (end > N) end = N;
      var m = end - st;
      for (l = 0; l < L.length; l++) { L[l].gW.fill(0); L[l].gB.fill(0); }

      for (q = st; q < end; q++) {
        var idx = order[q], x = data.X[idx], y = data.Y[idx];
        var p = forward(net, x);
        var pc = p < 1e-9 ? 1e-9 : (p > 1 - 1e-9 ? 1 - 1e-9 : p);
        total += -(y * Math.log(pc) + (1 - y) * Math.log(1 - pc));

        /* dLoss/dz for a sigmoid output under cross-entropy is just (p - y) */
        var lastI = L.length - 1;
        L[lastI].d[0] = p - y;
        for (l = lastI; l >= 0; l--) {
          var la = L[l], prev = (l === 0) ? x : L[l - 1].a;
          for (j = 0; j < la.fo; j++) {
            var dj = la.d[j], off = j * la.fi;
            la.gB[j] += dj;
            for (ii = 0; ii < la.fi; ii++) la.gW[off + ii] += dj * prev[ii];
          }
          if (l > 0) {
            var pl = L[l - 1];
            for (ii = 0; ii < la.fi; ii++) {
              var s = 0;
              for (j = 0; j < la.fo; j++) s += la.W[j * la.fi + ii] * la.d[j];
              pl.d[ii] = (net.act === "relu") ? (pl.z[ii] > 0 ? s : 0)
                                              : s * (1 - pl.a[ii] * pl.a[ii]);
            }
          }
        }
      }

      /* Adam step on the batch-averaged gradient */
      net.t++;
      var c1 = 1 - Math.pow(B1, net.t), c2 = 1 - Math.pow(B2, net.t), inv = 1 / m;
      for (l = 0; l < L.length; l++) {
        var lb = L[l], w, g;
        for (w = 0; w < lb.W.length; w++) {
          g = lb.gW[w] * inv;
          lb.mW[w] = B1 * lb.mW[w] + (1 - B1) * g;
          lb.vW[w] = B2 * lb.vW[w] + (1 - B2) * g * g;
          lb.W[w] -= lr * (lb.mW[w] / c1) / (Math.sqrt(lb.vW[w] / c2) + EPS);
        }
        for (w = 0; w < lb.B.length; w++) {
          g = lb.gB[w] * inv;
          lb.mB[w] = B1 * lb.mB[w] + (1 - B1) * g;
          lb.vB[w] = B2 * lb.vB[w] + (1 - B2) * g * g;
          lb.B[w] -= lr * (lb.mB[w] / c1) / (Math.sqrt(lb.vB[w] / c2) + EPS);
        }
      }
    }
    return total / N;
  }

  function scoreAll(net, data) {
    var right = 0, loss = 0, N = data.n;
    for (var i = 0; i < N; i++) {
      var p = forward(net, data.X[i]), y = data.Y[i];
      var pc = p < 1e-9 ? 1e-9 : (p > 1 - 1e-9 ? 1 - 1e-9 : p);
      loss += -(y * Math.log(pc) + (1 - y) * Math.log(1 - pc));
      pred[i] = p;
      if ((p >= 0.5 ? 1 : 0) === y) right++;
    }
    return { right: right, loss: loss / N };
  }

  /* ═══ state ═════════════════════════════════════════════════════════════ */
  var DOMAIN = 1.15;          /* half-width of the visible map, in data units */
  var EPOCHS_PER_SEC = 200;   /* paced by wall clock so it looks the same everywhere */

  var st = {
    dataId: "spiral",
    layers: 2,
    neurons: 8,
    act: "tanh",
    lr: 0.03,
    playing: !reduced,
    epoch: 0,
    loss: 0.693,
    right: 0,
    diverged: false,
    seed: 7
  };
  var data = makeData(st.dataId, 7);
  var pred = new Float64Array(data.n);
  var net, order, shuffleRnd, hist = [], rate = 0, rateCount = 0, rateT0 = 0;

  function sizesOf() {
    var s = [2];
    for (var i = 0; i < st.layers; i++) s.push(st.neurons);
    s.push(1);
    return s;
  }

  function rebuild(newSeed) {
    if (newSeed != null) st.seed = newSeed;
    net = makeNet(sizesOf(), st.act, st.seed);
    order = new Int32Array(data.n);
    for (var i = 0; i < data.n; i++) order[i] = i;
    shuffleRnd = mulberry32(st.seed * 7919 + 13);
    st.epoch = 0; st.diverged = false;
    hist = [];
    var s = scoreAll(net, data);
    st.loss = s.loss; st.right = s.right;
    hist.push(s.loss);
    rate = 0; rateCount = 0; rateT0 = 0;
    allocActs();
    syncUI();
  }

