-
Notifications
You must be signed in to change notification settings - Fork 1
Expand file tree
/
Copy patheval-driven-development.html
More file actions
1086 lines (1041 loc) · 38.7 KB
/
eval-driven-development.html
File metadata and controls
1086 lines (1041 loc) · 38.7 KB
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
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Eval-Driven Development — End-to-End Process</title>
<style>
:root {
--navy: #1B3A5C;
--blue: #2E75B6;
--light-blue: #4A9BD9;
--sky: #E8F2FB;
--teal: #0D8A72;
--green: #2E7D32;
--amber: #E8890C;
--red: #C62828;
--purple: #6A1B9A;
--gray-50: #FAFAFA;
--gray-100: #F5F5F5;
--gray-200: #EEEEEE;
--gray-300: #E0E0E0;
--gray-500: #9E9E9E;
--gray-700: #616161;
--gray-900: #212121;
--radius: 12px;
--radius-sm: 8px;
--shadow: 0 2px 12px rgba(0,0,0,0.08);
--shadow-lg: 0 8px 32px rgba(0,0,0,0.12);
}
* { margin: 0; padding: 0; box-sizing: border-box; }
body {
font-family: 'Segoe UI', -apple-system, BlinkMacSystemFont, sans-serif;
color: var(--gray-900);
background: var(--gray-50);
line-height: 1.6;
}
/* ── Header ── */
.hero {
background: linear-gradient(135deg, var(--navy) 0%, #1a4a7a 50%, var(--blue) 100%);
color: white;
padding: 4rem 2rem 3.5rem;
text-align: center;
position: relative;
overflow: hidden;
}
.hero::before {
content: '';
position: absolute;
top: -50%;
right: -20%;
width: 600px;
height: 600px;
background: radial-gradient(circle, rgba(255,255,255,0.06) 0%, transparent 70%);
border-radius: 50%;
}
.hero h1 {
font-size: 2.6rem;
font-weight: 700;
margin-bottom: 0.6rem;
letter-spacing: -0.5px;
}
.hero .subtitle {
font-size: 1.15rem;
opacity: 0.85;
max-width: 700px;
margin: 0 auto 1.5rem;
}
.hero .badge {
display: inline-block;
background: rgba(255,255,255,0.15);
border: 1px solid rgba(255,255,255,0.25);
border-radius: 20px;
padding: 0.3rem 1rem;
font-size: 0.85rem;
letter-spacing: 0.3px;
}
.container {
max-width: 1100px;
margin: 0 auto;
padding: 0 2rem;
}
/* ── Overview Strip ── */
.overview {
background: white;
border-radius: var(--radius);
box-shadow: var(--shadow);
margin: -2rem auto 3rem;
padding: 2rem 2.5rem;
max-width: 1100px;
position: relative;
z-index: 2;
}
.overview h2 {
font-size: 1.1rem;
text-transform: uppercase;
letter-spacing: 1.5px;
color: var(--gray-500);
margin-bottom: 1.2rem;
}
.overview p {
font-size: 1.05rem;
color: var(--gray-700);
max-width: 800px;
}
.overview p strong { color: var(--gray-900); }
/* ── Pipeline Visual ── */
.pipeline {
display: flex;
align-items: stretch;
gap: 0;
margin: 2.5rem auto 3.5rem;
max-width: 1100px;
padding: 0 2rem;
position: relative;
}
.pipeline::before {
content: '';
position: absolute;
top: 50%;
left: 4rem;
right: 4rem;
height: 3px;
background: var(--gray-300);
z-index: 0;
transform: translateY(-50%);
}
.pipe-stage {
flex: 1;
text-align: center;
position: relative;
z-index: 1;
padding: 0 0.3rem;
}
.pipe-icon {
width: 64px;
height: 64px;
border-radius: 50%;
margin: 0 auto 0.7rem;
display: flex;
align-items: center;
justify-content: center;
font-size: 1.6rem;
font-weight: 700;
color: white;
box-shadow: 0 3px 12px rgba(0,0,0,0.15);
position: relative;
}
.pipe-icon .num {
position: absolute;
top: -6px;
right: -6px;
background: var(--navy);
color: white;
width: 22px;
height: 22px;
border-radius: 50%;
font-size: 0.7rem;
display: flex;
align-items: center;
justify-content: center;
border: 2px solid white;
}
.pipe-stage .label {
font-size: 0.85rem;
font-weight: 700;
color: var(--gray-900);
margin-bottom: 0.2rem;
}
.pipe-stage .desc {
font-size: 0.75rem;
color: var(--gray-500);
line-height: 1.4;
}
.pipe-arrow {
flex: 0 0 28px;
display: flex;
align-items: center;
justify-content: center;
z-index: 1;
color: var(--gray-300);
font-size: 1.2rem;
}
/* ── Stage Cards ── */
