sofhiaazzhr Claude Opus 5 commited on
Commit
bd6659b
·
1 Parent(s): 60cf4a8

[NOTICKET] chore: add missing planner and readiness eval results

Browse files

Fills gaps in the committed eval history (readiness 2026-06-23 and 2026-07-02,
planner 2026-07-23). Same shape as the already-tracked results; the timestamped
files are a standing baseline and are never overwritten.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

eval/planner/results/planner_result_2026-07-23_134947.json ADDED
@@ -0,0 +1,662 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "timestamp": "2026-07-23_134947",
3
+ "total": 27,
4
+ "passed": 20,
5
+ "cases": [
6
+ {
7
+ "id": "count_zero_pa",
8
+ "category": "count",
9
+ "lang": "en",
10
+ "carried_over": true,
11
+ "question": "how many records have PA_Percent = 0?",
12
+ "passed": true,
13
+ "checks": [
14
+ {
15
+ "check": "select_agg=count",
16
+ "ok": true,
17
+ "detail": "aggs=['count']"
18
+ },
19
+ {
20
+ "check": "has_filter",
21
+ "ok": true,
22
+ "detail": "filter_present=True"
23
+ },
24
+ {
25
+ "check": "no_group_by",
26
+ "ok": true,
27
+ "detail": "group_by_present=False"
28
+ }
29
+ ],
30
+ "error": null,
31
+ "latency_ms": 3591,
32
+ "tokens": 14570
33
+ },
34
+ {
35
+ "id": "count_mttr_gt20_id",
36
+ "category": "count",
37
+ "lang": "id",
38
+ "carried_over": true,
39
+ "question": "berapa banyak record dengan MTTR di atas 20?",
40
+ "passed": true,
41
+ "checks": [
42
+ {
43
+ "check": "select_agg=count",
44
+ "ok": true,
45
+ "detail": "aggs=['count']"
46
+ },
47
+ {
48
+ "check": "has_filter",
49
+ "ok": true,
50
+ "detail": "filter_present=True"
51
+ }
52
+ ],
53
+ "error": null,
54
+ "latency_ms": 2154,
55
+ "tokens": 14571
56
+ },
57
+ {
58
+ "id": "count_section_hauler",
59
+ "category": "count",
60
+ "lang": "en",
61
+ "carried_over": true,
62
+ "question": "how many rows are in section OB HAULER?",
63
+ "passed": true,
64
+ "checks": [
65
+ {
66
+ "check": "select_agg=count",
67
+ "ok": true,
68
+ "detail": "aggs=['count']"
69
+ },
70
+ {
71
+ "check": "has_filter",
72
+ "ok": true,
73
+ "detail": "filter_present=True"
74
+ }
75
+ ],
76
+ "error": null,
77
+ "latency_ms": 2063,
78
+ "tokens": 14573
79
+ },
80
+ {
81
+ "id": "rank_units_worst_pa_id",
82
+ "category": "ranking",
83
+ "lang": "id",
84
+ "carried_over": false,
85
+ "question": "5 unit dengan PA terburuk?",
86
+ "passed": true,
87
+ "checks": [
88
+ {
89
+ "check": "group_by",
90
+ "ok": true,
91
+ "detail": "group_by_present=True"
92
+ },
93
+ {
94
+ "check": "group_by_col=Equipment_Number",
95
+ "ok": true,
96
+ "detail": "want_id=c_equipment_number"
97
+ },
98
+ {
99
+ "check": "select_agg=avg",
100
+ "ok": true,
101
+ "detail": "aggs=['avg']"
102
+ },
103
+ {
104
+ "check": "order_dir=asc",
105
+ "ok": true,
106
+ "detail": "dirs=['asc']"
107
+ },
108
+ {
109
+ "check": "limit=5",
110
+ "ok": true,
111
+ "detail": "limits=[5]"
112
+ }
113
+ ],
114
+ "error": null,
115
+ "latency_ms": 2664,
116
+ "tokens": 14784
117
+ },
118
+ {
119
+ "id": "rank_models_top_mttr_id",
120
+ "category": "ranking",
121
+ "lang": "id",
122
+ "carried_over": true,
123
+ "question": "top 3 model dengan MTTR tertinggi?",
124
+ "passed": true,
125
+ "checks": [
126
+ {
127
+ "check": "group_by",
128
+ "ok": true,
129
+ "detail": "group_by_present=True"
130
+ },
131
+ {
132
+ "check": "group_by_col=Model_Unit",
133
+ "ok": true,
134
+ "detail": "want_id=c_model_unit"
135
+ },
136
+ {
137
+ "check": "select_agg=avg",
138
+ "ok": true,
139
+ "detail": "aggs=['avg']"
140
+ },
141
+ {
142
+ "check": "order_dir=desc",
143
+ "ok": true,
144
+ "detail": "dirs=['desc']"
145
+ },
146
+ {
147
+ "check": "limit=3",
148
+ "ok": true,
149
+ "detail": "limits=[3]"
150
+ }
151
+ ],
152
+ "error": null,
153
+ "latency_ms": 2757,
154
+ "tokens": 14719
155
+ },
156
+ {
157
+ "id": "rank_sections_lowest_pa_en",
158
+ "category": "ranking",
159
+ "lang": "en",
160
+ "carried_over": true,
161
+ "question": "which section has the lowest average PA?",
162
+ "passed": false,
163
+ "checks": [
164
+ {
165
+ "check": "group_by",
166
+ "ok": false,
167
+ "detail": "group_by_present=False"
168
+ },
169
+ {
170
+ "check": "group_by_col=Section",
171
+ "ok": false,
172
+ "detail": "want_id=c_section"
173
+ },
174
+ {
175
+ "check": "select_agg=avg",
176
+ "ok": false,
177
+ "detail": "aggs=[]"
178
+ },
179
+ {
180
+ "check": "order_dir=asc",
181
+ "ok": false,
182
+ "detail": "dirs=[]"
183
+ }
184
+ ],
185
+ "error": null,
186
+ "latency_ms": 3366,
187
+ "tokens": 14834
188
+ },
189
+ {
190
+ "id": "rank_units_most_breakdown_id",
191
+ "category": "ranking",
192
+ "lang": "id",
193
+ "carried_over": false,
194
+ "question": "unit mana yang paling sering breakdown?",
195
+ "passed": true,
196
+ "checks": [
197
+ {
198
+ "check": "group_by",
199
+ "ok": true,
200
+ "detail": "group_by_present=True"
201
+ },
202
+ {
203
+ "check": "group_by_col=Equipment_Number",
204
+ "ok": true,
205
+ "detail": "want_id=c_equipment_number"
206
+ },
207
+ {
208
+ "check": "order_dir=desc",
209
+ "ok": true,
210
+ "detail": "dirs=['desc']"
211
+ }
212
+ ],
213
+ "error": null,
214
+ "latency_ms": 2925,
215
+ "tokens": 14736
216
+ },
217
+ {
218
+ "id": "rank_units_worst_pa_en",
219
+ "category": "ranking",
220
+ "lang": "en",
221
+ "carried_over": false,
222
+ "question": "list the 10 worst units by availability",
223
+ "passed": true,
224
+ "checks": [
225
+ {
226
+ "check": "group_by",
227
+ "ok": true,
228
+ "detail": "group_by_present=True"
229
+ },
230
+ {
231
+ "check": "group_by_col=Equipment_Number",
232
+ "ok": true,
233
+ "detail": "want_id=c_equipment_number"
234
+ },
235
+ {
236
+ "check": "select_agg=avg",
237
+ "ok": true,
238
+ "detail": "aggs=['avg']"
239
+ },
240
+ {
241
+ "check": "order_dir=asc",
242
+ "ok": true,
243
+ "detail": "dirs=['asc']"
244
+ },
245
+ {
246
+ "check": "limit=10",
247
+ "ok": true,
248
+ "detail": "limits=[10]"
249
+ }
250
+ ],
251
+ "error": null,
252
+ "latency_ms": 3001,
253
+ "tokens": 14758
254
+ },
255
+ {
256
+ "id": "fuzzy_model_777_id",
257
+ "category": "fuzzy_filter",
258
+ "lang": "id",
259
+ "carried_over": false,
260
+ "question": "berapa banyak model 777?",
261
+ "passed": false,
262
+ "checks": [
263
+ {
264
+ "check": "select_agg=count",
265
+ "ok": true,
266
+ "detail": "aggs=['count']"
267
+ },
268
+ {
269
+ "check": "filter_op=like",
270
+ "ok": false,
271
+ "detail": "ops=['=']"
272
+ },
273
+ {
274
+ "check": "no_filter_op=in",
275
+ "ok": true,
276
+ "detail": "ops=['=']"
277
+ }
278
+ ],
279
+ "error": null,
280
+ "latency_ms": 1831,
281
+ "tokens": 14578
282
+ },
283
+ {
284
+ "id": "fuzzy_model_hd785_id",
285
+ "category": "fuzzy_filter",
286
+ "lang": "id",
287
+ "carried_over": true,
288
+ "question": "berapa banyak unit HD785?",
289
+ "passed": false,
290
+ "checks": [
291
+ {
292
+ "check": "select_agg=count",
293
+ "ok": true,
294
+ "detail": "aggs=['count']"
295
+ },
296
+ {
297
+ "check": "filter_op=like",
298
+ "ok": false,
299
+ "detail": "ops=['=']"
300
+ }
301
+ ],
302
+ "error": null,
303
+ "latency_ms": 2219,
304
+ "tokens": 14578
305
+ },
306
+ {
307
+ "id": "fuzzy_model_ex_en",
308
+ "category": "fuzzy_filter",
309
+ "lang": "en",
310
+ "carried_over": false,
311
+ "question": "how many EX excavator units are there?",
312
+ "passed": true,
313
+ "checks": [
314
+ {
315
+ "check": "filter_op=like",
316
+ "ok": true,
317
+ "detail": "ops=['like']"
318
+ }
319
+ ],
320
+ "error": null,
321
+ "latency_ms": 5243,
322
+ "tokens": 29547
323
+ },
324
+ {
325
+ "id": "agg_pa_per_section_id",
326
+ "category": "aggregate",
327
+ "lang": "id",
328
+ "carried_over": true,
329
+ "question": "berapa rata-rata PA per section?",
330
+ "passed": false,
331
+ "checks": [
332
+ {
333
+ "check": "group_by",
334
+ "ok": false,
335
+ "detail": "group_by_present=False"
336
+ },
337
+ {
338
+ "check": "group_by_col=Section",
339
+ "ok": false,
340
+ "detail": "want_id=c_section"
341
+ },
342
+ {
343
+ "check": "select_agg=avg",
344
+ "ok": false,
345
+ "detail": "aggs=[]"
346
+ }
347
+ ],
348
+ "error": null,
349
+ "latency_ms": 2817,
350
+ "tokens": 14757
351
+ },
352
+ {
353
+ "id": "agg_mttr_per_model_en",
354
+ "category": "aggregate",
355
+ "lang": "en",
356
+ "carried_over": true,
357
+ "question": "what is the average MTTR per model unit?",
358
+ "passed": false,
359
+ "checks": [
360
+ {
361
+ "check": "group_by",
362
+ "ok": false,
363
+ "detail": "group_by_present=False"
364
+ },
365
+ {
366
+ "check": "group_by_col=Model_Unit",
367
+ "ok": false,
368
+ "detail": "want_id=c_model_unit"
369
+ },
370
+ {
371
+ "check": "select_agg=avg",
372
+ "ok": false,
373
+ "detail": "aggs=[]"
374
+ }
375
+ ],
376
+ "error": null,
377
+ "latency_ms": 2859,
378
+ "tokens": 14773
379
+ },
380
+ {
381
+ "id": "agg_downhour_per_group_id",
382
+ "category": "aggregate",
383
+ "lang": "id",
384
+ "carried_over": true,
385
+ "question": "total down hour per equipment group?",
386
+ "passed": false,
387
+ "checks": [
388
+ {
389
+ "check": "group_by",
390
+ "ok": false,
391
+ "detail": "group_by_present=False"
392
+ },
393
+ {
394
+ "check": "group_by_col=Equipment_Group_ID",
395
+ "ok": false,
396
+ "detail": "want_id=c_equipment_group_id"
397
+ },
398
+ {
399
+ "check": "select_agg=sum",
400
+ "ok": false,
401
+ "detail": "aggs=[]"
402
+ }
403
+ ],
404
+ "error": null,
405
+ "latency_ms": 2855,
406
+ "tokens": 14769
407
+ },
408
+ {
409
+ "id": "desc_mttr_stats_id",
410
+ "category": "descriptive",
411
+ "lang": "id",
412
+ "carried_over": true,
413
+ "question": "berikan ringkasan statistik MTTR",
414
+ "passed": true,
415
+ "checks": [
416
+ {
417
+ "check": "has_tool=analyze_descriptive",
418
+ "ok": true,
419
+ "detail": "tools=['analyze_descriptive', 'check_data', 'retrieve_data']"
420
+ }
421
+ ],
422
+ "error": null,
423
+ "latency_ms": 2928,
424
+ "tokens": 14714
425
+ },
426
+ {
427
+ "id": "desc_pa_stats_en",
428
+ "category": "descriptive",
429
+ "lang": "en",
430
+ "carried_over": true,
431
+ "question": "give me the summary statistics for PA_Percent",
432
+ "passed": true,
433
+ "checks": [
434
+ {
435
+ "check": "has_tool=analyze_descriptive",
436
+ "ok": true,
437
+ "detail": "tools=['analyze_descriptive', 'check_data', 'retrieve_data']"
438
+ }
439
+ ],
440
+ "error": null,
441
+ "latency_ms": 2862,
442
+ "tokens": 14704
443
+ },
444
+ {
445
+ "id": "corr_mttr_pa_id",
446
+ "category": "correlation",
447
+ "lang": "id",
448
+ "carried_over": true,
449
+ "question": "apakah ada korelasi antara MTTR dan PA?",
450
+ "passed": true,
451
+ "checks": [
452
+ {
453
+ "check": "has_tool=analyze_correlation",
454
+ "ok": true,
455
+ "detail": "tools=['analyze_correlation', 'check_data', 'retrieve_data']"
456
+ }
457
+ ],
458
+ "error": null,
459
+ "latency_ms": 3171,
460
+ "tokens": 14866
461
+ },
462
+ {
463
+ "id": "corr_freq_pa_en",
464
+ "category": "correlation",
465
+ "lang": "en",
466
+ "carried_over": true,
467
+ "question": "is breakdown frequency correlated with availability?",
468
+ "passed": true,
469
+ "checks": [
470
+ {
471
+ "check": "has_tool=analyze_correlation",
472
+ "ok": true,
473
+ "detail": "tools=['analyze_correlation', 'check_data', 'retrieve_data']"
474
+ }
475
+ ],
476
+ "error": null,
477
+ "latency_ms": 3259,
478
+ "tokens": 14921
479
+ },
480
+ {
481
+ "id": "trend_pa_daily_id",
482
+ "category": "trend",
483
+ "lang": "id",
484
+ "carried_over": true,
485
+ "question": "bagaimana trend PA harian?",
486
+ "passed": false,
487
+ "checks": [],
488
+ "error": "ValidationError: 6 validation errors for TaskList\ntasks.2.id\n Field required [type=missing, input_value={'tool': 'analyze_trend',...: 'day', 'agg': 'mean'}}, input_type=dict]\n For further information visit https://errors.pydantic.dev/2.10/v/missing\ntasks.2.stage\n Field required [type=missing, input_value={'tool': 'analyze_trend',...: 'day', 'agg': 'mean'}}, input_type=dict]\n For further information visit https://errors.pydantic.dev/2.10/v/missing\ntasks.2.objective\n Field required [type=missing, input_value={'tool': 'analyze_trend',...: 'day', 'agg': 'mean'}}, input_type=dict]\n For further information visit https://errors.pydantic.dev/2.10/v/missing\ntasks.2.tool_calls\n Field required [type=missing, input_value={'tool': 'analyze_trend',...: 'day', 'agg': 'mean'}}, input_type=dict]\n For further information visit https://errors.pydantic.dev/2.10/v/missing\ntasks.2.expected_output\n Field required [type=missing, input_value={'tool': 'analyze_trend',...: 'day', 'agg': 'mean'}}, input_type=dict]\n For further information visit https://errors.pydantic.dev/2.10/v/missing\ntasks.2.success_criteria\n Field required [type=missing, input_value={'tool': 'analyze_trend',...: 'day', 'agg': 'mean'}}, input_type=dict]\n For further information visit https://errors.pydantic.dev/2.10/v/missing",