  /* ═══ canvases ══════════════════════════════════════════════════════════ */
  var field = document.getElementById("field");
  var board = document.getElementById("board");
  var bctx  = board.getContext("2d");
  var spark = document.getElementById("spark");
  var sctx  = spark.getContext("2d");
  var netc  = document.getElementById("net");
  var nctx  = netc.getContext("2d");

  var off = document.createElement("canvas");
  var octx = off.getContext("2d");
  var GW = 104, GH = 88, gimg = null;
  off.width = GW; off.height = GH;

  var bW = 0, bH = 0, dpr = 1;

  function fitCanvas(cv, ctx, cssW, cssH) {
    dpr = Math.min(window.devicePixelRatio || 1, 2);
    var w = Math.max(1, Math.round(cssW * dpr)), h = Math.max(1, Math.round(cssH * dpr));
    if (cv.width !== w || cv.height !== h) { cv.width = w; cv.height = h; }
    ctx.setTransform(dpr, 0, 0, dpr, 0, 0);
  }

  function layout() {
    var r = field.getBoundingClientRect();
    bW = Math.max(120, Math.round(r.width));
    bH = Math.max(120, Math.round(r.height));
    fitCanvas(board, bctx, bW, bH);

    /* keep the boundary grid roughly square-celled */
    var target = 104;
    if (bW >= bH) { GW = target; GH = Math.max(24, Math.round(target * bH / bW)); }
    else { GH = target; GW = Math.max(24, Math.round(target * bW / bH)); }
    off.width = GW; off.height = GH;
    gimg = octx.createImageData(GW, GH);

    var sr = spark.getBoundingClientRect();
    fitCanvas(spark, sctx, Math.max(60, Math.round(sr.width)), 46);

    var nr = netc.getBoundingClientRect();
    var nH = 66 + Math.max(3, st.neurons) * 30;
    if (nH > 330) nH = 330;
    netc.style.height = nH + "px";
    fitCanvas(netc, nctx, Math.max(200, Math.round(nr.width)), nH);
  }

  /* pixel <-> data mapping: the square domain fits the short side of the box */
  function unitsPerPx() { return (2 * DOMAIN) / Math.min(bW, bH); }
  function px2x(px) { return (px - bW / 2) * unitsPerPx(); }
  function px2y(py) { return -(py - bH / 2) * unitsPerPx(); }
  function x2px(x)  { return bW / 2 + x / unitsPerPx(); }
  function y2px(y)  { return bH / 2 - y / unitsPerPx(); }

  /* ═══ drawing: the map ══════════════════════════════════════════════════ */
  function drawBoundary() {
    if (!gimg) return;
    var d = gimg.data, u = unitsPerPx();
    var halfW = bW / 2, halfH = bH / 2;
    var i = 0;
    for (var gy = 0; gy < GH; gy++) {
      var py = (gy + 0.5) * bH / GH;
      var yy = -(py - halfH) * u;
      for (var gx = 0; gx < GW; gx++) {
        var pxx = (gx + 0.5) * bW / GW;
        var xx = (pxx - halfW) * u;
        var p = predictAt(net, xx, yy);
        if (p !== p) p = 0.5;                       /* NaN guard */
        var t = (p - 0.5) * 2;                      /* -1 .. 1 */
        var mag = Math.pow(Math.abs(t), 0.72) * COL.fieldA;
        var c = mix(COL.surface, t < 0 ? COL.a : COL.b, mag);
        d[i++] = c[0]; d[i++] = c[1]; d[i++] = c[2]; d[i++] = 255;
      }
    }
    octx.putImageData(gimg, 0, 0);
  }

  function drawMap() {
    bctx.save();
    bctx.clearRect(0, 0, bW, bH);
    bctx.imageSmoothingEnabled = true;
    if (bctx.imageSmoothingQuality) bctx.imageSmoothingQuality = "high";
    bctx.drawImage(off, 0, 0, bW, bH);