.stages { margin-bottom: 3rem; }
.stage-card {
background: white;
border-radius: var(--radius);
box-shadow: var(--shadow);
margin-bottom: 2rem;
overflow: hidden;
border-left: 5px solid transparent;
}
.stage-header {
display: flex;
align-items: center;
gap: 1.2rem;
padding: 1.5rem 2rem;
cursor: default;
}
.stage-number {
width: 48px;
height: 48px;
border-radius: 50%;
display: flex;
align-items: center;
justify-content: center;
font-size: 1.3rem;
font-weight: 700;
color: white;
flex-shrink: 0;
}
.stage-title-group { flex: 1; }
.stage-title {
font-size: 1.3rem;
font-weight: 700;
color: var(--gray-900);
}
.stage-tagline {
font-size: 0.9rem;
color: var(--gray-500);
}
.stage-badges {
display: flex;
gap: 0.5rem;
flex-wrap: wrap;
}
.stage-badge {
font-size: 0.72rem;
padding: 0.2rem 0.6rem;
border-radius: 4px;
font-weight: 600;
letter-spacing: 0.3px;
text-transform: uppercase;
white-space: nowrap;
}
.stage-body {
padding: 0 2rem 2rem;
}
.stage-body h4 {
font-size: 0.85rem;
text-transform: uppercase;
letter-spacing: 1px;
color: var(--gray-500);
margin-bottom: 0.6rem;
margin-top: 1.2rem;
}
.stage-body h4:first-child { margin-top: 0; }
.stage-body ul {
list-style: none;
padding: 0;
}
.stage-body li {
padding: 0.3rem 0 0.3rem 1.4rem;
position: relative;
font-size: 0.95rem;
color: var(--gray-700);
}
.stage-body li::before {
content: '';
position: absolute;
left: 0;
top: 0.7rem;
width: 6px;
height: 6px;
border-radius: 50%;
background: var(--gray-300);
}
/* Artifact chips */
.artifacts {
display: flex;
flex-wrap: wrap;
gap: 0.5rem;
margin-top: 0.5rem;
}
.artifact {
display: inline-flex;
align-items: center;
gap: 0.35rem;
background: var(--gray-100);
border: 1px solid var(--gray-200);
border-radius: 6px;
padding: 0.3rem 0.7rem;
font-size: 0.8rem;
color: var(--gray-700);
font-weight: 500;
}
.artifact .icon { font-size: 0.9rem; }
/* Prompt chip */
.prompt-chip {
display: inline-block;
background: var(--navy);
color: white;
font-family: 'Cascadia Code', 'Fira Code', monospace;
font-size: 0.8rem;
padding: 0.2rem 0.6rem;
border-radius: 4px;
margin-top: 0.5rem;
}
/* Key insight callout */
.insight {
background: var(--sky);
border-left: 4px solid var(--blue);
border-radius: 0 var(--radius-sm) var(--radius-sm) 0;
padding: 0.8rem 1.2rem;
margin-top: 0.8rem;
font-size: 0.9rem;
color: var(--navy);
}
.insight strong { color: var(--navy); }
/* ── Stage-specific colors ── */
.stage-0 { border-left-color: var(--purple); }
.stage-0 .stage-number { background: var(--purple); }
.stage-0 li::before { background: var(--purple); }
.stage-1 { border-left-color: var(--blue); }
.stage-1 .stage-number { background: var(--blue); }
.stage-1 li::before { background: var(--blue); }
.stage-2 { border-left-color: var(--teal); }
.stage-2 .stage-number { background: var(--teal); }
.stage-2 li::before { background: var(--teal); }
.stage-3 { border-left-color: var(--amber); }
.stage-3 .stage-number { background: var(--amber); }
.stage-3 li::before { background: var(--amber); }
.stage-4 { border-left-color: var(--red); }
.stage-4 .stage-number { background: var(--red); }
.stage-4 li::before { background: var(--red); }
/* Pipe icon colors */
.bg-purple { background: var(--purple); }
.bg-blue { background: var(--blue); }
.bg-teal { background: var(--teal); }
.bg-amber { background: var(--amber); }
.bg-red { background: var(--red); }
/* ── Microsoft Mapping ── */
.ms-mapping {
background: white;
border-radius: var(--radius);
box-shadow: var(--shadow);
padding: 2rem 2.5rem;
margin-bottom: 3rem;
}
.ms-mapping h2 {
font-size: 1.3rem;
font-weight: 700;
color: var(--navy);
margin-bottom: 0.5rem;
}
.ms-mapping .intro {
color: var(--gray-700);
margin-bottom: 1.5rem;
font-size: 0.95rem;
}
.ms-grid {
display: grid;
grid-template-columns: 1fr 1fr;
gap: 1rem;
}
.ms-card {