489
+ "latency_ms": 3008,
490
+ "tokens": 14775
491
+ },
492
+ {
493
+ "id": "trend_downhour_en",
494
+ "category": "trend",
495
+ "lang": "en",
496
+ "carried_over": true,
497
+ "question": "show the trend of total down hours over time",
498
+ "passed": true,
499
+ "checks": [
500
+ {
501
+ "check": "has_tool=analyze_trend",
502
+ "ok": true,
503
+ "detail": "tools=['analyze_trend', 'check_data', 'retrieve_data']"
504
+ }
505
+ ],
506
+ "error": null,
507
+ "latency_ms": 3193,
508
+ "tokens": 14842
509
+ },
510
+ {
511
+ "id": "merge_worst_pa_and_mttr_id",
512
+ "category": "merge",
513
+ "lang": "id",
514
+ "carried_over": true,
515
+ "question": "model mana yang PA-nya paling buruk sekaligus MTTR-nya paling tinggi?",
516
+ "passed": true,
517
+ "checks": [
518
+ {
519
+ "check": "group_by",
520
+ "ok": true,
521
+ "detail": "group_by_present=True"
522
+ },
523
+ {
524
+ "check": "group_by_col=Model_Unit",
525
+ "ok": true,
526
+ "detail": "want_id=c_model_unit"
527
+ }
528
+ ],
529
+ "error": null,
530
+ "latency_ms": 4061,
531
+ "tokens": 15116
532
+ },
533
+ {
534
+ "id": "raw_rows_low_pa_id",
535
+ "category": "counter_raw_rows",
536
+ "lang": "id",
537
+ "carried_over": true,
538
+ "question": "tampilkan 10 record dengan PA di bawah 50",
539
+ "passed": true,
540
+ "checks": [
541
+ {
542
+ "check": "no_group_by",
543
+ "ok": true,
544
+ "detail": "group_by_present=False"
545
+ },
546
+ {
547
+ "check": "has_filter",
548
+ "ok": true,
549
+ "detail": "filter_present=True"
550
+ },
551
+ {
552
+ "check": "limit=10",
553
+ "ok": true,
554
+ "detail": "limits=[10]"
555
+ }
556
+ ],
557
+ "error": null,
558
+ "latency_ms": 2944,
559
+ "tokens": 14868
560
+ },
561
+ {
562
+ "id": "raw_rows_head_en",
563
+ "category": "counter_raw_rows",
564
+ "lang": "en",
565
+ "carried_over": true,
566
+ "question": "show me the first 5 rows of the data",
567
+ "passed": true,
568
+ "checks": [
569
+ {
570
+ "check": "no_group_by",
571
+ "ok": true,
572
+ "detail": "group_by_present=False"
573
+ }
574
+ ],
575
+ "error": null,
576
+ "latency_ms": 4427,
577
+ "tokens": 15315
578
+ },
579
+ {
580
+ "id": "exact_model_777d_id",
581
+ "category": "counter_exact_filter",
582
+ "lang": "id",
583
+ "carried_over": true,
584
+ "question": "berapa banyak record untuk model 777D?",
585
+ "passed": true,
586
+ "checks": [
587
+ {
588
+ "check": "select_agg=count",
589
+ "ok": true,
590
+ "detail": "aggs=['count']"
591
+ },
592
+ {
593
+ "check": "has_filter",
594
+ "ok": true,
595
+ "detail": "filter_present=True"
596
+ }
597
+ ],
598
+ "error": null,
599
+ "latency_ms": 2314,
600
+ "tokens": 14568
601
+ },
602
+ {
603
+ "id": "exact_section_loader_en",
604
+ "category": "counter_exact_filter",
605
+ "lang": "en",
606
+ "carried_over": true,
607
+ "question": "how many records are in the OB LOADER section?",
608
+ "passed": true,
609
+ "checks": [
610
+ {
611
+ "check": "select_agg=count",
612
+ "ok": true,
613
+ "detail": "aggs=['count']"
614
+ },
615
+ {
616
+ "check": "has_filter",
617
+ "ok": true,
618
+ "detail": "filter_present=True"
619
+ }
620
+ ],
621
+ "error": null,
622
+ "latency_ms": 2216,
623
+ "tokens": 14595
624
+ },
625
+ {
626
+ "id": "infeasible_churn_id",
627
+ "category": "infeasible",
628
+ "lang": "id",
629
+ "carried_over": true,
630
+ "question": "berapa churn rate pelanggan?",
631
+ "passed": true,
632
+ "checks": [
633
+ {
634
+ "check": "infeasible",
635
+ "ok": true,
636
+ "detail": "infeasible=True"
637
+ }
638
+ ],
639
+ "error": null,
640
+ "latency_ms": 1775,
641
+ "tokens": 14477
642
+ },
643
+ {
644
+ "id": "infeasible_profit_en",
645
+ "category": "infeasible",
646
+ "lang": "en",
647
+ "carried_over": true,
648
+ "question": "what is the monthly profit margin?",
649
+ "passed": true,
650
+ "checks": [
651
+ {
652
+ "check": "infeasible",
653
+ "ok": true,
654
+ "detail": "infeasible=True"
655
+ }
656
+ ],
657
+ "error": null,
658
+ "latency_ms": 2282,
659
+ "tokens": 14551
660
+ }
661
+ ]
662
+ }
eval/planner/results/planner_result_2026-07-23_141931.json ADDED
@@ -0,0 +1,1818 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "timestamp": "2026-07-23_141931",
3
+ "total": 27,
4
+ "passed": 25,
5
+ "cases": [
6
+ {
7
+ "id": "count_zero_pa",
8
+ "category": "count",
9
+ "lang": "en",
10
+ "carried_over": true,
11
+ "question": "how many records have PA_Percent = 0?",
12
+ "passed": true,
13
+ "checks": [
14
+ {
15
+ "check": "select_agg=count",
16
+ "ok": true,
17
+ "detail": "aggs=['count']"
18
+ },
19
+ {
20
+ "check": "has_filter",
21
+ "ok": true,
22
+ "detail": "filter_present=True"
23
+ },
24
+ {
25
+ "check": "no_group_by",
26
+ "ok": true,
27
+ "detail": "grouped=False"
28
+ }
29
+ ],
30
+ "facts": {
31
+ "tools": [
32
+ "retrieve_data"
33
+ ],
34
+ "irs": [
35
+ {
36
+ "source_id": "src_pa",
37
+ "table_id": "t_pa",
38
+ "select": [
39
+ {
40
+ "kind": "agg",
41
+ "fn": "count",
42
+ "alias": "record_count"
43
+ }
44
+ ],
45
+ "filters": [
46
+ {
47
+ "column_id": "c_pa_percent",
48
+ "op": "=",
49
+ "value": 0,
50
+ "value_type": "decimal"
51
+ }
52
+ ]
53
+ }
54
+ ],
55
+ "agg_args": [],
56
+ "infeasible": false
57
+ },
58
+ "error": null,
59
+ "latency_ms": 3652,
60
+ "tokens": 14570
61
+ },
62
+ {
63
+ "id": "count_mttr_gt20_id",
64
+ "category": "count",
65
+ "lang": "id",
66
+ "carried_over": true,
67
+ "question": "berapa banyak record dengan MTTR di atas 20?",
68
+ "passed": true,
69
+ "checks": [
70
+ {
71
+ "check": "select_agg=count",
72
+ "ok": true,
73
+ "detail": "aggs=['count']"
74
+ },
75
+ {
76
+ "check": "has_filter",
77
+ "ok": true,
78
+ "detail": "filter_present=True"
79
+ }
80
+ ],
81
+ "facts": {
82
+ "tools": [
83
+ "retrieve_data"
84
+ ],
85
+ "irs": [
86
+ {
87
+ "source_id": "src_pa",
88
+ "table_id": "t_pa",
89
+ "select": [
90
+ {
91
+ "kind": "agg",
92
+ "fn": "count",
93
+ "alias": "record_count"
94
+ }
95
+ ],
96
+ "filters": [
97
+ {
98
+ "column_id": "c_mttr",
99
+ "op": ">",
100
+ "value": 20,
101
+ "value_type": "decimal"
102
+ }
103
+ ]
104
+ }
105
+ ],
106
+ "agg_args": [],
107
+ "infeasible": false
108
+ },
109
+ "error": null,
110
+ "latency_ms": 2176,
111
+ "tokens": 14571
112
+ },
113
+ {
114
+ "id": "count_section_hauler",
115
+ "category": "count",
116
+ "lang": "en",
117
+ "carried_over": true,
118
+ "question": "how many rows are in section OB HAULER?",
119
+ "passed": true,
120
+ "checks": [
121
+ {
122
+ "check": "select_agg=count",
123
+ "ok": true,
124
+ "detail": "aggs=['count']"
125
+ },
126
+ {
127
+ "check": "has_filter",
128
+ "ok": true,
129
+ "detail": "filter_present=True"
130
+ }
131
+ ],
132
+ "facts": {
133
+ "tools": [
134
+ "retrieve_data"
135
+ ],
136
+ "irs": [
137
+ {
138
+ "source_id": "src_pa",
139
+ "table_id": "t_pa",
140
+ "select": [
141
+ {
142
+ "kind": "agg",
143
+ "fn": "count",
144
+ "alias": "row_count"
145
+ }
146
+ ],
147
+ "filters": [
148
+ {
149
+ "column_id": "c_section",
150
+ "op": "=",
151
+ "value": "OB HAULER",
152
+ "value_type": "string"
153
+ }
154
+ ]
155
+ }
156
+ ],
157
+ "agg_args": [],
158
+ "infeasible": false
159
+ },
160
+ "error": null,
161
+ "latency_ms": 2342,
162
+ "tokens": 14571
163
+ },
164
+ {
165
+ "id": "rank_units_worst_pa_id",
166
+ "category": "ranking",
167
+ "lang": "id",
168
+ "carried_over": false,
169
+ "question": "5 unit dengan PA terburuk?",
170
+ "passed": true,
171
+ "checks": [
172
+ {
173
+ "check": "group_by",
174
+ "ok": true,
175
+ "detail": "grouped=True"
176
+ },
177
+ {
178
+ "check": "group_by_col=Equipment_Number",
179
+ "ok": true,
180
+ "detail": "ids=['c_equipment_number'] aliases=[]"
181
+ },
182
+ {
183
+ "check": "select_agg=avg",
184
+ "ok": true,
185
+ "detail": "aggs=['avg']"
186
+ },
187
+ {
188
+ "check": "order_dir=asc",
189
+ "ok": true,
190
+ "detail": "dirs=['asc']"
191
+ },
192
+ {
193
+ "check": "limit=5",
194
+ "ok": true,
195
+ "detail": "limits=[5]"
196
+ }
197
+ ],
198
+ "facts": {
199
+ "tools": [
200
+ "check_data",
201
+ "retrieve_data"
202
+ ],
203
+ "irs": [
204
+ {
205
+ "source_id": "src_pa",
206
+ "table_id": "t_pa",
207
+ "select": [
208
+ {
209
+ "kind": "column",
210
+ "column_id": "c_equipment_number",
211
+ "alias": "equipment_number"
212
+ },
213
+ {
214
+ "kind": "agg",
215
+ "fn": "avg",
216
+ "column_id": "c_pa_percent",
217
+ "alias": "avg_pa_percent"
218
+ }
219
+ ],
220
+ "group_by": [
221
+ "c_equipment_number"
222
+ ],
223
+ "order_by": [
224
+ {
225
+ "column_id": "avg_pa_percent",
226
+ "dir": "asc"
227
+ }
228
+ ],
229
+ "limit": 5
230
+ }
231
+ ],
232
+ "agg_args": [],
233
+ "infeasible": false
234
+ },
235
+ "error": null,
236
+ "latency_ms": 3519,
237
+ "tokens": 14733
238
+ },
239
+ {
240
+ "id": "rank_models_top_mttr_id",
241
+ "category": "ranking",
242
+ "lang": "id",
243
+ "carried_over": true,
244
+ "question": "top 3 model dengan MTTR tertinggi?",
245
+ "passed": true,
246
+ "checks": [
247
+ {
248
+ "check": "group_by",
249
+ "ok": true,
250
+ "detail": "grouped=True"
251
+ },
252
+ {
253
+ "check": "group_by_col=Model_Unit",
254
+ "ok": true,
255
+ "detail": "ids=['c_model_unit'] aliases=[]"
256
+ },
257
+ {
258
+ "check": "select_agg=avg",
259
+ "ok": true,
260
+ "detail": "aggs=['avg']"
261
+ },
262
+ {
263
+ "check": "order_dir=desc",
264
+ "ok": true,
265
+ "detail": "dirs=['desc']"
266
+ },
267
+ {
268
+ "check": "limit=3",
269
+ "ok": true,
270
+ "detail": "limits=[3]"
271
+ }
272
+ ],
273
+ "facts": {
274
+ "tools": [
275
+ "check_data",
276
+ "retrieve_data"
277
+ ],
278
+ "irs": [
279
+ {
280
+ "source_id": "src_pa",
281
+ "table_id": "t_pa",
282
+ "select": [
283
+ {
284
+ "kind": "column",
285
+ "column_id": "c_model_unit",
286
+ "alias": "model"
287
+ },
288
+ {
289
+ "kind": "agg",
290
+ "fn": "avg",
291
+ "column_id": "c_mttr",
292
+ "alias": "avg_mttr"
293
+ }
294
+ ],
295
+ "group_by": [
296
+ "c_model_unit"
297
+ ],
298
+ "order_by": [
299
+ {
300
+ "column_id": "avg_mttr",
301
+ "dir": "desc"
302
+ }
303
+ ],
304
+ "limit": 3
305
+ }
306
+ ],
307
+ "agg_args": [],
308
+ "infeasible": false
309
+ },
310
+ "error": null,
311
+ "latency_ms": 2928,
312
+ "tokens": 14723
313
+ },
314
+ {
315
+ "id": "rank_sections_lowest_pa_en",
316
+ "category": "ranking",
317
+ "lang": "en",
318
+ "carried_over": true,
319
+ "question": "which section has the lowest average PA?",
320
+ "passed": false,
321
+ "checks": [
322
+ {
323
+ "check": "group_by",
324
+ "ok": true,
325
+ "detail": "grouped=True"
326
+ },
327
+ {
328
+ "check": "group_by_col=Section",
329
+ "ok": true,
330
+ "detail": "ids=[] aliases=['section']"
331
+ },
332
+ {
333
+ "check": "select_agg=avg",
334
+ "ok": true,
335
+ "detail": "aggs=['mean']"
336
+ },
337
+ {
338
+ "check": "order_dir=asc",
339
+ "ok": false,
340
+ "detail": "dirs=[]"
341
+ }
342
+ ],
343
+ "facts": {
344
+ "tools": [
345
+ "analyze_aggregate",
346
+ "check_data",
347
+ "retrieve_data"
348
+ ],
349
+ "irs": [
350
+ {
351
+ "source_id": "src_pa",
352
+ "table_id": "t_pa",
353
+ "select": [
354
+ {
355
+ "kind": "column",
356
+ "column_id": "c_section",
357
+ "alias": "section"
358
+ },
359
+ {
360
+ "kind": "column",
361
+ "column_id": "c_pa_percent",
362
+ "alias": "pa_percent"
363
+ }
364
+ ],
365
+ "filters": [
366
+ {
367
+ "column_id": "c_month_id",
368
+ "op": "=",
369
+ "value": 202604,
370
+ "value_type": "int"
371
+ },
372
+ {
373
+ "column_id": "c_isdeleted",
374
+ "op": "=",
375
+ "value": 0,
376
+ "value_type": "int"
377
+ }
378
+ ],
379
+ "limit": 10000
380
+ }
381
+ ],
382
+ "agg_args": [
383
+ {
384
+ "data": "${t2}",
385
+ "aggregations": {
386
+ "pa_percent": [
387
+ "mean"
388
+ ]
389
+ },
390
+ "group_by": [
391
+ "section"
392
+ ]
393
+ }
394
+ ],
395
+ "infeasible": false
396
+ },
397
+ "error": null,
398
+ "latency_ms": 3445,
399
+ "tokens": 14833
400
+ },
401
+ {
402
+ "id": "rank_units_most_breakdown_id",
403
+ "category": "ranking",
404
+ "lang": "id",
405
+ "carried_over": false,
406
+ "question": "unit mana yang paling sering breakdown?",
407
+ "passed": true,
408
+ "checks": [
409
+ {
410
+ "check": "group_by",
411
+ "ok": true,
412
+ "detail": "grouped=True"
413
+ },
414
+ {
415
+ "check": "group_by_col=Equipment_Number",
416
+ "ok": true,
417
+ "detail": "ids=['c_equipment_number'] aliases=[]"
418
+ },
419
+ {
420
+ "check": "order_dir=desc",
421
+ "ok": true,
422
+ "detail": "dirs=['desc']"
423
+ }
424
+ ],