    /* faint axes through the origin */
    bctx.strokeStyle = rgba(COL.ink, 0.10);
    bctx.lineWidth = 1;
    bctx.beginPath();
    bctx.moveTo(0, Math.round(y2px(0)) + 0.5); bctx.lineTo(bW, Math.round(y2px(0)) + 0.5);
    bctx.moveTo(Math.round(x2px(0)) + 0.5, 0); bctx.lineTo(Math.round(x2px(0)) + 0.5, bH);
    bctx.stroke();

    /* points: filled once the network puts them on the right side */
    var r = bW < 380 ? 3 : 3.7;
    for (var i = 0; i < data.n; i++) {
      var X = data.X[i], y = data.Y[i];
      var cx = x2px(X[0]), cy = y2px(X[1]);
      var ok = (pred[i] >= 0.5 ? 1 : 0) === y;
      var c = y === 0 ? COL.a : COL.b;
      bctx.beginPath();
      bctx.arc(cx, cy, r, 0, 6.28318530718);
      if (ok) {
        bctx.fillStyle = css(c); bctx.fill();
        bctx.lineWidth = 1.1; bctx.strokeStyle = rgba(COL.surface, 0.9); bctx.stroke();
      } else {
        bctx.fillStyle = rgba(COL.surface, 0.72); bctx.fill();
        bctx.lineWidth = 1.6; bctx.strokeStyle = css(c); bctx.stroke();
      }
    }
    bctx.restore();
  }

  /* ═══ drawing: the loss sparkline ═══════════════════════════════════════ */
  function drawSpark() {
    var w = spark.width / dpr, h = spark.height / dpr;
    sctx.clearRect(0, 0, w, h);
    var n = hist.length;
    if (n < 2) return;
    var top = 0, i;
    for (i = 0; i < n; i++) if (hist[i] > top && isFinite(hist[i])) top = hist[i];
    if (top < 0.05) top = 0.05;
    var pad = 4, gh = h - pad * 2;
    function yOf(v) {
      if (!isFinite(v)) v = top;
      var t = v / top; if (t > 1) t = 1; if (t < 0) t = 0;
      return pad + (1 - t) * gh;
    }
    /* One pixel column covers many epochs once the run is long, so keep BOTH
       the lowest and the highest loss in the column. When training settles the
       two collapse onto a single line; when the learning rate is too big the
       gap opens into a visible ribbon. Plotting only the minimum (as a plain
       sparkline does) would draw a thrashing run as a smooth flat line. */
    var pts = [], W = Math.max(2, Math.floor(w));
    for (var px = 0; px < W; px++) {
      var a = Math.floor(px / W * n), b = Math.floor((px + 1) / W * n);
      if (b <= a) b = a + 1; if (b > n) b = n;
      var lo = Infinity, hi = -Infinity;
      for (i = a; i < b; i++) {
        var hv = hist[i];
        if (!isFinite(hv)) hv = top;
        if (hv < lo) lo = hv;
        if (hv > hi) hi = hv;
      }
      if (lo === Infinity) { lo = hi = hist[n - 1]; }
      pts.push([px, yOf(lo), yOf(hi)]);
    }
    var curve = st.diverged ? tok("--danger", [201, 60, 55]) : COL.accent;
    /* area under the low edge */
    sctx.beginPath();
    sctx.moveTo(pts[0][0], h);
    for (i = 0; i < pts.length; i++) sctx.lineTo(pts[i][0], pts[i][1]);
    sctx.lineTo(pts[pts.length - 1][0], h);
    sctx.closePath();
    sctx.fillStyle = rgba(curve, 0.11);
    sctx.fill();
    /* min-to-max ribbon: thin when it settles, fat when it thrashes */
    sctx.beginPath();
    sctx.moveTo(pts[0][0], pts[0][2]);
    for (i = 1; i < pts.length; i++) sctx.lineTo(pts[i][0], pts[i][2]);
    for (i = pts.length - 1; i >= 0; i--) sctx.lineTo(pts[i][0], pts[i][1]);
    sctx.closePath();
    sctx.fillStyle = rgba(curve, 0.28);
    sctx.fill();
    /* line */
    sctx.beginPath();
    for (i = 0; i < pts.length; i++) (i ? sctx.lineTo(pts[i][0], pts[i][1]) : sctx.moveTo(pts[i][0], pts[i][1]));
    sctx.lineWidth = 1.4;
    sctx.strokeStyle = css(curve);
    sctx.stroke();
    /* head marker */
    var last = pts[pts.length - 1];
    sctx.beginPath();
    sctx.arc(last[0] - 1, last[1], 2, 0, 6.28318530718);
    sctx.fillStyle = css(COL.accent);
    sctx.fill();
  }