border: 1px solid var(--gray-200);
border-radius: var(--radius-sm);
padding: 1.2rem;
position: relative;
}
.ms-card-num {
font-size: 0.7rem;
font-weight: 700;
text-transform: uppercase;
letter-spacing: 1px;
margin-bottom: 0.3rem;
}
.ms-card h3 {
font-size: 1rem;
font-weight: 700;
color: var(--gray-900);
margin-bottom: 0.4rem;
}
.ms-card p {
font-size: 0.85rem;
color: var(--gray-700);
margin-bottom: 0.6rem;
}
.ms-maps-to {
display: flex;
flex-wrap: wrap;
gap: 0.3rem;
}
.ms-maps-to span {
font-size: 0.72rem;
padding: 0.15rem 0.5rem;
border-radius: 4px;
font-weight: 600;
color: white;
}
/* ── Architecture Scoping ── */
.arch-section {
background: white;
border-radius: var(--radius);
box-shadow: var(--shadow);
padding: 2rem 2.5rem;
margin-bottom: 3rem;
}
.arch-section h2 {
font-size: 1.3rem;
font-weight: 700;
color: var(--navy);
margin-bottom: 0.5rem;
}
.arch-section .intro {
color: var(--gray-700);
margin-bottom: 1.5rem;
font-size: 0.95rem;
}
.arch-stack {
display: flex;
flex-direction: column;
gap: 0;
}
.arch-layer {
border: 2px solid var(--gray-200);
border-radius: var(--radius-sm);
padding: 1rem 1.5rem;
position: relative;
margin-top: -1px;
}
.arch-layer:first-child { border-radius: var(--radius-sm) var(--radius-sm) 0 0; }
.arch-layer:last-child { border-radius: 0 0 var(--radius-sm) var(--radius-sm); }
.arch-layer h3 {
font-size: 1rem;
font-weight: 700;
margin-bottom: 0.2rem;
}
.arch-layer .arch-desc {
font-size: 0.85rem;
color: var(--gray-700);
margin-bottom: 0.5rem;
}
.arch-tests {
display: flex;
flex-wrap: wrap;
gap: 0.35rem;
}
.arch-tests span {
font-size: 0.75rem;
padding: 0.15rem 0.5rem;
border-radius: 4px;
background: var(--gray-100);
color: var(--gray-700);
border: 1px solid var(--gray-200);
}
.arch-layer.agentic { border-color: var(--red); background: #FFF5F5; }
.arch-layer.agentic h3 { color: var(--red); }
.arch-layer.agentic .arch-tests span { background: #FFEBEE; border-color: #FFCDD2; color: var(--red); }
.arch-layer.rag { border-color: var(--teal); background: #F0FBF8; }
.arch-layer.rag h3 { color: var(--teal); }
.arch-layer.rag .arch-tests span { background: #E0F2EF; border-color: #B2DFDB; color: var(--teal); }
.arch-layer.prompt { border-color: var(--blue); background: #F5F9FF; }
.arch-layer.prompt h3 { color: var(--blue); }
.arch-layer.prompt .arch-tests span { background: #E3F0FC; border-color: #BBDEFB; color: var(--blue); }
/* ── Iteration Loop Visual ── */
.loop-section {
background: white;
border-radius: var(--radius);
box-shadow: var(--shadow);
padding: 2rem 2.5rem;
margin-bottom: 3rem;
}
.loop-section h2 {
font-size: 1.3rem;
font-weight: 700;
color: var(--navy);
margin-bottom: 0.5rem;
}
.loop-section .intro {
color: var(--gray-700);
margin-bottom: 1.5rem;
font-size: 0.95rem;
}
.loop-diagram {
display: flex;
align-items: center;
justify-content: center;
gap: 0.5rem;
flex-wrap: wrap;
padding: 1.5rem 0;
}
.loop-node {
background: var(--navy);
color: white;
padding: 0.7rem 1.2rem;
border-radius: var(--radius-sm);
font-size: 0.9rem;
font-weight: 600;
text-align: center;
min-width: 110px;
}
.loop-arrow {
font-size: 1.4rem;
color: var(--gray-400);
}
.loop-node.change { background: var(--teal); }
.loop-node.run { background: var(--amber); }
.loop-node.interpret { background: var(--red); }
.loop-node.fix { background: var(--blue); }
.loop-caption {
text-align: center;
color: var(--gray-500);
font-size: 0.85rem;
margin-top: 0.5rem;
}
/* ── Verdicts Strip ── */
.verdict-section {
background: white;
border-radius: var(--radius);
box-shadow: var(--shadow);
padding: 2rem 2.5rem;
margin-bottom: 3rem;
}
.verdict-section h2 {
font-size: 1.3rem;
font-weight: 700;
color: var(--navy);
margin-bottom: 1.2rem;
}
.verdicts {
display: grid;
grid-template-columns: repeat(4, 1fr);
gap: 1rem;
}