425
+ "facts": {
426
+ "tools": [
427
+ "check_data",
428
+ "retrieve_data"
429
+ ],
430
+ "irs": [
431
+ {
432
+ "source_id": "src_pa",
433
+ "table_id": "t_pa",
434
+ "select": [
435
+ {
436
+ "kind": "column",
437
+ "column_id": "c_equipment_number",
438
+ "alias": "equipment_number"
439
+ },
440
+ {
441
+ "kind": "agg",
442
+ "fn": "sum",
443
+ "column_id": "c_total_frequency_down",
444
+ "alias": "total_breakdown_frequency"
445
+ }
446
+ ],
447
+ "filters": [
448
+ {
449
+ "column_id": "c_month_id",
450
+ "op": "=",
451
+ "value": 202604,
452
+ "value_type": "int"
453
+ },
454
+ {
455
+ "column_id": "c_isdeleted",
456
+ "op": "=",
457
+ "value": 0,
458
+ "value_type": "int"
459
+ }
460
+ ],
461
+ "group_by": [
462
+ "c_equipment_number"
463
+ ],
464
+ "order_by": [
465
+ {
466
+ "column_id": "total_breakdown_frequency",
467
+ "dir": "desc"
468
+ }
469
+ ],
470
+ "limit": 1
471
+ }
472
+ ],
473
+ "agg_args": [],
474
+ "infeasible": false
475
+ },
476
+ "error": null,
477
+ "latency_ms": 2590,
478
+ "tokens": 14760
479
+ },
480
+ {
481
+ "id": "rank_units_worst_pa_en",
482
+ "category": "ranking",
483
+ "lang": "en",
484
+ "carried_over": false,
485
+ "question": "list the 10 worst units by availability",
486
+ "passed": true,
487
+ "checks": [
488
+ {
489
+ "check": "group_by",
490
+ "ok": true,
491
+ "detail": "grouped=True"
492
+ },
493
+ {
494
+ "check": "group_by_col=Equipment_Number",
495
+ "ok": true,
496
+ "detail": "ids=['c_equipment_number'] aliases=[]"
497
+ },
498
+ {
499
+ "check": "select_agg=avg",
500
+ "ok": true,
501
+ "detail": "aggs=['avg']"
502
+ },
503
+ {
504
+ "check": "order_dir=asc",
505
+ "ok": true,
506
+ "detail": "dirs=['asc']"
507
+ },
508
+ {
509
+ "check": "limit=10",
510
+ "ok": true,
511
+ "detail": "limits=[10]"
512
+ }
513
+ ],
514
+ "facts": {
515
+ "tools": [
516
+ "check_data",
517
+ "retrieve_data"
518
+ ],
519
+ "irs": [
520
+ {
521
+ "source_id": "src_pa",
522
+ "table_id": "t_pa",
523
+ "select": [
524
+ {
525
+ "kind": "column",
526
+ "column_id": "c_equipment_number",
527
+ "alias": "equipment_number"
528
+ },
529
+ {
530
+ "kind": "agg",
531
+ "fn": "avg",
532
+ "column_id": "c_pa_percent",
533
+ "alias": "avg_pa_percent"
534
+ }
535
+ ],
536
+ "filters": [
537
+ {
538
+ "column_id": "c_month_id",
539
+ "op": "=",
540
+ "value": 202604,
541
+ "value_type": "int"
542
+ },
543
+ {
544
+ "column_id": "c_isdeleted",
545
+ "op": "=",
546
+ "value": 0,
547
+ "value_type": "int"
548
+ }
549
+ ],
550
+ "group_by": [
551
+ "c_equipment_number"
552
+ ],
553
+ "order_by": [
554
+ {
555
+ "column_id": "avg_pa_percent",
556
+ "dir": "asc"
557
+ }
558
+ ],
559
+ "limit": 10
560
+ }
561
+ ],
562
+ "agg_args": [],
563
+ "infeasible": false
564
+ },
565
+ "error": null,
566
+ "latency_ms": 3221,
567
+ "tokens": 14791
568
+ },
569
+ {
570
+ "id": "fuzzy_model_777_id",
571
+ "category": "fuzzy_filter",
572
+ "lang": "id",
573
+ "carried_over": false,
574
+ "question": "berapa banyak model 777?",
575
+ "passed": true,
576
+ "checks": [
577
+ {
578
+ "check": "select_agg=count",
579
+ "ok": true,
580
+ "detail": "aggs=['count']"
581
+ },
582
+ {
583
+ "check": "no_filter_op=in",
584
+ "ok": true,
585
+ "detail": "ops=['=']"
586
+ }
587
+ ],
588
+ "facts": {
589
+ "tools": [
590
+ "retrieve_data"
591
+ ],
592
+ "irs": [
593
+ {
594
+ "source_id": "src_pa",
595
+ "table_id": "t_pa",
596
+ "select": [
597
+ {
598
+ "kind": "agg",
599
+ "fn": "count",
600
+ "alias": "record_count"
601
+ }
602
+ ],
603
+ "filters": [
604
+ {
605
+ "column_id": "c_model_unit",
606
+ "op": "=",
607
+ "value": "777",
608
+ "value_type": "string"
609
+ }
610
+ ]
611
+ }
612
+ ],
613
+ "agg_args": [],
614
+ "infeasible": false
615
+ },
616
+ "error": null,
617
+ "latency_ms": 2332,
618
+ "tokens": 14576
619
+ },
620
+ {
621
+ "id": "fuzzy_model_hd785_id",
622
+ "category": "fuzzy_filter",
623
+ "lang": "id",
624
+ "carried_over": true,
625
+ "question": "berapa banyak unit HD785?",
626
+ "passed": true,
627
+ "checks": [
628
+ {
629
+ "check": "select_agg=count",
630
+ "ok": true,
631
+ "detail": "aggs=['count']"
632
+ },
633
+ {
634
+ "check": "no_filter_op=in",
635
+ "ok": true,
636
+ "detail": "ops=['=']"
637
+ }
638
+ ],
639
+ "facts": {
640
+ "tools": [
641
+ "retrieve_data"
642
+ ],
643
+ "irs": [
644
+ {
645
+ "source_id": "src_pa",
646
+ "table_id": "t_pa",
647
+ "select": [
648
+ {
649
+ "kind": "agg",
650
+ "fn": "count",
651
+ "alias": "unit_count"
652
+ }
653
+ ],
654
+ "filters": [
655
+ {
656
+ "column_id": "c_model_unit",
657
+ "op": "=",
658
+ "value": "HD785",
659
+ "value_type": "string"
660
+ }
661
+ ],
662
+ "limit": 10000
663
+ }
664
+ ],
665
+ "agg_args": [],
666
+ "infeasible": false
667
+ },
668
+ "error": null,
669
+ "latency_ms": 4613,
670
+ "tokens": 29329
671
+ },
672
+ {
673
+ "id": "fuzzy_model_ex_en",
674
+ "category": "fuzzy_filter",
675
+ "lang": "en",
676
+ "carried_over": false,
677
+ "question": "how many EX excavator units are there?",
678
+ "passed": true,
679
+ "checks": [
680
+ {
681
+ "check": "no_filter_op=in",
682
+ "ok": true,
683
+ "detail": "ops=['like']"
684
+ }
685
+ ],
686
+ "facts": {
687
+ "tools": [
688
+ "retrieve_data"
689
+ ],
690
+ "irs": [
691
+ {
692
+ "source_id": "src_pa",
693
+ "table_id": "t_pa",
694
+ "select": [
695
+ {
696
+ "kind": "agg",
697
+ "fn": "count",
698
+ "alias": "unit_count"
699
+ }
700
+ ],
701
+ "filters": [
702
+ {
703
+ "column_id": "c_equipment_number",
704
+ "op": "like",
705
+ "value": "EX%",
706
+ "value_type": "string"
707
+ }
708
+ ]
709
+ }
710
+ ],
711
+ "agg_args": [],
712
+ "infeasible": false
713
+ },
714
+ "error": null,
715
+ "latency_ms": 5577,
716
+ "tokens": 29385
717
+ },
718
+ {
719
+ "id": "agg_pa_per_section_id",
720
+ "category": "aggregate",
721
+ "lang": "id",
722
+ "carried_over": true,
723
+ "question": "berapa rata-rata PA per section?",
724
+ "passed": true,
725
+ "checks": [
726
+ {
727
+ "check": "group_by",
728
+ "ok": true,
729
+ "detail": "grouped=True"
730
+ },
731
+ {
732
+ "check": "group_by_col=Section",
733
+ "ok": true,
734
+ "detail": "ids=[] aliases=['section']"
735
+ },
736
+ {
737
+ "check": "select_agg=avg",
738
+ "ok": true,
739
+ "detail": "aggs=['mean']"
740
+ }
741
+ ],
742
+ "facts": {
743
+ "tools": [
744
+ "analyze_aggregate",
745
+ "check_data",
746
+ "retrieve_data"
747
+ ],
748
+ "irs": [
749
+ {
750
+ "source_id": "src_pa",
751
+ "table_id": "t_pa",
752
+ "select": [
753
+ {
754
+ "kind": "column",
755
+ "column_id": "c_section",
756
+ "alias": "section"
757
+ },
758
+ {
759
+ "kind": "column",
760
+ "column_id": "c_pa_percent",
761
+ "alias": "pa_percent"
762
+ }
763
+ ],
764
+ "limit": 10000
765
+ }
766
+ ],
767
+ "agg_args": [
768
+ {
769
+ "data": "${t2}",
770
+ "aggregations": {
771
+ "pa_percent": [
772
+ "mean"
773
+ ]
774
+ },
775
+ "group_by": [
776
+ "section"
777
+ ]
778
+ }
779
+ ],
780
+ "infeasible": false
781
+ },
782
+ "error": null,
783
+ "latency_ms": 3785,
784
+ "tokens": 14754
785
+ },
786
+ {
787
+ "id": "agg_mttr_per_model_en",
788
+ "category": "aggregate",
789
+ "lang": "en",
790
+ "carried_over": true,
791
+ "question": "what is the average MTTR per model unit?",
792
+ "passed": true,
793
+ "checks": [
794
+ {
795
+ "check": "group_by",
796
+ "ok": true,
797
+ "detail": "grouped=True"
798
+ },
799
+ {
800
+ "check": "group_by_col=Model_Unit",
801
+ "ok": true,
802
+ "detail": "ids=[] aliases=['model_unit']"
803
+ },
804
+ {
805
+ "check": "select_agg=avg",
806
+ "ok": true,
807
+ "detail": "aggs=['mean']"
808
+ }
809
+ ],
810
+ "facts": {
811
+ "tools": [
812
+ "analyze_aggregate",
813
+ "check_data",
814
+ "retrieve_data"
815
+ ],
816
+ "irs": [
817
+ {
818
+ "source_id": "src_pa",
819
+ "table_id": "t_pa",
820
+ "select": [
821
+ {
822
+ "kind": "column",
823
+ "column_id": "c_model_unit",
824
+ "alias": "model_unit"
825
+ },
826
+ {
827
+ "kind": "column",
828
+ "column_id": "c_mttr",
829
+ "alias": "mttr"
830
+ }
831
+ ],
832
+ "limit": 10000
833
+ }
834
+ ],
835
+ "agg_args": [
836
+ {
837
+ "data": "${t2}",
838
+ "aggregations": {
839
+ "mttr": [
840
+ "mean"
841
+ ]
842
+ },
843
+ "group_by": [
844
+ "model_unit"
845
+ ]
846
+ }
847
+ ],
848
+ "infeasible": false
849
+ },
850
+ "error": null,
851
+ "latency_ms": 2925,
852
+ "tokens": 14790
853
+ },
854
+ {
855
+ "id": "agg_downhour_per_group_id",
856
+ "category": "aggregate",
857
+ "lang": "id",
858
+ "carried_over": true,
859
+ "question": "total down hour per equipment group?",
860
+ "passed": false,
861
+ "checks": [
862
+ {
863
+ "check": "group_by",
864
+ "ok": true,
865
+ "detail": "grouped=True"
866
+ },
867
+ {
868
+ "check": "group_by_col=Equipment_Group_ID",
869
+ "ok": false,
870
+ "detail": "ids=[] aliases=['equipment_group']"
871
+ },
872
+ {
873
+ "check": "select_agg=sum",
874
+ "ok": true,
875
+ "detail": "aggs=['sum']"
876
+ }
877
+ ],
878
+ "facts": {
879
+ "tools": [
880
+ "analyze_aggregate",
881
+ "check_data",
882
+ "retrieve_data"
883
+ ],
884
+ "irs": [
885
+ {
886
+ "source_id": "src_pa",
887
+ "table_id": "t_pa",
888
+ "select": [
889
+ {
890
+ "kind": "column",
891
+ "column_id": "c_equipment_group_id",
892
+ "alias": "equipment_group"
893
+ },
894
+ {
895
+ "kind": "column",
896
+ "column_id": "c_total_down_hour",
897
+ "alias": "total_down_hour"
898
+ }
899
+ ],
900
+ "limit": 10000
901
+ }
902
+ ],
903
+ "agg_args": [
904
+ {
905
+ "data": "${t2}",
906
+ "aggregations": {
907
+ "total_down_hour": [
908
+ "sum"
909
+ ]
910
+ },
911
+ "group_by": [
912
+ "equipment_group"
913
+ ]
914
+ }
915
+ ],
916
+ "infeasible": false
917
+ },
918
+ "error": null,
919
+ "latency_ms": 3449,
920
+ "tokens": 14785
921
+ },
922
+ {
923
+ "id": "desc_mttr_stats_id",
924
+ "category": "descriptive",
925
+ "lang": "id",
926
+ "carried_over": true,
927
+ "question": "berikan ringkasan statistik MTTR",
928
+ "passed": true,
929
+ "checks": [
930
+ {
931
+ "check": "has_tool=analyze_descriptive",
932
+ "ok": true,
933
+ "detail": "tools=['analyze_descriptive', 'check_data', 'retrieve_data']"
934
+ }
935
+ ],
936
+ "facts": {
937
+ "tools": [
938
+ "analyze_descriptive",
939
+ "check_data",
940
+ "retrieve_data"
941
+ ],
942
+ "irs": [
943
+ {
944
+ "source_id": "src_pa",
945
+ "table_id": "t_pa",
946
+ "select": [
947
+ {
948
+ "kind": "column",
949
+ "column_id": "c_mttr",
950
+ "alias": "mttr"
951
+ }
952
+ ],
953
+ "filters": [
954
+ {
955
+ "column_id": "c_month_id",
956
+ "op": "=",
957
+ "value": 202604,
958
+ "value_type": "int"
959
+ },
960
+ {
961
+ "column_id": "c_isdeleted",
962
+ "op": "=",
963
+ "value": 0,
964
+ "value_type": "int"
965
+ }
966
+ ],
967
+ "limit": 10000
968
+ }
969
+ ],
970
+ "agg_args": [],
971
+ "infeasible": false
972
+ },
973
+ "error": null,
974
+ "latency_ms": 2895,
975
+ "tokens": 14767
976
+ },
977
+ {
978
+ "id": "desc_pa_stats_en",
979
+ "category": "descriptive",
980
+ "lang": "en",
981
+ "carried_over": true,
982
+ "question": "give me the summary statistics for PA_Percent",
983
+ "passed": true,
984
+ "checks": [
985
+ {
986
+ "check": "has_tool=analyze_descriptive",
987
+ "ok": true,
988
+ "detail": "tools=['analyze_descriptive', 'check_data', 'retrieve_data']"
989
+ }
990
+ ],
991
+ "facts": {
992
+ "tools": [
993
+ "analyze_descriptive",
994
+ "check_data",
995
+ "retrieve_data"
996
+ ],
997
+ "irs": [
998
+ {
999
+ "source_id": "src_pa",
1000
+ "table_id": "t_pa",
1001
+ "select": [
1002
+ {
1003
+ "kind": "column",
1004
+ "column_id": "c_pa_percent",
1005
+ "alias": "pa_percent"
1006
+ }
1007
+ ],
1008
+ "limit": 10000
1009
+ }
1010
+ ],
1011
+ "agg_args": [],
1012
+ "infeasible": false
1013
+ },
1014
+ "error": null,
1015
+ "latency_ms": 2850,
1016
+ "tokens": 14703
1017
+ },
1018
+ {
1019
+ "id": "corr_mttr_pa_id",
1020
+ "category": "correlation",
1021
+ "lang": "id",
1022
+ "carried_over": true,
1023
+ "question": "apakah ada korelasi antara MTTR dan PA?",
1024
+ "passed": true,
1025
+ "checks": [
1026
+ {
1027
+ "check": "has_tool=analyze_correlation",
1028
+ "ok": true,
1029
+ "detail": "tools=['analyze_correlation', 'check_data', 'retrieve_data']"
1030
+ }
1031
+ ],
1032
+ "facts": {
1033
+ "tools": [
1034
+ "analyze_correlation",
1035
+ "check_data",
1036
+ "retrieve_data"
1037
+ ],
1038
+ "irs": [
1039
+ {
1040
+ "source_id": "src_pa",
1041
+ "table_id": "t_pa",
1042
+ "select": [
1043
+ {
1044
+ "kind": "column",
1045
+ "column_id": "c_mttr",
1046
+ "alias": "mttr"
1047
+ },
1048
+ {
1049
+ "kind": "column",
1050
+ "column_id": "c_pa_percent",
1051
+ "alias": "pa_percent"
1052
+ }
1053
+ ],
1054
+ "filters": [
1055
+ {
1056
+ "column_id": "c_month_id",