  /* ═══ drawing: the network diagram ══════════════════════════════════════
     Every unit gets a small tile showing its own output across the same map,
     computed by running an 18x18 grid of points through the network once and
     keeping each layer's activations. */
  var NG = 18, gridPts = null, acts = null, actMax = null;
  function buildGridPts() {
    gridPts = new Float64Array(NG * NG * 2);
    var k = 0;
    for (var gy = 0; gy < NG; gy++) for (var gx = 0; gx < NG; gx++) {
      gridPts[k++] = (gx + 0.5) / NG * 2 * DOMAIN - DOMAIN;
      gridPts[k++] = DOMAIN - (gy + 0.5) / NG * 2 * DOMAIN;
    }
  }
  buildGridPts();
  function allocActs() {
    acts = []; actMax = [];
    for (var l = 0; l < net.L.length; l++) {
      acts.push(new Float32Array(net.L[l].fo * NG * NG));
      actMax.push(new Float32Array(net.L[l].fo));
    }
  }
  function computeActs() {
    var P = NG * NG, l, j;
    for (l = 0; l < actMax.length; l++) actMax[l].fill(1e-6);
    for (var g = 0; g < P; g++) {
      probe2[0] = gridPts[g * 2]; probe2[1] = gridPts[g * 2 + 1];
      forward(net, probe2);
      for (l = 0; l < net.L.length; l++) {
        var la = net.L[l], arr = acts[l];
        for (j = 0; j < la.fo; j++) {
          var v = la.a[j];
          if (v !== v) v = 0;
          arr[j * P + g] = v;
          var av = Math.abs(v);
          if (av > actMax[l][j]) actMax[l][j] = av;
        }
      }
    }
  }

  var tileCv = document.createElement("canvas");
  var tctx = tileCv.getContext("2d");
  tileCv.width = NG; tileCv.height = NG;
  var timg = tctx.createImageData(NG, NG);

  function paintTile(get) {
    var d = timg.data, i = 0;
    for (var g = 0; g < NG * NG; g++) {
      var v = get(g);                            /* -1 .. 1 */
      if (v !== v) v = 0;
      if (v > 1) v = 1; if (v < -1) v = -1;
      var c = mix(COL.surface, v < 0 ? COL.a : COL.b, Math.pow(Math.abs(v), 0.75) * 0.85);
      d[i++] = c[0]; d[i++] = c[1]; d[i++] = c[2]; d[i++] = 255;
    }
    tctx.putImageData(timg, 0, 0);
  }

  function drawNet() {
    var w = netc.width / dpr, h = netc.height / dpr;
    nctx.clearRect(0, 0, w, h);
    computeActs();

    var cols = 2 + st.layers;                     /* inputs + hidden + output */
    var counts = [2];
    for (var i = 0; i < st.layers; i++) counts.push(st.neurons);
    counts.push(1);
    var maxN = Math.max.apply(null, counts);

    var padTop = 10, padBot = 20;
    var tile = Math.min(30, Math.floor((h - padTop - padBot - (maxN - 1) * 6) / maxN));
    if (tile < 10) tile = 10;
    var gap = 6;
    var colW = w / cols;
    var xs = [];
    for (i = 0; i < cols; i++) xs.push(colW * (i + 0.5));

    function nodeY(ci, j) {
      var n = counts[ci];
      var totalH = n * tile + (n - 1) * gap;
      var top = padTop + (h - padTop - padBot - totalH) / 2;
      return top + j * (tile + gap);
    }

    /* edges first, under the tiles */
    for (var l = 0; l < net.L.length; l++) {
      var la = net.L[l];
      var x0 = xs[l] + tile / 2, x1 = xs[l + 1] - tile / 2;
      for (var j = 0; j < la.fo; j++) {
        var y1 = nodeY(l + 1, j) + tile / 2;
        for (var ii = 0; ii < la.fi; ii++) {
          var y0 = nodeY(l, ii) + tile / 2;
          var wv = la.W[j * la.fi + ii];
          if (wv !== wv) wv = 0;
          var mag = Math.abs(wv);
          var alpha = Math.min(0.8, 0.10 + mag * 0.30);
          nctx.beginPath();
          nctx.moveTo(x0, y0); nctx.lineTo(x1, y1);
          nctx.lineWidth = Math.max(0.5, Math.min(3.2, mag * 1.1));
          nctx.strokeStyle = rgba(wv >= 0 ? COL.b : COL.a, alpha);
          nctx.stroke();
        }
      }
    }