.verdict-card {
border-radius: var(--radius-sm);
padding: 1.2rem;
text-align: center;
}
.verdict-card h3 {
font-size: 1.1rem;
font-weight: 800;
margin-bottom: 0.3rem;
}
.verdict-card p {
font-size: 0.82rem;
line-height: 1.4;
}
.verdict-card.ship { background: #E8F5E9; color: var(--green); }
.verdict-card.ship-gaps { background: #FFF3E0; color: #E65100; }
.verdict-card.iterate { background: #FFF8E1; color: #F57F17; }
.verdict-card.block { background: #FFEBEE; color: var(--red); }
.verdict-card .verdict-threshold {
font-size: 0.75rem;
margin-top: 0.5rem;
opacity: 0.7;
}
/* ── Footer ── */
.footer {
text-align: center;
padding: 2rem;
color: var(--gray-500);
font-size: 0.85rem;
border-top: 1px solid var(--gray-200);
margin-top: 2rem;
}
.footer a {
color: var(--blue);
text-decoration: none;
}
.footer a:hover { text-decoration: underline; }
/* ── Two-col inside stage cards ── */
.two-col {
display: grid;
grid-template-columns: 1fr 1fr;
gap: 1.5rem;
}
/* ── Responsive ── */
@media (max-width: 768px) {
.hero h1 { font-size: 1.8rem; }
.pipeline { flex-direction: column; align-items: center; gap: 0.5rem; }
.pipeline::before { display: none; }
.pipe-arrow { transform: rotate(90deg); }
.ms-grid { grid-template-columns: 1fr; }
.verdicts { grid-template-columns: 1fr 1fr; }
.two-col { grid-template-columns: 1fr; }
}
</style>
</head>
<body>
<!-- ═══════════ HERO ═══════════ -->
<div class="hero">
<h1>Eval-Driven Development</h1>
<p class="subtitle">The end-to-end process for building AI agents that work — from defining "good" to shipping with confidence.</p>
<span class="badge">microsoft/eval-guide · 5 Stages · Grounded in Microsoft's Evaluation Framework</span>
</div>
<!-- ═══════════ OVERVIEW ═══════════ -->
<div class="container">
<div class="overview">
<h2>The Idea</h2>
<p>Most teams build agents first and test later — if they test at all. <strong>Eval-driven development flips this:</strong> define what "good" looks like before you build, write test cases before you write the prompt, then iterate with data instead of intuition. The eval suite becomes your specification, your regression safety net, and your ship/no-ship decision gate.</p>
</div>
</div>
<!-- ═══════════ PIPELINE VISUAL ═══════════ -->
<div class="pipeline">
<div class="pipe-stage">
<div class="pipe-icon bg-purple">
<span style="font-size:1.4rem">🔍</span>
<span class="num">0</span>
</div>
<div class="label">Discover</div>
<div class="desc">Define what the agent does & what success looks like</div>
</div>
<div class="pipe-arrow">→</div>
<div class="pipe-stage">
<div class="pipe-icon bg-blue">
<span style="font-size:1.4rem">📋</span>
<span class="num">1</span>
</div>
<div class="label">Plan</div>
<div class="desc">Zone Model, scenarios, acceptance criteria, success metrics</div>
</div>
<div class="pipe-arrow">→</div>
<div class="pipe-stage">
<div class="pipe-icon bg-teal">
<span style="font-size:1.4rem">⚙</span>
<span class="num">2</span>
</div>
<div class="label">Generate</div>
<div class="desc">Zone-grouped test cases with VERIFY highlights</div>
</div>
<div class="pipe-arrow">→</div>
<div class="pipe-stage">
<div class="pipe-icon bg-amber">
<span style="font-size:1.4rem">▶</span>
<span class="num">3</span>
</div>
<div class="label">Run</div>
<div class="desc">Execute tests against the live agent</div>
</div>
<div class="pipe-arrow">→</div>
<div class="pipe-stage">
<div class="pipe-icon bg-red">
<span style="font-size:1.4rem">📊</span>
<span class="num">4</span>
</div>
<div class="label">Interpret</div>
<div class="desc">Success metrics status, human judgement, triage</div>
</div>
</div>
<!-- ═══════════ STAGE CARDS ═══════════ -->
<div class="container stages">
<!-- Stage 0 -->
<div class="stage-card stage-0">
<div class="stage-header">