1057
+ "op": "=",
1058
+ "value": 202604,
1059
+ "value_type": "int"
1060
+ },
1061
+ {
1062
+ "column_id": "c_isdeleted",
1063
+ "op": "=",
1064
+ "value": 0,
1065
+ "value_type": "int"
1066
+ }
1067
+ ],
1068
+ "limit": 10000
1069
+ }
1070
+ ],
1071
+ "agg_args": [],
1072
+ "infeasible": false
1073
+ },
1074
+ "error": null,
1075
+ "latency_ms": 3665,
1076
+ "tokens": 14850
1077
+ },
1078
+ {
1079
+ "id": "corr_freq_pa_en",
1080
+ "category": "correlation",
1081
+ "lang": "en",
1082
+ "carried_over": true,
1083
+ "question": "is breakdown frequency correlated with availability?",
1084
+ "passed": true,
1085
+ "checks": [
1086
+ {
1087
+ "check": "has_tool=analyze_correlation",
1088
+ "ok": true,
1089
+ "detail": "tools=['analyze_correlation', 'check_data', 'retrieve_data']"
1090
+ }
1091
+ ],
1092
+ "facts": {
1093
+ "tools": [
1094
+ "analyze_correlation",
1095
+ "check_data",
1096
+ "retrieve_data"
1097
+ ],
1098
+ "irs": [
1099
+ {
1100
+ "source_id": "src_pa",
1101
+ "table_id": "t_pa",
1102
+ "select": [
1103
+ {
1104
+ "kind": "column",
1105
+ "column_id": "c_total_frequency_breakdown_schedule",
1106
+ "alias": "breakdown_freq_schedule"
1107
+ },
1108
+ {
1109
+ "kind": "column",
1110
+ "column_id": "c_total_frequency_breakdown_unschedule",
1111
+ "alias": "breakdown_freq_unschedule"
1112
+ },
1113
+ {
1114
+ "kind": "column",
1115
+ "column_id": "c_pa_percent",
1116
+ "alias": "pa_percent"
1117
+ }
1118
+ ],
1119
+ "filters": [
1120
+ {
1121
+ "column_id": "c_isdeleted",
1122
+ "op": "=",
1123
+ "value": 0,
1124
+ "value_type": "int"
1125
+ }
1126
+ ],
1127
+ "limit": 10000
1128
+ }
1129
+ ],
1130
+ "agg_args": [],
1131
+ "infeasible": false
1132
+ },
1133
+ "error": null,
1134
+ "latency_ms": 3299,
1135
+ "tokens": 14859
1136
+ },
1137
+ {
1138
+ "id": "trend_pa_daily_id",
1139
+ "category": "trend",
1140
+ "lang": "id",
1141
+ "carried_over": true,
1142
+ "question": "bagaimana trend PA harian?",
1143
+ "passed": true,
1144
+ "checks": [
1145
+ {
1146
+ "check": "has_tool=analyze_trend",
1147
+ "ok": true,
1148
+ "detail": "tools=['analyze_trend', 'check_data', 'retrieve_data']"
1149
+ }
1150
+ ],
1151
+ "facts": {
1152
+ "tools": [
1153
+ "analyze_trend",
1154
+ "check_data",
1155
+ "retrieve_data"
1156
+ ],
1157
+ "irs": [
1158
+ {
1159
+ "source_id": "src_pa",
1160
+ "table_id": "t_pa",
1161
+ "select": [
1162
+ {
1163
+ "kind": "column",
1164
+ "column_id": "c_from_date",
1165
+ "alias": "date"
1166
+ },
1167
+ {
1168
+ "kind": "column",
1169
+ "column_id": "c_pa_percent",
1170
+ "alias": "pa_percent"
1171
+ }
1172
+ ],
1173
+ "filters": [
1174
+ {
1175
+ "column_id": "c_month_id",
1176
+ "op": "=",
1177
+ "value": 202604,
1178
+ "value_type": "int"
1179
+ },
1180
+ {
1181
+ "column_id": "c_isdeleted",
1182
+ "op": "=",
1183
+ "value": 0,
1184
+ "value_type": "int"
1185
+ }
1186
+ ],
1187
+ "limit": 10000
1188
+ }
1189
+ ],
1190
+ "agg_args": [],
1191
+ "infeasible": false
1192
+ },
1193
+ "error": null,
1194
+ "latency_ms": 3472,
1195
+ "tokens": 14861
1196
+ },
1197
+ {
1198
+ "id": "trend_downhour_en",
1199
+ "category": "trend",
1200
+ "lang": "en",
1201
+ "carried_over": true,
1202
+ "question": "show the trend of total down hours over time",
1203
+ "passed": true,
1204
+ "checks": [
1205
+ {
1206
+ "check": "has_tool=analyze_trend",
1207
+ "ok": true,
1208
+ "detail": "tools=['analyze_trend', 'check_data', 'retrieve_data']"
1209
+ }
1210
+ ],
1211
+ "facts": {
1212
+ "tools": [
1213
+ "analyze_trend",
1214
+ "check_data",
1215
+ "retrieve_data"
1216
+ ],
1217
+ "irs": [
1218
+ {
1219
+ "source_id": "src_pa",
1220
+ "table_id": "t_pa",
1221
+ "select": [
1222
+ {
1223
+ "kind": "column",
1224
+ "column_id": "c_from_date",
1225
+ "alias": "date"
1226
+ },
1227
+ {
1228
+ "kind": "column",
1229
+ "column_id": "c_total_down_hour",
1230
+ "alias": "total_down_hours"
1231
+ }
1232
+ ],
1233
+ "filters": [
1234
+ {
1235
+ "column_id": "c_month_id",
1236
+ "op": "=",
1237
+ "value": 202604,
1238
+ "value_type": "int"
1239
+ }
1240
+ ],
1241
+ "limit": 10000
1242
+ }
1243
+ ],
1244
+ "agg_args": [],
1245
+ "infeasible": false
1246
+ },
1247
+ "error": null,
1248
+ "latency_ms": 3703,
1249
+ "tokens": 14825
1250
+ },
1251
+ {
1252
+ "id": "merge_worst_pa_and_mttr_id",
1253
+ "category": "merge",
1254
+ "lang": "id",
1255
+ "carried_over": true,
1256
+ "question": "model mana yang PA-nya paling buruk sekaligus MTTR-nya paling tinggi?",
1257
+ "passed": true,
1258
+ "checks": [
1259
+ {
1260
+ "check": "group_by",
1261
+ "ok": true,
1262
+ "detail": "grouped=True"
1263
+ },
1264
+ {
1265
+ "check": "group_by_col=Model_Unit",
1266
+ "ok": true,
1267
+ "detail": "ids=['c_model_unit', 'c_model_unit'] aliases=[]"
1268
+ }
1269
+ ],
1270
+ "facts": {
1271
+ "tools": [
1272
+ "analyze_merge",
1273
+ "check_data",
1274
+ "retrieve_data"
1275
+ ],
1276
+ "irs": [
1277
+ {
1278
+ "source_id": "src_pa",
1279
+ "table_id": "t_pa",
1280
+ "select": [
1281
+ {
1282
+ "kind": "column",
1283
+ "column_id": "c_model_unit",
1284
+ "alias": "model"
1285
+ },
1286
+ {
1287
+ "kind": "agg",
1288
+ "fn": "avg",
1289
+ "column_id": "c_pa_percent",
1290
+ "alias": "avg_pa"
1291
+ }
1292
+ ],
1293
+ "group_by": [
1294
+ "c_model_unit"
1295
+ ],
1296
+ "order_by": [
1297
+ {
1298
+ "column_id": "avg_pa",
1299
+ "dir": "asc"
1300
+ }
1301
+ ],
1302
+ "limit": 1
1303
+ },
1304
+ {
1305
+ "source_id": "src_pa",
1306
+ "table_id": "t_pa",
1307
+ "select": [
1308
+ {
1309
+ "kind": "column",
1310
+ "column_id": "c_model_unit",
1311
+ "alias": "model"
1312
+ },
1313
+ {
1314
+ "kind": "agg",
1315
+ "fn": "avg",
1316
+ "column_id": "c_mttr",
1317
+ "alias": "avg_mttr"
1318
+ }
1319
+ ],
1320
+ "group_by": [
1321
+ "c_model_unit"
1322
+ ],
1323
+ "order_by": [
1324
+ {
1325
+ "column_id": "avg_mttr",
1326
+ "dir": "desc"
1327
+ }
1328
+ ],
1329
+ "limit": 1
1330
+ }
1331
+ ],
1332
+ "agg_args": [],
1333
+ "infeasible": false
1334
+ },
1335
+ "error": null,
1336
+ "latency_ms": 4058,
1337
+ "tokens": 15042
1338
+ },
1339
+ {
1340
+ "id": "raw_rows_low_pa_id",
1341
+ "category": "counter_raw_rows",
1342
+ "lang": "id",
1343
+ "carried_over": true,
1344
+ "question": "tampilkan 10 record dengan PA di bawah 50",
1345
+ "passed": true,
1346
+ "checks": [
1347
+ {
1348
+ "check": "no_group_by",
1349
+ "ok": true,
1350
+ "detail": "grouped=False"
1351
+ },
1352
+ {
1353
+ "check": "has_filter",
1354
+ "ok": true,
1355
+ "detail": "filter_present=True"
1356
+ },
1357
+ {
1358
+ "check": "limit=10",
1359
+ "ok": true,
1360
+ "detail": "limits=[10]"
1361
+ }
1362
+ ],
1363
+ "facts": {
1364
+ "tools": [
1365
+ "retrieve_data"
1366
+ ],
1367
+ "irs": [
1368
+ {
1369
+ "source_id": "src_pa",
1370
+ "table_id": "t_pa",
1371
+ "select": [
1372
+ {
1373
+ "kind": "column",
1374
+ "column_id": "c_keyid",
1375
+ "alias": "key_id"
1376
+ },
1377
+ {
1378
+ "kind": "column",
1379
+ "column_id": "c_month_id",
1380
+ "alias": "month_id"
1381
+ },
1382
+ {
1383
+ "kind": "column",
1384
+ "column_id": "c_site_id",
1385
+ "alias": "site_id"
1386
+ },
1387
+ {
1388
+ "kind": "column",
1389
+ "column_id": "c_from_date",
1390
+ "alias": "from_date"
1391
+ },
1392
+ {
1393
+ "kind": "column",
1394
+ "column_id": "c_to_date",
1395
+ "alias": "to_date"
1396
+ },
1397
+ {
1398
+ "kind": "column",
1399
+ "column_id": "c_time_description",
1400
+ "alias": "time_description"
1401
+ },
1402
+ {
1403
+ "kind": "column",
1404
+ "column_id": "c_model_unit",
1405
+ "alias": "model_unit"
1406
+ },
1407
+ {
1408
+ "kind": "column",
1409
+ "column_id": "c_equipment_number",
1410
+ "alias": "equipment_number"
1411
+ },
1412
+ {
1413
+ "kind": "column",
1414
+ "column_id": "c_equipment_group_id",
1415
+ "alias": "equipment_group_id"
1416
+ },
1417
+ {
1418
+ "kind": "column",
1419
+ "column_id": "c_unit_status",
1420
+ "alias": "unit_status"
1421
+ },
1422
+ {
1423
+ "kind": "column",
1424
+ "column_id": "c_pa_percent",
1425
+ "alias": "pa_percent"
1426
+ }
1427
+ ],
1428
+ "filters": [
1429
+ {
1430
+ "column_id": "c_pa_percent",
1431
+ "op": "<",
1432
+ "value": 50,
1433
+ "value_type": "decimal"
1434
+ }
1435
+ ],
1436
+ "order_by": [
1437
+ {
1438
+ "column_id": "pa_percent",
1439
+ "dir": "asc"
1440
+ }
1441
+ ],
1442
+ "limit": 10
1443
+ }
1444
+ ],
1445
+ "agg_args": [],
1446
+ "infeasible": false
1447
+ },
1448
+ "error": null,
1449
+ "latency_ms": 3493,
1450
+ "tokens": 14790
1451
+ },
1452
+ {
1453
+ "id": "raw_rows_head_en",
1454
+ "category": "counter_raw_rows",
1455
+ "lang": "en",
1456
+ "carried_over": true,
1457
+ "question": "show me the first 5 rows of the data",
1458
+ "passed": true,
1459
+ "checks": [
1460
+ {
1461
+ "check": "no_group_by",
1462
+ "ok": true,
1463
+ "detail": "grouped=False"
1464
+ }
1465
+ ],
1466
+ "facts": {
1467
+ "tools": [
1468
+ "check_data",
1469
+ "retrieve_data"
1470
+ ],
1471
+ "irs": [
1472
+ {
1473
+ "source_id": "src_pa",
1474
+ "table_id": "t_pa",
1475
+ "select": [
1476
+ {
1477
+ "kind": "column",
1478
+ "column_id": "c_keyid",
1479
+ "alias": "KeyId"
1480
+ },
1481
+ {
1482
+ "kind": "column",
1483
+ "column_id": "c_month_id",
1484
+ "alias": "Month_ID"
1485
+ },
1486
+ {
1487
+ "kind": "column",
1488
+ "column_id": "c_site_id",
1489
+ "alias": "Site_ID"
1490
+ },
1491
+ {
1492
+ "kind": "column",
1493
+ "column_id": "c_from_date",
1494
+ "alias": "From_Date"
1495
+ },
1496
+ {
1497
+ "kind": "column",
1498
+ "column_id": "c_to_date",
1499
+ "alias": "To_Date"
1500
+ },
1501
+ {
1502
+ "kind": "column",
1503
+ "column_id": "c_time_description",
1504
+ "alias": "Time_Description"
1505
+ },
1506
+ {
1507
+ "kind": "column",
1508
+ "column_id": "c_model_unit",
1509
+ "alias": "Model_Unit"
1510
+ },
1511
+ {
1512
+ "kind": "column",
1513
+ "column_id": "c_equipment_number",
1514
+ "alias": "Equipment_Number"
1515
+ },
1516
+ {
1517
+ "kind": "column",
1518
+ "column_id": "c_equipment_group_id",
1519
+ "alias": "Equipment_Group_ID"
1520
+ },
1521
+ {
1522
+ "kind": "column",
1523
+ "column_id": "c_unit_status",
1524
+ "alias": "Unit_Status"
1525
+ },
1526
+ {
1527
+ "kind": "column",
1528
+ "column_id": "c_total_breakdown_schedule_hour",
1529
+ "alias": "Total_Breakdown_Schedule_Hour"
1530
+ },
1531
+ {
1532
+ "kind": "column",
1533
+ "column_id": "c_total_breakdown_unschedule_hour",
1534
+ "alias": "Total_Breakdown_Unschedule_Hour"
1535
+ },
1536
+ {
1537
+ "kind": "column",
1538
+ "column_id": "c_total_adj_breakdown_schedule_hour",
1539
+ "alias": "Total_Adj_Breakdown_Schedule_Hour"
1540
+ },
1541
+ {
1542
+ "kind": "column",
1543
+ "column_id": "c_total_adj_breakdown_unschedule_hour",
1544
+ "alias": "Total_Adj_Breakdown_Unschedule_Hour"
1545
+ },
1546
+ {
1547
+ "kind": "column",
1548
+ "column_id": "c_total_mtc_hour",
1549
+ "alias": "Total_MTC_Hour"
1550
+ },
1551
+ {
1552
+ "kind": "column",
1553
+ "column_id": "c_total_down_hour",
1554
+ "alias": "Total_Down_Hour"
1555
+ },
1556
+ {
1557
+ "kind": "column",
1558
+ "column_id": "c_total_frequency_breakdown_schedule",
1559
+ "alias": "Total_Frequency_Breakdown_Schedule"
1560
+ },
1561
+ {
1562
+ "kind": "column",
1563
+ "column_id": "c_total_frequency_breakdown_unschedule",
1564
+ "alias": "Total_Frequency_Breakdown_Unschedule"
1565
+ },
1566
+ {
1567
+ "kind": "column",
1568
+ "column_id": "c_total_frequency_maintenance",
1569
+ "alias": "Total_Frequency_Maintenance"
1570
+ },
1571
+ {
1572
+ "kind": "column",
1573
+ "column_id": "c_total_frequency_tire",
1574
+ "alias": "Total_Frequency_Tire"
1575
+ },
1576
+ {
1577
+ "kind": "column",
1578
+ "column_id": "c_total_frequency_down",
1579
+ "alias": "Total_Frequency_Down"
1580
+ },
1581
+ {
1582
+ "kind": "column",
1583
+ "column_id": "c_total_hours",
1584
+ "alias": "Total_Hours"
1585
+ },
1586
+ {
1587
+ "kind": "column",
1588
+ "column_id": "c_total_inpr_hour",
1589
+ "alias": "Total_INPR_Hour"
1590
+ },
1591
+ {
1592
+ "kind": "column",
1593
+ "column_id": "c_total_record_hm_hour",
1594
+ "alias": "Total_Record_HM_Hour"
1595
+ },
1596
+ {
1597
+ "kind": "column",
1598
+ "column_id": "c_total_hm_mtc_down_hour",
1599
+ "alias": "Total_HM_Mtc_Down_Hour"
1600
+ },
1601
+ {
1602
+ "kind": "column",
1603
+ "column_id": "c_plan_pa_percent",
1604
+ "alias": "Plan_PA_Percent"
1605
+ },
1606
+ {
1607
+ "kind": "column",
1608
+ "column_id": "c_pa_percent",
1609
+ "alias": "PA_Percent"
1610
+ },
1611
+ {
1612
+ "kind": "column",
1613
+ "column_id": "c_mtbs",
1614
+ "alias": "MTBS"
1615
+ },
1616
+ {
1617
+ "kind": "column",
1618
+ "column_id": "c_mttr",
1619
+ "alias": "MTTR"
1620
+ },
1621
+ {
1622
+ "kind": "column",
1623
+ "column_id": "c_sm_percent",