    nctx.imageSmoothingEnabled = true;
    if (nctx.imageSmoothingQuality) nctx.imageSmoothingQuality = "high";

    function blit(cx, cy) {
      nctx.save();
      nctx.beginPath();
      var rr = 4;
      roundRect(nctx, cx, cy, tile, tile, rr);
      nctx.clip();
      nctx.drawImage(tileCv, cx, cy, tile, tile);
      nctx.restore();
      nctx.beginPath();
      roundRect(nctx, cx + 0.5, cy + 0.5, tile - 1, tile - 1, 4);
      nctx.lineWidth = 1;
      nctx.strokeStyle = rgba(COL.ink, 0.16);
      nctx.stroke();
    }

    /* inputs: the raw coordinates */
    for (i = 0; i < 2; i++) {
      (function (which) {
        paintTile(function (g) { return gridPts[g * 2 + which] / DOMAIN; });
      })(i);
      blit(xs[0] - tile / 2, nodeY(0, i));
    }
    /* hidden + output units */
    var P = NG * NG;
    for (l = 0; l < net.L.length; l++) {
      var lay = net.L[l], arr = acts[l];
      for (j = 0; j < lay.fo; j++) {
        (function (a, jj, isLast, mx) {
          paintTile(function (g) {
            var v = a[jj * P + g];
            if (isLast) return (v - 0.5) * 2;
            if (st.act === "relu") return Math.min(1, v / mx);
            return v;
          });
        })(arr, j, lay.last, Math.max(1e-6, actMax[l][j]));
        blit(xs[l + 1] - tile / 2, nodeY(l + 1, j));
      }
    }

    /* column labels in mono */
    nctx.font = "9px " + getComputedStyle(root).getPropertyValue("--font-mono");
    nctx.textAlign = "center";
    nctx.textBaseline = "alphabetic";
    nctx.fillStyle = css(COL.muted);
    var labels = ["input"];
    for (i = 0; i < st.layers; i++) labels.push(st.layers === 1 ? "hidden" : "hidden " + (i + 1));
    labels.push("output");
    for (i = 0; i < cols; i++) nctx.fillText(labels[i].toUpperCase(), xs[i], h - 7);
  }

  function roundRect(ctx, x, y, w, h, r) {
    if (ctx.roundRect) { ctx.roundRect(x, y, w, h, r); return; }
    ctx.moveTo(x + r, y);
    ctx.arcTo(x + w, y, x + w, y + h, r);
    ctx.arcTo(x + w, y + h, x, y + h, r);
    ctx.arcTo(x, y + h, x, y, r);
    ctx.arcTo(x, y, x + w, y, r);
    ctx.closePath();
  }

  function drawAll() { drawBoundary(); drawMap(); drawSpark(); drawNet(); }

  /* ═══ UI wiring ═════════════════════════════════════════════════════════ */
  var $ = function (id) { return document.getElementById(id); };
  var epochV = $("epochV"), lossV = $("lossV"), accV = $("accV"), rateV = $("rateV");
  var paramsV = $("paramsV"), shapeV = $("shapeV"), pill = $("pill"), verdict = $("verdict");
  var tag = $("tag"), chip = $("chip"), live = $("live"), dataNote = $("dataNote");
  var playBtn = $("playBtn"), stepBtn = $("stepBtn"), resetBtn = $("resetBtn");
  var lrEl = $("lr"), lrV = $("lrV"), neuronsEl = $("neurons"), neuronsV = $("neuronsV");
  var netAlt = $("netAlt");

  function say(msg) { live.textContent = msg; }

  /* dataset buttons */
  var dataBtns = $("dataBtns");
  DATASETS.forEach(function (d) {
    var b = document.createElement("button");
    b.type = "button"; b.textContent = d.name;
    b.setAttribute("aria-pressed", String(d.id === st.dataId));
    b.addEventListener("click", function () {
      if (st.dataId === d.id) return;
      st.dataId = d.id;
      data = makeData(d.id, 7);
      pred = new Float64Array(data.n);
      rebuild((Math.random() * 1e9) | 0);
      drawAll();
      say(d.name + " loaded, network reset");
    });
    dataBtns.appendChild(b);
  });