<div class="stage-number">0</div>
<div class="stage-title-group">
<div class="stage-title">Discover</div>
<div class="stage-tagline">Articulate what "good" looks like — before writing a single test</div>
</div>
<div class="stage-badges">
<span class="stage-badge" style="background:#F3E5F5;color:var(--purple)">No agent needed</span>
</div>
</div>
<div class="stage-body">
<div class="two-col">
<div>
<h4>What happens</h4>
<ul>
<li>Define the agent's purpose, users, and context</li>
<li>Inventory knowledge sources and capabilities</li>
<li>Set explicit boundaries — what it must NOT do</li>
<li>Establish measurable success criteria</li>
<li>Assess risk profile (low / medium / high)</li>
<li>Determine if behavior varies by user role</li>
</ul>
</div>
<div>
<h4>Output: Agent Vision</h4>
<ul>
<li><strong>Purpose</strong> — one sentence on the core problem it solves</li>
<li><strong>Users</strong> — who, in what context</li>
<li><strong>Knowledge & Data</strong> — planned or actual sources</li>
<li><strong>Core Capabilities</strong> — 3-5 things the agent should do</li>
<li><strong>Boundaries</strong> — what it must never attempt</li>
<li><strong>Success Criteria</strong> — these become your eval targets</li>
</ul>
</div>
</div>
<div class="insight">
<strong>Why this matters:</strong> Most teams have never written down what "good" looks like for their agent. This Agent Vision becomes the foundation for everything — the eval plan, the test cases, and eventually the agent's system prompt. Every test you write ties back to what you define here.
</div>
<h4>Artifacts</h4>
<div class="artifacts">
<span class="artifact"><span class="icon">📄</span> Agent Vision document</span>
<span class="artifact"><span class="icon">📊</span> Risk profile assessment</span>
</div>
<div class="prompt-chip">#eval-guide</div>
</div>
</div>
<!-- Stage 1 -->
<div class="stage-card stage-1">
<div class="stage-header">
<div class="stage-number">1</div>
<div class="stage-title-group">
<div class="stage-title">Plan</div>
<div class="stage-tagline">Align business and dev on priorities, expectations, and what "good enough" means</div>
</div>
<div class="stage-badges">
<span class="stage-badge" style="background:#E3F2FD;color:var(--blue)">No agent needed</span>
</div>
</div>
<div class="stage-body">
<div class="two-col">
<div>
<h4>Zone Model for Prioritization</h4>
<ul>
<li>Classify scenarios into 3 zones based on <strong>Priority + Value + Effort</strong></li>
<li><strong>Zone 1: "This Must Work"</strong> — highest value, full investment, clear acceptance criteria. Eval every change.</li>
<li><strong>Zone 2: "Real but Lower Priority"</strong> — scoped capabilities, intentionally lower expectations. Eval before releases.</li>
<li><strong>Zone 3: "Exploratory"</strong> — emergent patterns from production usage. Eval periodically.</li>
<li>Drag scenarios between zones in the interactive dashboard to reprioritize</li>
</ul>
</div>
<div>
<h4>Dashboard sections</h4>
<ul>
<li><strong>Scenarios & Acceptance Criteria:</strong> Each scenario gets Pass = and Fail = conditions specific enough for two reviewers to agree</li>
<li><strong>Quality Dimensions to Test:</strong> Group scenarios by dimension (Policy Accuracy, Hallucination Prevention, Privacy Protection, etc.) with draggable chips</li>
<li><strong>Success Metrics:</strong> Zone + Scenario + Quality Dimension + Target Pass Rate — the agreement on what "good enough" means</li>
</ul>
</div>
</div>
<div class="insight">
<strong>The Zone Model is the shared language.</strong> Business wants to ship. Dev wants quality. The Zone Model gives both a way to set expectations together — Zone 1 must work, Zone 2 is intentionally lower investment, Zone 3 is exploration. If everything is Zone 1, nothing is prioritized.