1624
+ "alias": "SM_Percent"
1625
+ },
1626
+ {
1627
+ "kind": "column",
1628
+ "column_id": "c_unschedule_sm_percent",
1629
+ "alias": "Unschedule_SM_Percent"
1630
+ },
1631
+ {
1632
+ "kind": "column",
1633
+ "column_id": "c_isdeleted",
1634
+ "alias": "IsDeleted"
1635
+ },
1636
+ {
1637
+ "kind": "column",
1638
+ "column_id": "c_section",
1639
+ "alias": "Section"
1640
+ },
1641
+ {
1642
+ "kind": "column",
1643
+ "column_id": "c_week_id",
1644
+ "alias": "Week_ID"
1645
+ },
1646
+ {
1647
+ "kind": "column",
1648
+ "column_id": "c_plan_pa_percent_2",
1649
+ "alias": "Plan_PA_Percent_2"
1650
+ },
1651
+ {
1652
+ "kind": "column",
1653
+ "column_id": "c_updated_date",
1654
+ "alias": "Updated_Date"
1655
+ }
1656
+ ],
1657
+ "limit": 5
1658
+ }
1659
+ ],
1660
+ "agg_args": [],
1661
+ "infeasible": false
1662
+ },
1663
+ "error": null,
1664
+ "latency_ms": 4440,
1665
+ "tokens": 15314
1666
+ },
1667
+ {
1668
+ "id": "exact_model_777d_id",
1669
+ "category": "counter_exact_filter",
1670
+ "lang": "id",
1671
+ "carried_over": true,
1672
+ "question": "berapa banyak record untuk model 777D?",
1673
+ "passed": true,
1674
+ "checks": [
1675
+ {
1676
+ "check": "select_agg=count",
1677
+ "ok": true,
1678
+ "detail": "aggs=['count']"
1679
+ },
1680
+ {
1681
+ "check": "has_filter",
1682
+ "ok": true,
1683
+ "detail": "filter_present=True"
1684
+ }
1685
+ ],
1686
+ "facts": {
1687
+ "tools": [
1688
+ "retrieve_data"
1689
+ ],
1690
+ "irs": [
1691
+ {
1692
+ "source_id": "src_pa",
1693
+ "table_id": "t_pa",
1694
+ "select": [
1695
+ {
1696
+ "kind": "agg",
1697
+ "fn": "count",
1698
+ "alias": "record_count"
1699
+ }
1700
+ ],
1701
+ "filters": [
1702
+ {
1703
+ "column_id": "c_model_unit",
1704
+ "op": "=",
1705
+ "value": "777D",
1706
+ "value_type": "string"
1707
+ }
1708
+ ]
1709
+ }
1710
+ ],
1711
+ "agg_args": [],
1712
+ "infeasible": false
1713
+ },
1714
+ "error": null,
1715
+ "latency_ms": 2035,
1716
+ "tokens": 14580
1717
+ },
1718
+ {
1719
+ "id": "exact_section_loader_en",
1720
+ "category": "counter_exact_filter",
1721
+ "lang": "en",
1722
+ "carried_over": true,
1723
+ "question": "how many records are in the OB LOADER section?",
1724
+ "passed": true,
1725
+ "checks": [
1726
+ {
1727
+ "check": "select_agg=count",
1728
+ "ok": true,
1729
+ "detail": "aggs=['count']"
1730
+ },
1731
+ {
1732
+ "check": "has_filter",
1733
+ "ok": true,
1734
+ "detail": "filter_present=True"
1735
+ }
1736
+ ],
1737
+ "facts": {
1738
+ "tools": [
1739
+ "retrieve_data"
1740
+ ],
1741
+ "irs": [
1742
+ {
1743
+ "source_id": "src_pa",
1744
+ "table_id": "t_pa",
1745
+ "select": [
1746
+ {
1747
+ "kind": "agg",
1748
+ "fn": "count",
1749
+ "alias": "record_count"
1750
+ }
1751
+ ],
1752
+ "filters": [
1753
+ {
1754
+ "column_id": "c_section",
1755
+ "op": "=",
1756
+ "value": "OB LOADER",
1757
+ "value_type": "string"
1758
+ }
1759
+ ]
1760
+ }
1761
+ ],
1762
+ "agg_args": [],
1763
+ "infeasible": false
1764
+ },
1765
+ "error": null,
1766
+ "latency_ms": 2133,
1767
+ "tokens": 14570
1768
+ },
1769
+ {
1770
+ "id": "infeasible_churn_id",
1771
+ "category": "infeasible",
1772
+ "lang": "id",
1773
+ "carried_over": true,
1774
+ "question": "berapa churn rate pelanggan?",
1775
+ "passed": true,
1776
+ "checks": [
1777
+ {
1778
+ "check": "infeasible",
1779
+ "ok": true,
1780
+ "detail": "infeasible=True"
1781
+ }
1782
+ ],
1783
+ "facts": {
1784
+ "tools": [],
1785
+ "irs": [],
1786
+ "agg_args": [],
1787
+ "infeasible": true
1788
+ },
1789
+ "error": null,
1790
+ "latency_ms": 2586,
1791
+ "tokens": 14512
1792
+ },
1793
+ {
1794
+ "id": "infeasible_profit_en",
1795
+ "category": "infeasible",
1796
+ "lang": "en",
1797
+ "carried_over": true,
1798
+ "question": "what is the monthly profit margin?",
1799
+ "passed": true,
1800
+ "checks": [
1801
+ {
1802
+ "check": "infeasible",
1803
+ "ok": true,
1804
+ "detail": "infeasible=True"
1805
+ }
1806
+ ],
1807
+ "facts": {
1808
+ "tools": [],
1809
+ "irs": [],
1810
+ "agg_args": [],
1811
+ "infeasible": true
1812
+ },
1813
+ "error": null,
1814
+ "latency_ms": 2206,
1815
+ "tokens": 14551
1816
+ }
1817
+ ]
1818
+ }
eval/planner/results/planner_result_2026-07-23_152650.json ADDED
@@ -0,0 +1,2123 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "timestamp": "2026-07-23_152650",
3
+ "total": 31,
4
+ "passed": 30,
5
+ "cases": [
6
+ {
7
+ "id": "count_zero_pa",
8
+ "category": "count",
9
+ "lang": "en",
10
+ "carried_over": true,
11
+ "question": "how many records have PA_Percent = 0?",
12
+ "passed": true,
13
+ "checks": [
14
+ {
15
+ "check": "select_agg=count",
16
+ "ok": true,
17
+ "detail": "aggs=['count']"
18
+ },
19
+ {
20
+ "check": "has_filter",
21
+ "ok": true,
22
+ "detail": "filter_present=True"
23
+ },
24
+ {
25
+ "check": "no_group_by",
26
+ "ok": true,
27
+ "detail": "grouped=False"
28
+ }
29
+ ],
30
+ "facts": {
31
+ "tools": [
32
+ "retrieve_data"
33
+ ],
34
+ "irs": [
35
+ {
36
+ "source_id": "src_pa",
37
+ "table_id": "t_pa",
38
+ "select": [
39
+ {
40
+ "kind": "agg",
41
+ "fn": "count",
42
+ "alias": "record_count"
43
+ }
44
+ ],
45
+ "filters": [
46
+ {
47
+ "column_id": "c_pa_percent",
48
+ "op": "=",
49
+ "value": 0,
50
+ "value_type": "decimal"
51
+ }
52
+ ]
53
+ }
54
+ ],
55
+ "agg_args": [],
56
+ "infeasible": false
57
+ },
58
+ "error": null,
59
+ "latency_ms": 4042,
60
+ "tokens": 14569
61
+ },
62
+ {
63
+ "id": "count_mttr_gt20_id",
64
+ "category": "count",
65
+ "lang": "id",
66
+ "carried_over": true,
67
+ "question": "berapa banyak record dengan MTTR di atas 20?",
68
+ "passed": true,
69
+ "checks": [
70
+ {
71
+ "check": "select_agg=count",
72
+ "ok": true,
73
+ "detail": "aggs=['count']"
74
+ },
75
+ {
76
+ "check": "has_filter",
77
+ "ok": true,
78
+ "detail": "filter_present=True"
79
+ }
80
+ ],
81
+ "facts": {
82
+ "tools": [
83
+ "retrieve_data"
84
+ ],
85
+ "irs": [
86
+ {
87
+ "source_id": "src_pa",
88
+ "table_id": "t_pa",
89
+ "select": [
90
+ {
91
+ "kind": "agg",
92
+ "fn": "count",
93
+ "alias": "record_count"
94
+ }
95
+ ],
96
+ "filters": [
97
+ {
98
+ "column_id": "c_mttr",
99
+ "op": ">",
100
+ "value": 20,
101
+ "value_type": "decimal"
102
+ }
103
+ ]
104
+ }
105
+ ],
106
+ "agg_args": [],
107
+ "infeasible": false
108
+ },
109
+ "error": null,
110
+ "latency_ms": 2495,
111
+ "tokens": 14571
112
+ },
113
+ {
114
+ "id": "count_section_hauler",
115
+ "category": "count",
116
+ "lang": "en",
117
+ "carried_over": true,
118
+ "question": "how many rows are in section OB HAULER?",
119
+ "passed": true,
120
+ "checks": [
121
+ {
122
+ "check": "select_agg=count",
123
+ "ok": true,
124
+ "detail": "aggs=['count']"
125
+ },
126
+ {
127
+ "check": "has_filter",
128
+ "ok": true,
129
+ "detail": "filter_present=True"
130
+ }
131
+ ],
132
+ "facts": {
133
+ "tools": [
134
+ "retrieve_data"
135
+ ],
136
+ "irs": [
137
+ {
138
+ "source_id": "src_pa",
139
+ "table_id": "t_pa",
140
+ "select": [
141
+ {
142
+ "kind": "agg",
143
+ "fn": "count",
144
+ "alias": "row_count"
145
+ }
146
+ ],
147
+ "filters": [
148
+ {
149
+ "column_id": "c_section",
150
+ "op": "=",
151
+ "value": "OB HAULER",
152
+ "value_type": "string"
153
+ }
154
+ ]
155
+ }
156
+ ],
157
+ "agg_args": [],
158
+ "infeasible": false
159
+ },
160
+ "error": null,
161
+ "latency_ms": 2162,
162
+ "tokens": 14574
163
+ },
164
+ {
165
+ "id": "rank_units_worst_pa_id",
166
+ "category": "ranking",
167
+ "lang": "id",
168
+ "carried_over": false,
169
+ "question": "5 unit dengan PA terburuk?",
170
+ "passed": true,
171
+ "checks": [
172
+ {
173
+ "check": "group_by",
174
+ "ok": true,
175
+ "detail": "grouped=True"
176
+ },
177
+ {
178
+ "check": "group_by_col=Equipment_Number",
179
+ "ok": true,
180
+ "detail": "ids=['c_equipment_number'] aliases=[] resolved=['c_equipment_number']"
181
+ },
182
+ {
183
+ "check": "select_agg=avg",
184
+ "ok": true,
185
+ "detail": "aggs=['avg']"
186
+ },
187
+ {
188
+ "check": "order_dir=asc",
189
+ "ok": true,
190
+ "detail": "dirs=['asc']"
191
+ },
192
+ {
193
+ "check": "limit=5",
194
+ "ok": true,
195
+ "detail": "limits=[5]"
196
+ }
197
+ ],
198
+ "facts": {
199
+ "tools": [
200
+ "check_data",
201
+ "retrieve_data"
202
+ ],
203
+ "irs": [
204
+ {
205
+ "source_id": "src_pa",
206
+ "table_id": "t_pa",
207
+ "select": [
208
+ {
209
+ "kind": "column",
210
+ "column_id": "c_equipment_number",
211
+ "alias": "equipment_number"
212
+ },
213
+ {
214
+ "kind": "agg",
215
+ "fn": "avg",
216
+ "column_id": "c_pa_percent",
217
+ "alias": "avg_pa_percent"
218
+ }
219
+ ],
220
+ "group_by": [
221
+ "c_equipment_number"
222
+ ],
223
+ "order_by": [
224
+ {
225
+ "column_id": "avg_pa_percent",
226
+ "dir": "asc"
227
+ }
228
+ ],
229
+ "limit": 5
230
+ }
231
+ ],
232
+ "agg_args": [],
233
+ "infeasible": false
234
+ },
235
+ "error": null,
236
+ "latency_ms": 3918,
237
+ "tokens": 14745
238
+ },
239
+ {
240
+ "id": "rank_models_top_mttr_id",
241
+ "category": "ranking",
242
+ "lang": "id",
243
+ "carried_over": true,
244
+ "question": "top 3 model dengan MTTR tertinggi?",
245
+ "passed": true,
246
+ "checks": [
247
+ {
248
+ "check": "group_by",
249
+ "ok": true,
250
+ "detail": "grouped=True"
251
+ },
252
+ {
253
+ "check": "group_by_col=Model_Unit",
254
+ "ok": true,
255
+ "detail": "ids=['c_model_unit'] aliases=[] resolved=['c_model_unit']"
256
+ },
257
+ {
258
+ "check": "select_agg=avg",
259
+ "ok": true,
260
+ "detail": "aggs=['avg']"
261
+ },
262
+ {
263
+ "check": "order_dir=desc",
264
+ "ok": true,
265
+ "detail": "dirs=['desc']"
266
+ },
267
+ {
268
+ "check": "limit=3",
269
+ "ok": true,
270
+ "detail": "limits=[3]"
271
+ }
272
+ ],
273
+ "facts": {
274
+ "tools": [
275
+ "check_data",
276
+ "retrieve_data"
277
+ ],
278
+ "irs": [
279
+ {
280
+ "source_id": "src_pa",
281
+ "table_id": "t_pa",
282
+ "select": [
283
+ {
284
+ "kind": "column",
285
+ "column_id": "c_model_unit",
286
+ "alias": "model"
287
+ },
288
+ {
289
+ "kind": "agg",
290
+ "fn": "avg",
291
+ "column_id": "c_mttr",
292
+ "alias": "avg_mttr"
293
+ }
294
+ ],
295
+ "group_by": [
296
+ "c_model_unit"
297
+ ],
298
+ "order_by": [
299
+ {
300
+ "column_id": "avg_mttr",
301
+ "dir": "desc"
302
+ }
303
+ ],
304
+ "limit": 3
305
+ }
306
+ ],
307
+ "agg_args": [],
308
+ "infeasible": false
309
+ },
310
+ "error": null,
311
+ "latency_ms": 2820,
312
+ "tokens": 14719
313
+ },
314
+ {
315
+ "id": "rank_sections_lowest_pa_en",
316
+ "category": "ranking",
317
+ "lang": "en",
318
+ "carried_over": true,
319
+ "question": "which section has the lowest average PA?",
320
+ "passed": true,
321
+ "checks": [
322
+ {
323
+ "check": "group_by",
324
+ "ok": true,
325
+ "detail": "grouped=True"
326
+ },
327
+ {
328
+ "check": "group_by_col=Section",
329
+ "ok": true,
330
+ "detail": "ids=[] aliases=['section'] resolved=['c_section']"
331
+ },
332
+ {
333
+ "check": "select_agg=avg",
334
+ "ok": true,
335
+ "detail": "aggs=['mean']"
336
+ }
337
+ ],
338
+ "facts": {
339
+ "tools": [
340
+ "analyze_aggregate",
341
+ "check_data",
342
+ "retrieve_data"
343
+ ],
344
+ "irs": [
345
+ {
346
+ "source_id": "src_pa",
347
+ "table_id": "t_pa",
348
+ "select": [
349
+ {
350
+ "kind": "column",
351
+ "column_id": "c_section",
352
+ "alias": "section"
353
+ },
354
+ {
355
+ "kind": "column",
356
+ "column_id": "c_pa_percent",
357
+ "alias": "pa_percent"
358
+ }
359
+ ],
360
+ "filters": [
361
+ {
362
+ "column_id": "c_month_id",
363
+ "op": "=",
364
+ "value": 202604,
365
+ "value_type": "int"
366
+ },
367
+ {
368
+ "column_id": "c_isdeleted",
369
+ "op": "=",
370
+ "value": 0,
371
+ "value_type": "int"
372
+ }
373
+ ],
374
+ "limit": 10000
375
+ }
376
+ ],
377
+ "agg_args": [
378
+ {
379
+ "data": "${t2}",
380
+ "aggregations": {
381
+ "pa_percent": [
382
+ "mean"
383
+ ]
384
+ },
385
+ "group_by": [
386
+ "section"
387
+ ]
388
+ }
389
+ ],
390
+ "infeasible": false
391
+ },
392
+ "error": null,
393
+ "latency_ms": 3103,
394
+ "tokens": 14818
395
+ },
396
+ {
397
+ "id": "rank_units_most_breakdown_id",
398
+ "category": "ranking",
399
+ "lang": "id",
400
+ "carried_over": false,
401
+ "question": "unit mana yang paling sering breakdown?",
402
+ "passed": true,
403
+ "checks": [
404
+ {
405
+ "check": "group_by",
406
+ "ok": true,
407
+ "detail": "grouped=True"
408
+ },
409
+ {
410
+ "check": "group_by_col=Equipment_Number",
411
+ "ok": true,
412
+ "detail": "ids=['c_equipment_number'] aliases=[] resolved=['c_equipment_number']"
413
+ },
414
+ {
415
+ "check": "order_dir=desc",
416
+ "ok": true,
417
+ "detail": "dirs=['desc']"
418
+ }
419
+ ],
420
+ "facts": {
421
+ "tools": [
422
+ "check_data",
423
+ "retrieve_data"
424
+ ],