  /* hidden-layer buttons */
  var layerBtns = $("layerBtns");
  [1, 2, 3].forEach(function (k) {
    var b = document.createElement("button");
    b.type = "button"; b.textContent = String(k);
    b.setAttribute("aria-label", k + " hidden layer" + (k > 1 ? "s" : ""));
    b.setAttribute("aria-pressed", String(k === st.layers));
    b.addEventListener("click", function () {
      if (st.layers === k) return;
      st.layers = k;
      layout(); rebuild(); drawAll();
      say(k + " hidden layer" + (k > 1 ? "s" : "") + ", network reset");
    });
    layerBtns.appendChild(b);
  });

  /* activation buttons */
  var actBtns = $("actBtns");
  [["tanh", "tanh"], ["relu", "relu"]].forEach(function (p) {
    var b = document.createElement("button");
    b.type = "button"; b.textContent = p[1];
    b.style.fontFamily = "var(--font-mono)";
    b.setAttribute("aria-pressed", String(p[0] === st.act));
    b.addEventListener("click", function () {
      if (st.act === p[0]) return;
      st.act = p[0];
      rebuild(); drawAll();
      say(p[1] + " activation, network reset");
    });
    actBtns.appendChild(b);
  });

  function pressGroup(container, isOn) {
    var bs = container.querySelectorAll("button");
    for (var i = 0; i < bs.length; i++) bs[i].setAttribute("aria-pressed", String(isOn(i)));
  }

  /* learning rate: log scale from 0.001 to 0.3 */
  function lrFromSlider(v) {
    var t = v / 1000;
    return Math.exp(Math.log(0.001) + t * (Math.log(0.3) - Math.log(0.001)));
  }
  function sliderFromLr(lr) {
    return Math.round(1000 * (Math.log(lr) - Math.log(0.001)) / (Math.log(0.3) - Math.log(0.001)));
  }
  lrEl.value = String(sliderFromLr(st.lr));
  lrEl.addEventListener("input", function () {
    st.lr = lrFromSlider(+lrEl.value);
    lrV.textContent = st.lr.toFixed(3);
  });

  neuronsEl.addEventListener("input", function () {
    var n = +neuronsEl.value;
    if (n === st.neurons) return;
    st.neurons = n;
    neuronsV.textContent = String(n);
    layout(); rebuild(); drawAll();
  });
  neuronsEl.addEventListener("change", function () {
    say(st.neurons + " neurons per layer, network reset");
  });

  playBtn.addEventListener("click", function () {
    if (st.diverged) { rebuild((Math.random() * 1e9) | 0); drawAll(); }
    st.playing = !st.playing;
    if (st.playing) { rateT0 = 0; rateCount = 0; }
    syncUI();
    say(st.playing ? "training" : "paused at epoch " + st.epoch);
  });

  stepBtn.addEventListener("click", function () {
    if (st.diverged) return;
    for (var i = 0; i < 200 && !st.diverged; i++) runEpoch();
    refreshScore(); syncUI(); drawAll();
    say("stepped to epoch " + st.epoch + ", " + st.right + " of " + data.n + " correct");
  });

  resetBtn.addEventListener("click", function () {
    rebuild((Math.random() * 1e9) | 0);
    drawAll();
    say("new random weights, back to epoch 0");
  });

  /* hover readout — supplementary, mouse only */
  function onProbe(e) {
    var r = board.getBoundingClientRect();
    var px = e.clientX - r.left, py = e.clientY - r.top;
    if (px < 0 || py < 0 || px > r.width || py > r.height) return;
    var p = predictAt(net, px2x(px * bW / r.width), px2y(py * bH / r.height));
    if (p !== p) p = 0.5;
    var side = p >= 0.5 ? "group B" : "group A";
    var conf = (p >= 0.5 ? p : 1 - p);
    chip.innerHTML = "guess here: <b>" + side + "</b> &nbsp;" + (conf * 100).toFixed(0) + "% sure";
  }
  board.addEventListener("pointermove", onProbe);
  board.addEventListener("pointerleave", function () {
    chip.textContent = PROBE_HINT;
  });