</div>
<h4>Artifacts</h4>
<div class="artifacts">
<span class="artifact"><span class="icon">📄</span> Eval plan report (.docx) with Zone Model</span>
<span class="artifact"><span class="icon">🎯</span> Quality dimension mapping</span>
<span class="artifact"><span class="icon">✅</span> Success metrics (target pass rates per scenario)</span>
<span class="artifact"><span class="icon">📊</span> Acceptance criteria (Pass/Fail)</span>
</div>
<div class="prompt-chip">#eval-suite-planner</div>
</div>
</div>
<!-- Stage 2 -->
<div class="stage-card stage-2">
<div class="stage-header">
<div class="stage-number">2</div>
<div class="stage-title-group">
<div class="stage-title">Generate & Baseline</div>
<div class="stage-tagline">Produce zone-aware test cases with factual verification markers</div>
</div>
<div class="stage-badges">
<span class="stage-badge" style="background:#E0F2EF;color:var(--teal)">No agent needed</span>
</div>
</div>
<div class="stage-body">
<div class="two-col">
<div>
<h4>Dashboard features</h4>
<ul>
<li><strong>Zone-colored tabs:</strong> Each quality dimension tab is color-coded by dominant zone (red=Zone 1, yellow=Zone 2, blue=Zone 3)</li>
<li><strong>Scenario groups:</strong> Test cases grouped by scenario, showing zone badge + acceptance criteria (Pass in green, Fail in red)</li>
<li><strong>VERIFY highlights:</strong> AI-generated factual content marked in yellow — click to edit and check against real knowledge sources</li>
<li><strong>Suggested test methods:</strong> Set at the dimension level, not per-row</li>
</ul>
</div>
<div>
<h4>What gets generated</h4>
<ul>
<li>Test cases organized by quality dimension with nested scenario groups</li>
<li>Expected responses with <span style="background:#fff3cd;padding:0 0.3rem;border-radius:2px;font-size:0.85rem;color:#856404;">[VERIFY]</span> markers for facts to check</li>
<li>Acceptance criteria carried forward from Stage 1 for each scenario group</li>
<li>CSV files importable into Copilot Studio evaluation</li>
</ul>
</div>
</div>
<div class="insight">
<strong>What you'd miss without this:</strong> Most teams write 5-10 happy-path tests. The generator produces 20-30 cases across adversarial attacks, hallucination traps, robustness tests, and safety boundaries. VERIFY markers flag AI-generated facts so you don't ship untested claims.
</div>
<h4>Artifacts</h4>
<div class="artifacts">
<span class="artifact"><span class="icon">📄</span> CSV files per quality dimension</span>
<span class="artifact"><span class="icon">📄</span> Test case report (.docx) with Zone Model</span>
<span class="artifact"><span class="icon">⚠</span> VERIFY-flagged factual content</span>
</div>
<div class="prompt-chip">#eval-generator</div>
</div>
</div>
<!-- Stage 3 -->
<div class="stage-card stage-3">
<div class="stage-header">
<div class="stage-number">3</div>
<div class="stage-title-group">
<div class="stage-title">Run</div>
<div class="stage-tagline">Execute tests against the live agent and collect scores</div>
</div>
<div class="stage-badges">
<span class="stage-badge" style="background:#FFF3E0;color:var(--amber)">Requires running agent</span>
</div>
</div>
<div class="stage-body">
<div class="two-col">
<div>
<h4>What happens</h4>
<ul>
<li>Import CSV test sets into Copilot Studio's Evaluation tab</li>
<li>Run evaluation against the published agent</li>
<li>Copilot Studio scores each case using the specified test method</li>
<li>Export results CSV with scores, pass/fail, and explanations</li>
<li>Alternative: run via DirectLine API with eval-runner.js for CI/CD integration</li>
</ul>
</div>
<div>
<h4>How scoring works</h4>
<ul>
<li><strong>General Quality:</strong> LLM judge scores on Relevance, Groundedness, Completeness, Abstention (0.0-1.0)</li>
<li><strong>Compare Meaning:</strong> Semantic equivalence between expected and actual (0.0-1.0)</li>
<li><strong>Keyword Match:</strong> Code-based string matching — are required terms present?</li>
<li><strong>Exact Match:</strong> Code-based string equality</li>
<li><strong>Custom:</strong> Domain-specific criteria with your own labels</li>
</ul>
</div>
</div>
<div class="insight">
<strong>Run at least 3 times.</strong> AI agents are non-deterministic. A single run might mislead you. Microsoft recommends 3 runs minimum: ±5% variance is normal, ±10% needs investigation. Always export results — they're only available in Copilot Studio for 89 days.