425
+ "irs": [
426
+ {
427
+ "source_id": "src_pa",
428
+ "table_id": "t_pa",
429
+ "select": [
430
+ {
431
+ "kind": "column",
432
+ "column_id": "c_equipment_number",
433
+ "alias": "equipment_number"
434
+ },
435
+ {
436
+ "kind": "agg",
437
+ "fn": "sum",
438
+ "column_id": "c_total_frequency_down",
439
+ "alias": "total_breakdown_frequency"
440
+ }
441
+ ],
442
+ "filters": [
443
+ {
444
+ "column_id": "c_month_id",
445
+ "op": "=",
446
+ "value": 202604,
447
+ "value_type": "int"
448
+ },
449
+ {
450
+ "column_id": "c_isdeleted",
451
+ "op": "=",
452
+ "value": 0,
453
+ "value_type": "int"
454
+ }
455
+ ],
456
+ "group_by": [
457
+ "c_equipment_number"
458
+ ],
459
+ "order_by": [
460
+ {
461
+ "column_id": "total_breakdown_frequency",
462
+ "dir": "desc"
463
+ }
464
+ ],
465
+ "limit": 1
466
+ }
467
+ ],
468
+ "agg_args": [],
469
+ "infeasible": false
470
+ },
471
+ "error": null,
472
+ "latency_ms": 2902,
473
+ "tokens": 14761
474
+ },
475
+ {
476
+ "id": "rank_units_worst_pa_en",
477
+ "category": "ranking",
478
+ "lang": "en",
479
+ "carried_over": false,
480
+ "question": "list the 10 worst units by availability",
481
+ "passed": true,
482
+ "checks": [
483
+ {
484
+ "check": "group_by",
485
+ "ok": true,
486
+ "detail": "grouped=True"
487
+ },
488
+ {
489
+ "check": "group_by_col=Equipment_Number",
490
+ "ok": true,
491
+ "detail": "ids=['c_equipment_number'] aliases=[] resolved=['c_equipment_number']"
492
+ },
493
+ {
494
+ "check": "select_agg=avg",
495
+ "ok": true,
496
+ "detail": "aggs=['avg']"
497
+ },
498
+ {
499
+ "check": "order_dir=asc",
500
+ "ok": true,
501
+ "detail": "dirs=['asc']"
502
+ },
503
+ {
504
+ "check": "limit=10",
505
+ "ok": true,
506
+ "detail": "limits=[10]"
507
+ }
508
+ ],
509
+ "facts": {
510
+ "tools": [
511
+ "check_data",
512
+ "retrieve_data"
513
+ ],
514
+ "irs": [
515
+ {
516
+ "source_id": "src_pa",
517
+ "table_id": "t_pa",
518
+ "select": [
519
+ {
520
+ "kind": "column",
521
+ "column_id": "c_equipment_number",
522
+ "alias": "equipment_number"
523
+ },
524
+ {
525
+ "kind": "agg",
526
+ "fn": "avg",
527
+ "column_id": "c_pa_percent",
528
+ "alias": "avg_pa_percent"
529
+ }
530
+ ],
531
+ "filters": [
532
+ {
533
+ "column_id": "c_isdeleted",
534
+ "op": "=",
535
+ "value": 0,
536
+ "value_type": "int"
537
+ },
538
+ {
539
+ "column_id": "c_month_id",
540
+ "op": "=",
541
+ "value": 202604,
542
+ "value_type": "int"
543
+ }
544
+ ],
545
+ "group_by": [
546
+ "c_equipment_number"
547
+ ],
548
+ "order_by": [
549
+ {
550
+ "column_id": "avg_pa_percent",
551
+ "dir": "asc"
552
+ }
553
+ ],
554
+ "limit": 10
555
+ }
556
+ ],
557
+ "agg_args": [],
558
+ "infeasible": false
559
+ },
560
+ "error": null,
561
+ "latency_ms": 3112,
562
+ "tokens": 14762
563
+ },
564
+ {
565
+ "id": "fuzzy_model_777_id",
566
+ "category": "fuzzy_filter",
567
+ "lang": "id",
568
+ "carried_over": false,
569
+ "question": "berapa banyak model 777?",
570
+ "passed": true,
571
+ "checks": [
572
+ {
573
+ "check": "select_agg=count",
574
+ "ok": true,
575
+ "detail": "aggs=['count']"
576
+ },
577
+ {
578
+ "check": "no_filter_op=in",
579
+ "ok": true,
580
+ "detail": "ops=['=']"
581
+ }
582
+ ],
583
+ "facts": {
584
+ "tools": [
585
+ "retrieve_data"
586
+ ],
587
+ "irs": [
588
+ {
589
+ "source_id": "src_pa",
590
+ "table_id": "t_pa",
591
+ "select": [
592
+ {
593
+ "kind": "agg",
594
+ "fn": "count",
595
+ "alias": "record_count"
596
+ }
597
+ ],
598
+ "filters": [
599
+ {
600
+ "column_id": "c_model_unit",
601
+ "op": "=",
602
+ "value": "777",
603
+ "value_type": "string"
604
+ }
605
+ ],
606
+ "limit": 10000
607
+ }
608
+ ],
609
+ "agg_args": [],
610
+ "infeasible": false
611
+ },
612
+ "error": null,
613
+ "latency_ms": 13617,
614
+ "tokens": 14582
615
+ },
616
+ {
617
+ "id": "fuzzy_model_hd785_id",
618
+ "category": "fuzzy_filter",
619
+ "lang": "id",
620
+ "carried_over": true,
621
+ "question": "berapa banyak unit HD785?",
622
+ "passed": true,
623
+ "checks": [
624
+ {
625
+ "check": "select_agg=count",
626
+ "ok": true,
627
+ "detail": "aggs=['count']"
628
+ },
629
+ {
630
+ "check": "no_filter_op=in",
631
+ "ok": true,
632
+ "detail": "ops=['=']"
633
+ }
634
+ ],
635
+ "facts": {
636
+ "tools": [
637
+ "check_data",
638
+ "retrieve_data"
639
+ ],
640
+ "irs": [
641
+ {
642
+ "source_id": "src_pa",
643
+ "table_id": "t_pa",
644
+ "select": [
645
+ {
646
+ "kind": "agg",
647
+ "fn": "count",
648
+ "alias": "unit_count"
649
+ }
650
+ ],
651
+ "filters": [
652
+ {
653
+ "column_id": "c_model_unit",
654
+ "op": "=",
655
+ "value": "HD785",
656
+ "value_type": "string"
657
+ }
658
+ ]
659
+ }
660
+ ],
661
+ "agg_args": [],
662
+ "infeasible": false
663
+ },
664
+ "error": null,
665
+ "latency_ms": 3472,
666
+ "tokens": 14660
667
+ },
668
+ {
669
+ "id": "fuzzy_model_ex_en",
670
+ "category": "fuzzy_filter",
671
+ "lang": "en",
672
+ "carried_over": false,
673
+ "question": "how many EX excavator units are there?",
674
+ "passed": true,
675
+ "checks": [
676
+ {
677
+ "check": "no_filter_op=in",
678
+ "ok": true,
679
+ "detail": "ops=['like']"
680
+ }
681
+ ],
682
+ "facts": {
683
+ "tools": [
684
+ "retrieve_data"
685
+ ],
686
+ "irs": [
687
+ {
688
+ "source_id": "src_pa",
689
+ "table_id": "t_pa",
690
+ "select": [
691
+ {
692
+ "kind": "agg",
693
+ "fn": "count",
694
+ "alias": "unit_count"
695
+ }
696
+ ],
697
+ "filters": [
698
+ {
699
+ "column_id": "c_model_unit",
700
+ "op": "like",
701
+ "value": "EX%",
702
+ "value_type": "string"
703
+ }
704
+ ],
705
+ "limit": 10000
706
+ }
707
+ ],
708
+ "agg_args": [],
709
+ "infeasible": false
710
+ },
711
+ "error": null,
712
+ "latency_ms": 5809,
713
+ "tokens": 29353
714
+ },
715
+ {
716
+ "id": "agg_pa_per_section_id",
717
+ "category": "aggregate",
718
+ "lang": "id",
719
+ "carried_over": true,
720
+ "question": "berapa rata-rata PA per section?",
721
+ "passed": true,
722
+ "checks": [
723
+ {
724
+ "check": "group_by",
725
+ "ok": true,
726
+ "detail": "grouped=True"
727
+ },
728
+ {
729
+ "check": "group_by_col=Section",
730
+ "ok": true,
731
+ "detail": "ids=[] aliases=['section'] resolved=['c_section']"
732
+ },
733
+ {
734
+ "check": "select_agg=avg",
735
+ "ok": true,
736
+ "detail": "aggs=['mean']"
737
+ }
738
+ ],
739
+ "facts": {
740
+ "tools": [
741
+ "analyze_aggregate",
742
+ "check_data",
743
+ "retrieve_data"
744
+ ],
745
+ "irs": [
746
+ {
747
+ "source_id": "src_pa",
748
+ "table_id": "t_pa",
749
+ "select": [
750
+ {
751
+ "kind": "column",
752
+ "column_id": "c_section",
753
+ "alias": "section"
754
+ },
755
+ {
756
+ "kind": "column",
757
+ "column_id": "c_pa_percent",
758
+ "alias": "pa_percent"
759
+ }
760
+ ],
761
+ "filters": [
762
+ {
763
+ "column_id": "c_isdeleted",
764
+ "op": "=",
765
+ "value": 0,
766
+ "value_type": "int"
767
+ }
768
+ ],
769
+ "limit": 10000
770
+ }
771
+ ],
772
+ "agg_args": [
773
+ {
774
+ "data": "${t2}",
775
+ "aggregations": {
776
+ "pa_percent": [
777
+ "mean"
778
+ ]
779
+ },
780
+ "group_by": [
781
+ "section"
782
+ ]
783
+ }
784
+ ],
785
+ "infeasible": false
786
+ },
787
+ "error": null,
788
+ "latency_ms": 3942,
789
+ "tokens": 14785
790
+ },
791
+ {
792
+ "id": "agg_mttr_per_model_en",
793
+ "category": "aggregate",
794
+ "lang": "en",
795
+ "carried_over": true,
796
+ "question": "what is the average MTTR per model unit?",
797
+ "passed": true,
798
+ "checks": [
799
+ {
800
+ "check": "group_by",
801
+ "ok": true,
802
+ "detail": "grouped=True"
803
+ },
804
+ {
805
+ "check": "group_by_col=Model_Unit",
806
+ "ok": true,
807
+ "detail": "ids=[] aliases=['model_unit'] resolved=['c_model_unit']"
808
+ },
809
+ {
810
+ "check": "select_agg=avg",
811
+ "ok": true,
812
+ "detail": "aggs=['mean']"
813
+ }
814
+ ],
815
+ "facts": {
816
+ "tools": [
817
+ "analyze_aggregate",
818
+ "check_data",
819
+ "retrieve_data"
820
+ ],
821
+ "irs": [
822
+ {
823
+ "source_id": "src_pa",
824
+ "table_id": "t_pa",
825
+ "select": [
826
+ {
827
+ "kind": "column",
828
+ "column_id": "c_model_unit",
829
+ "alias": "model_unit"
830
+ },
831
+ {
832
+ "kind": "column",
833
+ "column_id": "c_mttr",
834
+ "alias": "mttr"
835
+ }
836
+ ],
837
+ "limit": 10000
838
+ }
839
+ ],
840
+ "agg_args": [
841
+ {
842
+ "data": "${t2}",
843
+ "aggregations": {
844
+ "mttr": [
845
+ "mean"
846
+ ]
847
+ },
848
+ "group_by": [
849
+ "model_unit"
850
+ ]
851
+ }
852
+ ],
853
+ "infeasible": false
854
+ },
855
+ "error": null,
856
+ "latency_ms": 3648,
857
+ "tokens": 14781
858
+ },
859
+ {
860
+ "id": "agg_downhour_per_group_id",
861
+ "category": "aggregate",
862
+ "lang": "id",
863
+ "carried_over": true,
864
+ "question": "total down hour per equipment group?",
865
+ "passed": true,
866
+ "checks": [
867
+ {
868
+ "check": "group_by",
869
+ "ok": true,
870
+ "detail": "grouped=True"
871
+ },
872
+ {
873
+ "check": "group_by_col=Equipment_Group_ID",
874
+ "ok": true,
875
+ "detail": "ids=[] aliases=['equipment_group'] resolved=['c_equipment_group_id']"
876
+ },
877
+ {
878
+ "check": "select_agg=sum",
879
+ "ok": true,
880
+ "detail": "aggs=['sum']"
881
+ }
882
+ ],
883
+ "facts": {
884
+ "tools": [
885
+ "analyze_aggregate",
886
+ "check_data",
887
+ "retrieve_data"
888
+ ],
889
+ "irs": [
890
+ {
891
+ "source_id": "src_pa",
892
+ "table_id": "t_pa",
893
+ "select": [
894
+ {
895
+ "kind": "column",
896
+ "column_id": "c_equipment_group_id",
897
+ "alias": "equipment_group"
898
+ },
899
+ {
900
+ "kind": "column",
901
+ "column_id": "c_total_down_hour",
902
+ "alias": "down_hour"
903
+ }
904
+ ],
905
+ "filters": [
906
+ {
907
+ "column_id": "c_month_id",
908
+ "op": "=",
909
+ "value": 202604,
910
+ "value_type": "int"
911
+ },
912
+ {
913
+ "column_id": "c_isdeleted",
914
+ "op": "=",
915
+ "value": 0,
916
+ "value_type": "int"
917
+ }
918
+ ],
919
+ "limit": 10000
920
+ }
921
+ ],
922
+ "agg_args": [
923
+ {
924
+ "data": "${t2}",
925
+ "aggregations": {
926
+ "down_hour": [
927
+ "sum"
928
+ ]
929
+ },
930
+ "group_by": [
931
+ "equipment_group"
932
+ ]
933
+ }
934
+ ],
935
+ "infeasible": false
936
+ },
937
+ "error": null,
938
+ "latency_ms": 5057,
939
+ "tokens": 14830
940
+ },
941
+ {
942
+ "id": "desc_mttr_stats_id",
943
+ "category": "descriptive",
944
+ "lang": "id",
945
+ "carried_over": true,
946
+ "question": "berikan ringkasan statistik MTTR",
947
+ "passed": true,
948
+ "checks": [
949
+ {
950
+ "check": "has_tool=analyze_descriptive",
951
+ "ok": true,
952
+ "detail": "tools=['analyze_descriptive', 'check_data', 'retrieve_data']"
953
+ }
954
+ ],
955
+ "facts": {
956
+ "tools": [
957
+ "analyze_descriptive",
958
+ "check_data",
959
+ "retrieve_data"
960
+ ],
961
+ "irs": [
962
+ {
963
+ "source_id": "src_pa",
964
+ "table_id": "t_pa",
965
+ "select": [
966
+ {
967
+ "kind": "column",
968
+ "column_id": "c_mttr",
969
+ "alias": "mttr"
970
+ }
971
+ ],
972
+ "limit": 10000
973
+ }
974
+ ],
975
+ "agg_args": [],
976
+ "infeasible": false
977
+ },
978
+ "error": null,
979
+ "latency_ms": 3173,
980
+ "tokens": 14707
981
+ },
982
+ {
983
+ "id": "desc_pa_stats_en",
984
+ "category": "descriptive",
985
+ "lang": "en",
986
+ "carried_over": true,
987
+ "question": "give me the summary statistics for PA_Percent",
988
+ "passed": true,
989
+ "checks": [
990
+ {
991
+ "check": "has_tool=analyze_descriptive",
992
+ "ok": true,
993
+ "detail": "tools=['analyze_descriptive', 'check_data', 'retrieve_data']"
994
+ }
995
+ ],
996
+ "facts": {
997
+ "tools": [
998
+ "analyze_descriptive",
999
+ "check_data",
1000
+ "retrieve_data"
1001
+ ],
1002
+ "irs": [
1003
+ {
1004
+ "source_id": "src_pa",
1005
+ "table_id": "t_pa",
1006
+ "select": [
1007
+ {
1008
+ "kind": "column",
1009
+ "column_id": "c_pa_percent",
1010
+ "alias": "PA_Percent"
1011
+ }
1012
+ ],
1013
+ "limit": 10000
1014
+ }
1015
+ ],
1016
+ "agg_args": [],
1017
+ "infeasible": false
1018
+ },
1019
+ "error": null,
1020
+ "latency_ms": 3135,
1021
+ "tokens": 14717
1022
+ },
1023
+ {
1024
+ "id": "corr_mttr_pa_id",
1025
+ "category": "correlation",
1026
+ "lang": "id",
1027
+ "carried_over": true,
1028
+ "question": "apakah ada korelasi antara MTTR dan PA?",
1029
+ "passed": true,
1030
+ "checks": [
1031
+ {
1032
+ "check": "has_tool=analyze_correlation",
1033
+ "ok": true,
1034
+ "detail": "tools=['analyze_correlation', 'check_data', 'retrieve_data']"
1035
+ }
1036
+ ],
1037
+ "facts": {
1038
+ "tools": [
1039
+ "analyze_correlation",
1040
+ "check_data",
1041
+ "retrieve_data"
1042
+ ],
1043
+ "irs": [
1044
+ {
1045
+ "source_id": "src_pa",
1046
+ "table_id": "t_pa",
1047
+ "select": [
1048
+ {
1049
+ "kind": "column",
1050
+ "column_id": "c_mttr",
1051