  /* ═══ readout ═══════════════════════════════════════════════════════════ */
  function fmtLoss(v) {
    if (!isFinite(v)) return "—";
    return v < 0.001 ? v.toExponential(1).replace("e-", "e−") : v.toFixed(3);
  }

  function syncUI() {
    var solved = st.right === data.n;
    epochV.textContent = String(st.epoch);
    lossV.textContent = fmtLoss(st.loss);
    lossV.className = "v" + (st.diverged ? " bad" : "");
    accV.textContent = st.right + "/" + data.n;
    accV.className = "v" + (solved && !st.diverged ? " ok" : "");
    paramsV.textContent = net.params + " weights";
    shapeV.textContent = sizesOf().join("-");
    rateV.textContent = st.playing && rate > 0 ? Math.round(rate) + " EPOCHS/S" : "—";
    lrV.textContent = st.lr.toFixed(3);
    neuronsV.textContent = String(st.neurons);
    playBtn.textContent = st.diverged ? "Start over" : (st.playing ? "Pause" : (st.epoch ? "Resume" : "Train"));
    stepBtn.disabled = st.diverged;
    tag.textContent = DATASETS.filter(function (d) { return d.id === st.dataId; })[0].name;
    dataNote.textContent = DATASETS.filter(function (d) { return d.id === st.dataId; })[0].note;

    pressGroup(dataBtns, function (i) { return DATASETS[i].id === st.dataId; });
    pressGroup(layerBtns, function (i) { return i + 1 === st.layers; });
    pressGroup(actBtns, function (i) { return (i === 0 ? "tanh" : "relu") === st.act; });

    /* only call it a dead end once it has genuinely stopped improving */
    var slow = (st.epoch > 1400 && st.right / data.n < 0.9) ? recent() : null;
    if (slow && slow.drop >= 0.05) slow = null;

    var cls = "pill", txt = "ready", v;
    if (st.diverged) {
      cls = "pill bad"; txt = "blew up";
      v = "The nudges were too big and the weights ran off to infinity. Turn the <b>learning rate</b> down and start over.";
    } else if (solved) {
      cls = "pill done"; txt = "all " + data.n + " correct";
      v = "Every single point is on the right side of the line it drew. The loss keeps falling from here — the boundary just gets crisper.";
    } else if (!st.playing) {
      cls = "pill"; txt = st.epoch ? "paused" : "ready";
      v = st.epoch
        ? "Paused at epoch " + st.epoch + ". Press <b>Resume</b>, or take 200 passes at a time."
        : "Fresh random weights — the background is pure guesswork. Press <b>Train</b>.";
    } else if (slow) {
      /* the learning rate is only a plausible culprit when it is actually
         large — at a sane rate a flat curve means the shape is the limit */
      if (st.lr > 0.05 && (slow.spread > 0.10 || slow.mean > 0.60)) {
        cls = "pill bad"; txt = "thrashing";
        v = "The loss bounces instead of falling — every nudge overshoots the bottom. Turn the <b>learning rate</b> down.";
      } else if (st.lr < 0.005) {
        cls = "pill go"; txt = "crawling";
        v = "It is still learning, just far too slowly to watch. Turn the <b>learning rate</b> up.";
      } else {
        cls = "pill go"; txt = "stuck";
        v = "It has stopped improving. This shape can't bend the boundary enough — try <b>more neurons</b> or another hidden layer.";
      }
    } else {
      cls = "pill go"; txt = "training";
      v = st.epoch < 60
        ? "First passes: the boundary is still one soft smear across the map."
        : "Each pass nudges every weight a little against the slope of the loss. Watch the hollow dots fill in.";
    }
    pill.className = cls; pill.textContent = txt;
    verdict.innerHTML = v;

    netAlt.textContent = "Network shape " + sizesOf().join("-") + " with " + st.act +
      " activation, " + net.params + " weights, epoch " + st.epoch + ", " +
      st.right + " of " + data.n + " points correct.";
  }

  function refreshScore() {
    var s = scoreAll(net, data);
    st.loss = s.loss; st.right = s.right;
  }