</div>
<h4>Artifacts</h4>
<div class="artifacts">
<span class="artifact"><span class="icon">📈</span> Results CSV with scores</span>
<span class="artifact"><span class="icon">📊</span> Per-method pass/fail summary</span>
</div>
</div>
</div>
<!-- Stage 4 -->
<div class="stage-card stage-4">
<div class="stage-header">
<div class="stage-number">4</div>
<div class="stage-title-group">
<div class="stage-title">Interpret & Improve</div>
<div class="stage-tagline">Compare results against success metrics, triage failures, decide what to fix</div>
</div>
<div class="stage-badges">
<span class="stage-badge" style="background:#FFEBEE;color:var(--red)">Requires eval results</span>
</div>
</div>
<div class="stage-body">
<div class="two-col">
<div>
<h4>Dashboard features</h4>
<ul>
<li><strong>Success Metrics Status:</strong> Zone summary cards (e.g., Zone 1: 5/8 met) + scenario table comparing targets vs actuals</li>
<li><strong>Expandable rows:</strong> Click any scenario to see test case details — question, expected, actual, score, pass/fail</li>
<li><strong>Human Judgement:</strong> Agree/Disagree per test case — override the LLM judge when you know better</li>
<li><strong>Failure Triage:</strong> Root cause classification (Eval Setup / Agent Config / Platform Limitation)</li>
<li><strong>Top Actions & Patterns:</strong> Prioritized fixes + systemic issue detection</li>
</ul>
</div>
<div>
<h4>Human-in-the-loop</h4>
<ul>
<li>Each test case shows the LLM judge's <strong>Rationale</strong> explaining the score</li>
<li>Select <strong>Disagree</strong> to override — automatically moves the case to "Eval Setup — Human Disagrees" in triage</li>
<li>This prevents wasted effort fixing the agent when the eval setup was wrong</li>
<li><strong>At least 20% of failures in a new eval are eval setup issues, not agent issues</strong></li>
</ul>
</div>
</div>
<div class="insight">
<strong>No more SHIP/BLOCK verdicts.</strong> Instead, you see exactly which success metrics were met per zone. Zone 1: 5/8 tells you more than "ITERATE" — you know which 3 scenarios need work and what the target gap is. Deploy when Zone 1 targets are met at acceptable rates.