+ "alias": "mttr"
1052
+ },
1053
+ {
1054
+ "kind": "column",
1055
+ "column_id": "c_pa_percent",
1056
+ "alias": "pa_percent"
1057
+ }
1058
+ ],
1059
+ "filters": [
1060
+ {
1061
+ "column_id": "c_month_id",
1062
+ "op": "=",
1063
+ "value": 202604,
1064
+ "value_type": "int"
1065
+ },
1066
+ {
1067
+ "column_id": "c_isdeleted",
1068
+ "op": "=",
1069
+ "value": 0,
1070
+ "value_type": "int"
1071
+ }
1072
+ ],
1073
+ "limit": 10000
1074
+ }
1075
+ ],
1076
+ "agg_args": [],
1077
+ "infeasible": false
1078
+ },
1079
+ "error": null,
1080
+ "latency_ms": 3920,
1081
+ "tokens": 14798
1082
+ },
1083
+ {
1084
+ "id": "corr_freq_pa_en",
1085
+ "category": "correlation",
1086
+ "lang": "en",
1087
+ "carried_over": true,
1088
+ "question": "is breakdown frequency correlated with availability?",
1089
+ "passed": true,
1090
+ "checks": [
1091
+ {
1092
+ "check": "has_tool=analyze_correlation",
1093
+ "ok": true,
1094
+ "detail": "tools=['analyze_correlation', 'check_data', 'retrieve_data']"
1095
+ }
1096
+ ],
1097
+ "facts": {
1098
+ "tools": [
1099
+ "analyze_correlation",
1100
+ "check_data",
1101
+ "retrieve_data"
1102
+ ],
1103
+ "irs": [
1104
+ {
1105
+ "source_id": "src_pa",
1106
+ "table_id": "t_pa",
1107
+ "select": [
1108
+ {
1109
+ "kind": "column",
1110
+ "column_id": "c_total_frequency_down",
1111
+ "alias": "breakdown_frequency"
1112
+ },
1113
+ {
1114
+ "kind": "column",
1115
+ "column_id": "c_pa_percent",
1116
+ "alias": "availability"
1117
+ }
1118
+ ],
1119
+ "filters": [
1120
+ {
1121
+ "column_id": "c_month_id",
1122
+ "op": "=",
1123
+ "value": 202604,
1124
+ "value_type": "int"
1125
+ },
1126
+ {
1127
+ "column_id": "c_isdeleted",
1128
+ "op": "=",
1129
+ "value": 0,
1130
+ "value_type": "int"
1131
+ }
1132
+ ],
1133
+ "limit": 10000
1134
+ }
1135
+ ],
1136
+ "agg_args": [],
1137
+ "infeasible": false
1138
+ },
1139
+ "error": null,
1140
+ "latency_ms": 3774,
1141
+ "tokens": 14845
1142
+ },
1143
+ {
1144
+ "id": "trend_pa_daily_id",
1145
+ "category": "trend",
1146
+ "lang": "id",
1147
+ "carried_over": true,
1148
+ "question": "bagaimana trend PA harian?",
1149
+ "passed": true,
1150
+ "checks": [
1151
+ {
1152
+ "check": "has_tool=analyze_trend",
1153
+ "ok": true,
1154
+ "detail": "tools=['analyze_trend', 'check_data', 'retrieve_data']"
1155
+ }
1156
+ ],
1157
+ "facts": {
1158
+ "tools": [
1159
+ "analyze_trend",
1160
+ "check_data",
1161
+ "retrieve_data"
1162
+ ],
1163
+ "irs": [
1164
+ {
1165
+ "source_id": "src_pa",
1166
+ "table_id": "t_pa",
1167
+ "select": [
1168
+ {
1169
+ "kind": "column",
1170
+ "column_id": "c_from_date",
1171
+ "alias": "date"
1172
+ },
1173
+ {
1174
+ "kind": "column",
1175
+ "column_id": "c_pa_percent",
1176
+ "alias": "pa_percent"
1177
+ }
1178
+ ],
1179
+ "filters": [
1180
+ {
1181
+ "column_id": "c_month_id",
1182
+ "op": "=",
1183
+ "value": 202604,
1184
+ "value_type": "int"
1185
+ },
1186
+ {
1187
+ "column_id": "c_isdeleted",
1188
+ "op": "=",
1189
+ "value": 0,
1190
+ "value_type": "int"
1191
+ }
1192
+ ],
1193
+ "limit": 10000
1194
+ }
1195
+ ],
1196
+ "agg_args": [],
1197
+ "infeasible": false
1198
+ },
1199
+ "error": null,
1200
+ "latency_ms": 4485,
1201
+ "tokens": 14852
1202
+ },
1203
+ {
1204
+ "id": "trend_downhour_en",
1205
+ "category": "trend",
1206
+ "lang": "en",
1207
+ "carried_over": true,
1208
+ "question": "show the trend of total down hours over time",
1209
+ "passed": true,
1210
+ "checks": [
1211
+ {
1212
+ "check": "has_tool=analyze_trend",
1213
+ "ok": true,
1214
+ "detail": "tools=['analyze_trend', 'check_data', 'retrieve_data']"
1215
+ }
1216
+ ],
1217
+ "facts": {
1218
+ "tools": [
1219
+ "analyze_trend",
1220
+ "check_data",
1221
+ "retrieve_data"
1222
+ ],
1223
+ "irs": [
1224
+ {
1225
+ "source_id": "src_pa",
1226
+ "table_id": "t_pa",
1227
+ "select": [
1228
+ {
1229
+ "kind": "column",
1230
+ "column_id": "c_from_date",
1231
+ "alias": "from_date"
1232
+ },
1233
+ {
1234
+ "kind": "column",
1235
+ "column_id": "c_total_down_hour",
1236
+ "alias": "total_down_hour"
1237
+ }
1238
+ ],
1239
+ "filters": [
1240
+ {
1241
+ "column_id": "c_month_id",
1242
+ "op": "=",
1243
+ "value": 202604,
1244
+ "value_type": "int"
1245
+ }
1246
+ ],
1247
+ "limit": 10000
1248
+ }
1249
+ ],
1250
+ "agg_args": [],
1251
+ "infeasible": false
1252
+ },
1253
+ "error": null,
1254
+ "latency_ms": 3686,
1255
+ "tokens": 14832
1256
+ },
1257
+ {
1258
+ "id": "merge_worst_pa_and_mttr_id",
1259
+ "category": "merge",
1260
+ "lang": "id",
1261
+ "carried_over": true,
1262
+ "question": "model mana yang PA-nya paling buruk sekaligus MTTR-nya paling tinggi?",
1263
+ "passed": true,
1264
+ "checks": [
1265
+ {
1266
+ "check": "group_by",
1267
+ "ok": true,
1268
+ "detail": "grouped=True"
1269
+ },
1270
+ {
1271
+ "check": "group_by_col=Model_Unit",
1272
+ "ok": true,
1273
+ "detail": "ids=['c_model_unit', 'c_model_unit'] aliases=[] resolved=['c_model_unit', 'c_model_unit']"
1274
+ }
1275
+ ],
1276
+ "facts": {
1277
+ "tools": [
1278
+ "analyze_merge",
1279
+ "check_data",
1280
+ "retrieve_data"
1281
+ ],
1282
+ "irs": [
1283
+ {
1284
+ "source_id": "src_pa",
1285
+ "table_id": "t_pa",
1286
+ "select": [
1287
+ {
1288
+ "kind": "column",
1289
+ "column_id": "c_model_unit",
1290
+ "alias": "model"
1291
+ },
1292
+ {
1293
+ "kind": "agg",
1294
+ "fn": "avg",
1295
+ "column_id": "c_pa_percent",
1296
+ "alias": "avg_pa"
1297
+ }
1298
+ ],
1299
+ "filters": [
1300
+ {
1301
+ "column_id": "c_from_date",
1302
+ "op": "between",
1303
+ "value": [
1304
+ "2026-04-01",
1305
+ "2026-04-30"
1306
+ ],
1307
+ "value_type": "date"
1308
+ }
1309
+ ],
1310
+ "group_by": [
1311
+ "c_model_unit"
1312
+ ],
1313
+ "order_by": [
1314
+ {
1315
+ "column_id": "avg_pa",
1316
+ "dir": "asc"
1317
+ }
1318
+ ],
1319
+ "limit": 1
1320
+ },
1321
+ {
1322
+ "source_id": "src_pa",
1323
+ "table_id": "t_pa",
1324
+ "select": [
1325
+ {
1326
+ "kind": "column",
1327
+ "column_id": "c_model_unit",
1328
+ "alias": "model"
1329
+ },
1330
+ {
1331
+ "kind": "agg",
1332
+ "fn": "avg",
1333
+ "column_id": "c_mttr",
1334
+ "alias": "avg_mttr"
1335
+ }
1336
+ ],
1337
+ "filters": [
1338
+ {
1339
+ "column_id": "c_from_date",
1340
+ "op": "between",
1341
+ "value": [
1342
+ "2026-04-01",
1343
+ "2026-04-30"
1344
+ ],
1345
+ "value_type": "date"
1346
+ }
1347
+ ],
1348
+ "group_by": [
1349
+ "c_model_unit"
1350
+ ],
1351
+ "order_by": [
1352
+ {
1353
+ "column_id": "avg_mttr",
1354
+ "dir": "desc"
1355
+ }
1356
+ ],
1357
+ "limit": 1
1358
+ }
1359
+ ],
1360
+ "agg_args": [],
1361
+ "infeasible": false
1362
+ },
1363
+ "error": null,
1364
+ "latency_ms": 5546,
1365
+ "tokens": 15165
1366
+ },
1367
+ {
1368
+ "id": "raw_rows_low_pa_id",
1369
+ "category": "counter_raw_rows",
1370
+ "lang": "id",
1371
+ "carried_over": true,
1372
+ "question": "tampilkan 10 record dengan PA di bawah 50",
1373
+ "passed": true,
1374
+ "checks": [
1375
+ {
1376
+ "check": "no_group_by",
1377
+ "ok": true,
1378
+ "detail": "grouped=False"
1379
+ },
1380
+ {
1381
+ "check": "has_filter",
1382
+ "ok": true,
1383
+ "detail": "filter_present=True"
1384
+ },
1385
+ {
1386
+ "check": "limit=10",
1387
+ "ok": true,
1388
+ "detail": "limits=[10]"
1389
+ }
1390
+ ],
1391
+ "facts": {
1392
+ "tools": [
1393
+ "check_data",
1394
+ "retrieve_data"
1395
+ ],
1396
+ "irs": [
1397
+ {
1398
+ "source_id": "src_pa",
1399
+ "table_id": "t_pa",
1400
+ "select": [
1401
+ {
1402
+ "kind": "column",
1403
+ "column_id": "c_keyid",
1404
+ "alias": "key_id"
1405
+ },
1406
+ {
1407
+ "kind": "column",
1408
+ "column_id": "c_month_id",
1409
+ "alias": "month_id"
1410
+ },
1411
+ {
1412
+ "kind": "column",
1413
+ "column_id": "c_site_id",
1414
+ "alias": "site_id"
1415
+ },
1416
+ {
1417
+ "kind": "column",
1418
+ "column_id": "c_from_date",
1419
+ "alias": "from_date"
1420
+ },
1421
+ {
1422
+ "kind": "column",
1423
+ "column_id": "c_to_date",
1424
+ "alias": "to_date"
1425
+ },
1426
+ {
1427
+ "kind": "column",
1428
+ "column_id": "c_time_description",
1429
+ "alias": "time_description"
1430
+ },
1431
+ {
1432
+ "kind": "column",
1433
+ "column_id": "c_model_unit",
1434
+ "alias": "model_unit"
1435
+ },
1436
+ {
1437
+ "kind": "column",
1438
+ "column_id": "c_equipment_number",
1439
+ "alias": "equipment_number"
1440
+ },
1441
+ {
1442
+ "kind": "column",
1443
+ "column_id": "c_equipment_group_id",
1444
+ "alias": "equipment_group_id"
1445
+ },
1446
+ {
1447
+ "kind": "column",
1448
+ "column_id": "c_unit_status",
1449
+ "alias": "unit_status"
1450
+ },
1451
+ {
1452
+ "kind": "column",
1453
+ "column_id": "c_pa_percent",
1454
+ "alias": "pa_percent"
1455
+ }
1456
+ ],
1457
+ "filters": [
1458
+ {
1459
+ "column_id": "c_pa_percent",
1460
+ "op": "<",
1461
+ "value": 50,
1462
+ "value_type": "decimal"
1463
+ }
1464
+ ],
1465
+ "order_by": [
1466
+ {
1467
+ "column_id": "pa_percent",
1468
+ "dir": "asc"
1469
+ }
1470
+ ],
1471
+ "limit": 10
1472
+ }
1473
+ ],
1474
+ "agg_args": [],
1475
+ "infeasible": false
1476
+ },
1477
+ "error": null,
1478
+ "latency_ms": 3525,
1479
+ "tokens": 14840
1480
+ },
1481
+ {
1482
+ "id": "raw_rows_head_en",
1483
+ "category": "counter_raw_rows",
1484
+ "lang": "en",
1485
+ "carried_over": true,
1486
+ "question": "show me the first 5 rows of the data",
1487
+ "passed": true,
1488
+ "checks": [
1489
+ {
1490
+ "check": "no_group_by",
1491
+ "ok": true,
1492
+ "detail": "grouped=False"
1493
+ }
1494
+ ],
1495
+ "facts": {
1496
+ "tools": [
1497
+ "check_data",
1498
+ "retrieve_data"
1499
+ ],
1500
+ "irs": [
1501
+ {
1502
+ "source_id": "src_pa",
1503
+ "table_id": "t_pa",
1504
+ "select": [
1505
+ {
1506
+ "kind": "column",
1507
+ "column_id": "c_keyid",
1508
+ "alias": "KeyId"
1509
+ },
1510
+ {
1511
+ "kind": "column",
1512
+ "column_id": "c_month_id",
1513
+ "alias": "Month_ID"
1514
+ },
1515
+ {
1516
+ "kind": "column",
1517
+ "column_id": "c_site_id",
1518
+ "alias": "Site_ID"
1519
+ },
1520
+ {
1521
+ "kind": "column",
1522
+ "column_id": "c_from_date",
1523
+ "alias": "From_Date"
1524
+ },
1525
+ {
1526
+ "kind": "column",
1527
+ "column_id": "c_to_date",
1528
+ "alias": "To_Date"
1529
+ },
1530
+ {
1531
+ "kind": "column",
1532
+ "column_id": "c_time_description",
1533
+ "alias": "Time_Description"
1534
+ },
1535
+ {
1536
+ "kind": "column",
1537
+ "column_id": "c_model_unit",
1538
+ "alias": "Model_Unit"
1539
+ },
1540
+ {
1541
+ "kind": "column",
1542
+ "column_id": "c_equipment_number",
1543
+ "alias": "Equipment_Number"
1544
+ },
1545
+ {
1546
+ "kind": "column",
1547
+ "column_id": "c_equipment_group_id",
1548
+ "alias": "Equipment_Group_ID"
1549
+ },
1550
+ {
1551
+ "kind": "column",
1552
+ "column_id": "c_unit_status",
1553
+ "alias": "Unit_Status"
1554
+ },
1555
+ {
1556
+ "kind": "column",
1557
+ "column_id": "c_total_breakdown_schedule_hour",
1558
+ "alias": "Total_Breakdown_Schedule_Hour"
1559
+ },
1560
+ {
1561
+ "kind": "column",
1562
+ "column_id": "c_total_breakdown_unschedule_hour",
1563
+ "alias": "Total_Breakdown_Unschedule_Hour"
1564
+ },
1565
+ {
1566
+ "kind": "column",
1567
+ "column_id": "c_total_adj_breakdown_schedule_hour",
1568
+ "alias": "Total_Adj_Breakdown_Schedule_Hour"
1569
+ },
1570
+ {
1571
+ "kind": "column",
1572
+ "column_id": "c_total_adj_breakdown_unschedule_hour",
1573
+ "alias": "Total_Adj_Breakdown_Unschedule_Hour"
1574
+ },
1575
+ {
1576
+ "kind": "column",
1577
+ "column_id": "c_total_mtc_hour",
1578
+ "alias": "Total_MTC_Hour"
1579
+ },
1580
+ {
1581
+ "kind": "column",
1582
+ "column_id": "c_total_down_hour",
1583
+ "alias": "Total_Down_Hour"
1584
+ },
1585
+ {
1586
+ "kind": "column",
1587
+ "column_id": "c_total_frequency_breakdown_schedule",
1588
+ "alias": "Total_Frequency_Breakdown_Schedule"
1589
+ },
1590
+ {
1591
+ "kind": "column",
1592
+ "column_id": "c_total_frequency_breakdown_unschedule",
1593
+ "alias": "Total_Frequency_Breakdown_Unschedule"
1594
+ },
1595
+ {
1596
+ "kind": "column",
1597
+ "column_id": "c_total_frequency_maintenance",
1598
+ "alias": "Total_Frequency_Maintenance"
1599
+ },
1600
+ {
1601
+ "kind": "column",
1602
+ "column_id": "c_total_frequency_tire",
1603
+ "alias": "Total_Frequency_Tire"
1604
+ },
1605
+ {
1606
+ "kind": "column",
1607
+ "column_id": "c_total_frequency_down",
1608
+ "alias": "Total_Frequency_Down"
1609
+ },
1610
+ {
1611
+ "kind": "column",
1612
+ "column_id": "c_total_hours",
1613
+ "alias": "Total_Hours"
1614
+ },
1615
+ {
1616
+ "kind": "column",
1617
+ "column_id": "c_total_inpr_hour",
1618
+ "alias": "Total_INPR_Hour"
1619
+ },
1620
+ {
1621
+ "kind": "column",
1622
+ "column_id": "c_total_record_hm_hour",
1623
+ "alias": "Total_Record_HM_Hour"