  /* How the loss has behaved over the last stretch of epochs. Used to tell
     "the nudges are too big" (curve bounces, or sits pinned at chance) apart
     from "this shape cannot bend far enough" (curve smooth and flat) and from
     "the nudges are too small" (curve still falling, imperceptibly). */
  function recent() {
    var n = hist.length, a = n - 150; if (a < 0) a = 0;
    var w = n - a; if (w < 120) return null;
    var mn = Infinity, mx = -Infinity, sum = 0, i, v;
    for (i = a; i < n; i++) {
      v = hist[i]; if (!isFinite(v)) v = 20;
      if (v < mn) mn = v;
      if (v > mx) mx = v;
      sum += v;
    }
    var mean = sum / w, mid = a + (w >> 1), s1 = 0, s2 = 0;
    for (i = a; i < mid; i++) { v = hist[i]; s1 += isFinite(v) ? v : 20; }
    for (i = mid; i < n; i++) { v = hist[i]; s2 += isFinite(v) ? v : 20; }
    s1 /= (mid - a); s2 /= (n - mid);
    return {
      spread: (mx - mn) / Math.max(mean, 0.02),
      mean: mean,
      drop: (s1 - s2) / Math.max(s1, 1e-6)
    };
  }

  function runEpoch() {
    var l = trainEpoch(net, data, st.lr, order, shuffleRnd);
    st.epoch++;
    if (!isFinite(l) || Math.abs(l) > 1e6) {
      st.diverged = true; st.playing = false; st.loss = l;
      hist.push(isFinite(l) ? l : (hist.length ? hist[hist.length - 1] * 2 : 1));
      return;
    }
    st.loss = l;
    if (hist.length < 60000) hist.push(l);
  }

  /* ═══ main loop ═════════════════════════════════════════════════════════ */
  var acc = 0, lastT = 0, frame = 0, uiFrame = 0, bCost = 0, bEvery = 1;
  function tick(now) {
    requestAnimationFrame(tick);
    if (!lastT) lastT = now;
    var dt = Math.min(0.25, (now - lastT) / 1000);
    lastT = now;

    if (st.playing && !st.diverged) {
      acc += dt * EPOCHS_PER_SEC;
      var budget = Math.min(24, Math.floor(acc));
      acc -= budget;
      for (var i = 0; i < budget && !st.diverged; i++) runEpoch();
      if (budget > 0) {
        rateCount += budget;
        if (!rateT0) rateT0 = now;
        else if (now - rateT0 > 500) { rate = rateCount / ((now - rateT0) / 1000); rateCount = 0; rateT0 = now; }
        refreshScore();
        /* the boundary grid is the expensive part — if this machine is slow,
           recompute it every other frame instead of dropping the frame rate */
        if (frame % bEvery === 0) {
          var bt = performance.now();
          drawBoundary();
          var el = performance.now() - bt;
          bCost = bCost ? bCost * 0.85 + el * 0.15 : el;
          bEvery = bCost > 6 ? 2 : 1;
        }
        drawMap();
        drawSpark();
        frame++;
        if (frame % (reduced ? 12 : 6) === 0) drawNet();
        uiFrame++;
        if (uiFrame % 3 === 0) syncUI();
      }
      if (st.diverged) { refreshScore(); drawAll(); syncUI(); say("training blew up — lower the learning rate"); }
    } else {
      acc = 0;
    }
  }

  /* ═══ resize ════════════════════════════════════════════════════════════ */
  var rt = null;
  function onResize() {
    if (rt) clearTimeout(rt);
    rt = setTimeout(function () { layout(); drawAll(); }, 80);
  }
  if (window.ResizeObserver) {
    var ro = new ResizeObserver(onResize);
    ro.observe(field);
  }
  window.addEventListener("resize", onResize);

  /* ═══ boot ══════════════════════════════════════════════════════════════ */
  var solvedAnnounced = false;
  rebuild(7);
  layout();
  lrV.textContent = st.lr.toFixed(3);
  chip.textContent = PROBE_HINT;
  if (reduced) {
    /* Reduced motion means nothing moves until asked. An untrained network is
       a grey smear, so run the opening stretch up front (about 100ms) and open
       on a half-learned boundary with a real loss curve, paused. */
    for (var boot = 0; boot < 260 && !st.diverged; boot++) runEpoch();
    refreshScore();
    chip.textContent = "paused for reduced motion — press Resume to keep training";
  }
  drawAll();
  syncUI();
  requestAnimationFrame(tick);

  /* announce the win once per run */
  setInterval(function () {
    if (st.right === data.n && !st.diverged) {
      if (!solvedAnnounced) { solvedAnnounced = true; say("solved — all " + data.n + " points correct at epoch " + st.epoch); }
    } else solvedAnnounced = false;
  }, 700);

})();
</script>
</body>
</html>