</div>
<h4>Artifacts</h4>
<div class="artifacts">
<span class="artifact"><span class="icon">📄</span> Triage report (.docx) with zone breakdown</span>
<span class="artifact"><span class="icon">📊</span> Success metrics status (target vs actual)</span>
<span class="artifact"><span class="icon">🛠</span> Prioritized action items</span>
<span class="artifact"><span class="icon">📈</span> Zone-aware pattern analysis</span>
</div>
<div class="prompt-chip">#eval-result-interpreter</div>
<span style="margin-left:0.5rem" class="prompt-chip">#eval-triage-and-improvement</span>
</div>
</div>
</div>
<!-- ═══════════ ITERATION LOOP ═══════════ -->
<div class="container">
<div class="loop-section">
<h2>The Iteration Loop</h2>
<p class="intro">Stages 0-2 happen once (with expansions later). Stages 3-4 repeat on every change. This is where eval-driven development becomes a habit — not a one-time activity.</p>
<div class="loop-diagram">
<div class="loop-node change">Make a Change</div>
<span class="loop-arrow">→</span>
<div class="loop-node run">Run Evals</div>
<span class="loop-arrow">→</span>
<div class="loop-node interpret">Interpret Results</div>
<span class="loop-arrow">→</span>
<div class="loop-node fix">Fix or Ship</div>
<span class="loop-arrow">↺</span>
</div>
<div class="loop-caption">Changed the prompt? Run the eval. Updated knowledge sources? Run the eval. New model version? Run the eval.</div>
<div style="margin-top:1.5rem;display:grid;grid-template-columns:1fr 1fr 1fr;gap:1rem;">
<div style="background:#FFEBEE;border-radius:var(--radius-sm);padding:1rem;text-align:center;">
<div style="font-size:0.75rem;font-weight:700;color:var(--red);text-transform:uppercase;letter-spacing:0.5px;">Zone 1</div>
<div style="font-size:0.85rem;font-weight:700;margin:0.3rem 0 0.2rem">Every Change</div>
<div style="font-size:0.78rem;color:var(--gray-500)">Regression baseline — failures here block deployment</div>
</div>
<div style="background:#FFF8E1;border-radius:var(--radius-sm);padding:1rem;text-align:center;">
<div style="font-size:0.75rem;font-weight:700;color:#996600;text-transform:uppercase;letter-spacing:0.5px;">Zone 2</div>
<div style="font-size:0.85rem;font-weight:700;margin:0.3rem 0 0.2rem">Before Releases</div>
<div style="font-size:0.78rem;color:var(--gray-500)">Track improvement trends — failures are signals, not emergencies</div>
</div>
<div style="background:#E3F2FD;border-radius:var(--radius-sm);padding:1rem;text-align:center;">
<div style="font-size:0.75rem;font-weight:700;color:var(--blue);text-transform:uppercase;letter-spacing:0.5px;">Zone 3</div>
<div style="font-size:0.85rem;font-weight:700;margin:0.3rem 0 0.2rem">Periodic / Discovery</div>
<div style="font-size:0.78rem;color:var(--gray-500)">Watch for rising usage — promote to higher zones as patterns emerge</div>
</div>
</div>
</div>
</div>
<!-- ═══════════ VERDICTS ═══════════ -->
<div class="container">
<div class="verdict-section">
<h2>The Deploy Decision: Zone-Based</h2>
<p style="color:var(--gray-700);margin-bottom:1.2rem;font-size:0.95rem;">Instead of a single SHIP/BLOCK verdict, deployment decisions are made per zone. The question is: "At what point does this agent create more value than it costs?"</p>
<div class="verdicts">
<div class="verdict-card" style="background:#FFEBEE;color:var(--red);">
<h3>Zone 1 Met?</h3>
<p>All "must work" scenarios meet their targets. Safety + core business passing at acceptable rates.</p>
<div class="verdict-threshold">Deploy when Zone 1 targets are met</div>
</div>
<div class="verdict-card" style="background:#FFF8E1;color:#996600;">
<h3>Zone 2 Tracked</h3>
<p>Lower-priority gaps are documented, not blockers. Improvement signals for the next iteration.</p>
<div class="verdict-threshold">Zone 2 gaps don't block deployment</div>
</div>
<div class="verdict-card" style="background:#E3F2FD;color:var(--blue);">
<h3>Zone 3 Explored</h3>
<p>Emerging patterns monitored. Promote rising use cases to higher zones over time.</p>
<div class="verdict-threshold">Research & discovery, not operations</div>
</div>
<div class="verdict-card" style="background:#F3E5F5;color:var(--purple);">
<h3>Human Override</h3>
<p>Disagree with the LLM judge? Override any test case. Eval setup issues don't count against the agent.</p>
<div class="verdict-threshold">Human judgement > automated scoring</div>
</div>
</div>
</div>
</div>
<!-- ═══════════ ARCHITECTURE SCOPING ═══════════ -->
<div class="container">
<div class="arch-section">
<h2>Architecture-Aware Eval Scoping</h2>
<p class="intro">Different agent architectures need different eval layers. The toolkit automatically scopes what to test based on complexity — each layer inherits all tests from layers below it.</p>
<div class="arch-stack">
<div class="arch-layer agentic">
<h3>Agentic (multi-step, tool use, orchestration)</h3>
<div class="arch-desc">Agent calls APIs, uses connectors, chains actions, makes decisions across steps</div>
<div class="arch-tests">
<span>Tool selection accuracy</span>
<span>Action correctness</span>
<span>Error recovery</span>
<span>Multi-turn context retention</span>
<span>Task completion rate</span>
</div>
</div>