1624
+ },
1625
+ {
1626
+ "kind": "column",
1627
+ "column_id": "c_total_hm_mtc_down_hour",
1628
+ "alias": "Total_HM_Mtc_Down_Hour"
1629
+ },
1630
+ {
1631
+ "kind": "column",
1632
+ "column_id": "c_plan_pa_percent",
1633
+ "alias": "Plan_PA_Percent"
1634
+ },
1635
+ {
1636
+ "kind": "column",
1637
+ "column_id": "c_pa_percent",
1638
+ "alias": "PA_Percent"
1639
+ },
1640
+ {
1641
+ "kind": "column",
1642
+ "column_id": "c_mtbs",
1643
+ "alias": "MTBS"
1644
+ },
1645
+ {
1646
+ "kind": "column",
1647
+ "column_id": "c_mttr",
1648
+ "alias": "MTTR"
1649
+ },
1650
+ {
1651
+ "kind": "column",
1652
+ "column_id": "c_sm_percent",
1653
+ "alias": "SM_Percent"
1654
+ },
1655
+ {
1656
+ "kind": "column",
1657
+ "column_id": "c_unschedule_sm_percent",
1658
+ "alias": "Unschedule_SM_Percent"
1659
+ },
1660
+ {
1661
+ "kind": "column",
1662
+ "column_id": "c_isdeleted",
1663
+ "alias": "IsDeleted"
1664
+ },
1665
+ {
1666
+ "kind": "column",
1667
+ "column_id": "c_section",
1668
+ "alias": "Section"
1669
+ },
1670
+ {
1671
+ "kind": "column",
1672
+ "column_id": "c_week_id",
1673
+ "alias": "Week_ID"
1674
+ },
1675
+ {
1676
+ "kind": "column",
1677
+ "column_id": "c_plan_pa_percent_2",
1678
+ "alias": "Plan_PA_Percent_2"
1679
+ },
1680
+ {
1681
+ "kind": "column",
1682
+ "column_id": "c_updated_date",
1683
+ "alias": "Updated_Date"
1684
+ }
1685
+ ],
1686
+ "limit": 5
1687
+ }
1688
+ ],
1689
+ "agg_args": [],
1690
+ "infeasible": false
1691
+ },
1692
+ "error": null,
1693
+ "latency_ms": 6662,
1694
+ "tokens": 15316
1695
+ },
1696
+ {
1697
+ "id": "exact_model_777d_id",
1698
+ "category": "counter_exact_filter",
1699
+ "lang": "id",
1700
+ "carried_over": true,
1701
+ "question": "berapa banyak record untuk model 777D?",
1702
+ "passed": true,
1703
+ "checks": [
1704
+ {
1705
+ "check": "select_agg=count",
1706
+ "ok": true,
1707
+ "detail": "aggs=['count']"
1708
+ },
1709
+ {
1710
+ "check": "has_filter",
1711
+ "ok": true,
1712
+ "detail": "filter_present=True"
1713
+ }
1714
+ ],
1715
+ "facts": {
1716
+ "tools": [
1717
+ "retrieve_data"
1718
+ ],
1719
+ "irs": [
1720
+ {
1721
+ "source_id": "src_pa",
1722
+ "table_id": "t_pa",
1723
+ "select": [
1724
+ {
1725
+ "kind": "agg",
1726
+ "fn": "count",
1727
+ "alias": "record_count"
1728
+ }
1729
+ ],
1730
+ "filters": [
1731
+ {
1732
+ "column_id": "c_model_unit",
1733
+ "op": "=",
1734
+ "value": "777D",
1735
+ "value_type": "string"
1736
+ }
1737
+ ]
1738
+ }
1739
+ ],
1740
+ "agg_args": [],
1741
+ "infeasible": false
1742
+ },
1743
+ "error": null,
1744
+ "latency_ms": 2422,
1745
+ "tokens": 14573
1746
+ },
1747
+ {
1748
+ "id": "exact_section_loader_en",
1749
+ "category": "counter_exact_filter",
1750
+ "lang": "en",
1751
+ "carried_over": true,
1752
+ "question": "how many records are in the OB LOADER section?",
1753
+ "passed": true,
1754
+ "checks": [
1755
+ {
1756
+ "check": "select_agg=count",
1757
+ "ok": true,
1758
+ "detail": "aggs=['count']"
1759
+ },
1760
+ {
1761
+ "check": "has_filter",
1762
+ "ok": true,
1763
+ "detail": "filter_present=True"
1764
+ }
1765
+ ],
1766
+ "facts": {
1767
+ "tools": [
1768
+ "retrieve_data"
1769
+ ],
1770
+ "irs": [
1771
+ {
1772
+ "source_id": "src_pa",
1773
+ "table_id": "t_pa",
1774
+ "select": [
1775
+ {
1776
+ "kind": "agg",
1777
+ "fn": "count",
1778
+ "alias": "record_count"
1779
+ }
1780
+ ],
1781
+ "filters": [
1782
+ {
1783
+ "column_id": "c_section",
1784
+ "op": "=",
1785
+ "value": "OB LOADER",
1786
+ "value_type": "string"
1787
+ }
1788
+ ]
1789
+ }
1790
+ ],
1791
+ "agg_args": [],
1792
+ "infeasible": false
1793
+ },
1794
+ "error": null,
1795
+ "latency_ms": 2378,
1796
+ "tokens": 14570
1797
+ },
1798
+ {
1799
+ "id": "disambig_trend_pa_id",
1800
+ "category": "column_disambiguation",
1801
+ "lang": "id",
1802
+ "carried_over": false,
1803
+ "question": "bagaimana trend PA?",
1804
+ "passed": true,
1805
+ "checks": [
1806
+ {
1807
+ "check": "selects_col=PA_Percent",
1808
+ "ok": true,
1809
+ "detail": "selected=['c_from_date', 'c_pa_percent']"
1810
+ },
1811
+ {
1812
+ "check": "not_selects_col=Plan_PA_Percent",
1813
+ "ok": true,
1814
+ "detail": "selected=['c_from_date', 'c_pa_percent']"
1815
+ }
1816
+ ],
1817
+ "facts": {
1818
+ "tools": [
1819
+ "analyze_trend",
1820
+ "check_data",
1821
+ "retrieve_data"
1822
+ ],
1823
+ "irs": [
1824
+ {
1825
+ "source_id": "src_pa",
1826
+ "table_id": "t_pa",
1827
+ "select": [
1828
+ {
1829
+ "kind": "column",
1830
+ "column_id": "c_from_date",
1831
+ "alias": "date"
1832
+ },
1833
+ {
1834
+ "kind": "column",
1835
+ "column_id": "c_pa_percent",
1836
+ "alias": "pa_percent"
1837
+ }
1838
+ ],
1839
+ "filters": [
1840
+ {
1841
+ "column_id": "c_month_id",
1842
+ "op": "=",
1843
+ "value": 202604,
1844
+ "value_type": "int"
1845
+ },
1846
+ {
1847
+ "column_id": "c_isdeleted",
1848
+ "op": "=",
1849
+ "value": 0,
1850
+ "value_type": "int"
1851
+ }
1852
+ ],
1853
+ "limit": 10000
1854
+ }
1855
+ ],
1856
+ "agg_args": [],
1857
+ "infeasible": false
1858
+ },
1859
+ "error": null,
1860
+ "latency_ms": 3527,
1861
+ "tokens": 14860
1862
+ },
1863
+ {
1864
+ "id": "disambig_avg_pa_en",
1865
+ "category": "column_disambiguation",
1866
+ "lang": "en",
1867
+ "carried_over": false,
1868
+ "question": "what is the average PA per section?",
1869
+ "passed": true,
1870
+ "checks": [
1871
+ {
1872
+ "check": "selects_col=PA_Percent",
1873
+ "ok": true,
1874
+ "detail": "selected=['c_section', 'c_pa_percent']"
1875
+ },
1876
+ {
1877
+ "check": "not_selects_col=Plan_PA_Percent",
1878
+ "ok": true,
1879
+ "detail": "selected=['c_section', 'c_pa_percent']"
1880
+ }
1881
+ ],
1882
+ "facts": {
1883
+ "tools": [
1884
+ "analyze_aggregate",
1885
+ "check_data",
1886
+ "retrieve_data"
1887
+ ],
1888
+ "irs": [
1889
+ {
1890
+ "source_id": "src_pa",
1891
+ "table_id": "t_pa",
1892
+ "select": [
1893
+ {
1894
+ "kind": "column",
1895
+ "column_id": "c_section",
1896
+ "alias": "section"
1897
+ },
1898
+ {
1899
+ "kind": "column",
1900
+ "column_id": "c_pa_percent",
1901
+ "alias": "pa_percent"
1902
+ }
1903
+ ],
1904
+ "filters": [
1905
+ {
1906
+ "column_id": "c_month_id",
1907
+ "op": "=",
1908
+ "value": 202604,
1909
+ "value_type": "int"
1910
+ },
1911
+ {
1912
+ "column_id": "c_isdeleted",
1913
+ "op": "=",
1914
+ "value": 0,
1915
+ "value_type": "int"
1916
+ }
1917
+ ],
1918
+ "limit": 10000
1919
+ }
1920
+ ],
1921
+ "agg_args": [
1922
+ {
1923
+ "data": "${t2}",
1924
+ "aggregations": {
1925
+ "pa_percent": [
1926
+ "mean"
1927
+ ]
1928
+ },
1929
+ "group_by": [
1930
+ "section"
1931
+ ]
1932
+ }
1933
+ ],
1934
+ "infeasible": false
1935
+ },
1936
+ "error": null,
1937
+ "latency_ms": 3972,
1938
+ "tokens": 14810
1939
+ },
1940
+ {
1941
+ "id": "chart_trend_pa_id",
1942
+ "category": "chart",
1943
+ "lang": "id",
1944
+ "carried_over": false,
1945
+ "question": "bagaimana visualisasi trend PA?",
1946
+ "passed": false,
1947
+ "checks": [
1948
+ {
1949
+ "check": "has_tool=render_chart",
1950
+ "ok": true,
1951
+ "detail": "tools=['check_data', 'render_chart', 'retrieve_data']"
1952
+ },
1953
+ {
1954
+ "check": "any_tool=['analyze_aggregate', 'analyze_trend']",
1955
+ "ok": false,
1956
+ "detail": "tools=['check_data', 'render_chart', 'retrieve_data']"
1957
+ },
1958
+ {
1959
+ "check": "selects_col=PA_Percent",
1960
+ "ok": true,
1961
+ "detail": "selected=['c_from_date', 'c_pa_percent']"
1962
+ },
1963
+ {
1964
+ "check": "not_selects_col=Plan_PA_Percent",
1965
+ "ok": true,
1966
+ "detail": "selected=['c_from_date', 'c_pa_percent']"
1967
+ }
1968
+ ],
1969
+ "facts": {
1970
+ "tools": [
1971
+ "check_data",
1972
+ "render_chart",
1973
+ "retrieve_data"
1974
+ ],
1975
+ "irs": [
1976
+ {
1977
+ "source_id": "src_pa",
1978
+ "table_id": "t_pa",
1979
+ "select": [
1980
+ {
1981
+ "kind": "column",
1982
+ "column_id": "c_from_date",
1983
+ "alias": "from_date"
1984
+ },
1985
+ {
1986
+ "kind": "column",
1987
+ "column_id": "c_pa_percent",
1988
+ "alias": "pa_percent"
1989
+ }
1990
+ ],
1991
+ "filters": [
1992
+ {
1993
+ "column_id": "c_month_id",
1994
+ "op": "=",
1995
+ "value": 202604,
1996
+ "value_type": "int"
1997
+ },
1998
+ {
1999
+ "column_id": "c_isdeleted",
2000
+ "op": "=",
2001
+ "value": 0,
2002
+ "value_type": "int"
2003
+ }
2004
+ ],
2005
+ "limit": 10000
2006
+ }
2007
+ ],
2008
+ "agg_args": [],
2009
+ "infeasible": false
2010
+ },
2011
+ "error": null,
2012
+ "latency_ms": 3641,
2013
+ "tokens": 14859
2014
+ },
2015
+ {
2016
+ "id": "chart_avg_pa_by_section_en",
2017
+ "category": "chart",
2018
+ "lang": "en",
2019
+ "carried_over": false,
2020
+ "question": "show me a bar chart of average PA per section",
2021
+ "passed": true,
2022
+ "checks": [
2023
+ {
2024
+ "check": "has_tool=render_chart",
2025
+ "ok": true,
2026
+ "detail": "tools=['check_data', 'render_chart', 'retrieve_data']"
2027
+ },
2028
+ {
2029
+ "check": "group_by",
2030
+ "ok": true,
2031
+ "detail": "grouped=True"
2032
+ },
2033
+ {
2034
+ "check": "selects_col=PA_Percent",
2035
+ "ok": true,
2036
+ "detail": "selected=['c_section', 'c_pa_percent']"
2037
+ }
2038
+ ],
2039
+ "facts": {
2040
+ "tools": [
2041
+ "check_data",
2042
+ "render_chart",
2043
+ "retrieve_data"
2044
+ ],
2045
+ "irs": [
2046
+ {
2047
+ "source_id": "src_pa",
2048
+ "table_id": "t_pa",
2049
+ "select": [
2050
+ {
2051
+ "kind": "column",
2052
+ "column_id": "c_section",
2053
+ "alias": "section"
2054
+ },
2055
+ {
2056
+ "kind": "agg",
2057
+ "fn": "avg",
2058
+ "column_id": "c_pa_percent",
2059
+ "alias": "avg_pa"
2060
+ }
2061
+ ],
2062
+ "group_by": [
2063
+ "c_section"
2064
+ ]
2065
+ }
2066
+ ],
2067
+ "agg_args": [],
2068
+ "infeasible": false
2069
+ },
2070
+ "error": null,
2071
+ "latency_ms": 3166,
2072
+ "tokens": 14769
2073
+ },
2074
+ {
2075
+ "id": "infeasible_churn_id",
2076
+ "category": "infeasible",
2077
+ "lang": "id",
2078
+ "carried_over": true,
2079
+ "question": "berapa churn rate pelanggan?",
2080
+ "passed": true,
2081
+ "checks": [
2082
+ {
2083
+ "check": "infeasible",
2084
+ "ok": true,
2085
+ "detail": "infeasible=True"
2086
+ }
2087
+ ],
2088
+ "facts": {
2089
+ "tools": [],
2090
+ "irs": [],
2091
+ "agg_args": [],
2092
+ "infeasible": true
2093
+ },
2094
+ "error": null,
2095
+ "latency_ms": 2055,
2096
+ "tokens": 14486
2097
+ },
2098
+ {
2099
+ "id": "infeasible_profit_en",
2100
+ "category": "infeasible",
2101
+ "lang": "en",
2102
+ "carried_over": true,
2103
+ "question": "what is the monthly profit margin?",
2104
+ "passed": true,
2105
+ "checks": [
2106
+ {
2107
+ "check": "infeasible",
2108
+ "ok": true,
2109
+ "detail": "infeasible=True"
2110
+ }
2111
+ ],
2112
+ "facts": {
2113
+ "tools": [],
2114
+ "irs": [],
2115
+ "agg_args": [],
2116
+ "infeasible": true
2117
+ },
2118
+ "error": null,
2119
+ "latency_ms": 2072,
2120
+ "tokens": 14501
2121
+ }
2122
+ ]
2123
+ }
eval/readiness/results/readiness_result_2026-06-23_094622.json ADDED
@@ -0,0 +1,284 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "run": {
3
+ "timestamp": "2026-06-23T09:46:22",
4
+ "dataset": "readiness_dataset.json",
5
+ "target": "src/agents/report/readiness.is_report_ready",
6
+ "total": 17,
7
+ "passed": 17,
8
+ "accuracy": 1.0,
9
+ "runtime_avg_ms": 0.0
10
+ },
11
+ "alignment_gap": {
12
+ "count": 2,
13
+ "ids": [
14
+ "align_01",
15
+ "align_02"
16
+ ]
17
+ },
18
+ "by_group": {
19
+ "floor": {
20
+ "n": 8,
21
+ "passed": 8,
22
+ "accuracy": 1.0
23
+ },
24
+ "delta": {
25
+ "n": 5,
26
+ "passed": 5,
27
+ "accuracy": 1.0
28
+ },
29
+ "edge": {
30
+ "n": 1,
31
+ "passed": 1,
32
+ "accuracy": 1.0
33
+ },
34
+ "alignment": {
35
+ "n": 3,
36
+ "passed": 3,
37
+ "accuracy": 1.0
38
+ }
39
+ },
40
+ "cases": [
41
+ {
42
+ "id": "floor_01",
43
+ "group": "floor",
44
+ "expected_ready": false,
45
+ "got_ready": false,
46
+ "expected_missing": [
47
+ "a validated problem statement",
48
+ "at least one completed analysis"
49
+ ],
50
+ "got_missing": [
51
+ "a validated problem statement",
52
+ "at least one completed analysis"
53
+ ],
54
+ "correct": true,
55
+ "aligned": false,
56
+ "gap": false,
57
+ "latency_ms": 0.0
58
+ },
59
+ {
60
+ "id": "floor_02",
61
+ "group": "floor",
62
+ "expected_ready": false,
63
+ "got_ready": false,
64
+ "expected_missing": [
65
+ "a validated problem statement"
66
+ ],
67
+ "got_missing": [
68
+ "a validated problem statement"
69
+ ],
70
+ "correct": true,
71
+ "aligned": true,
72
+ "gap": false,
73
+ "latency_ms": 0.0
74
+ },
75
+ {
76
+ "id": "floor_03",
77
+ "group": "floor",
78
+ "expected_ready": false,
79
+ "got_ready": false,
80
+ "expected_missing": [
81
+ "at least one completed analysis"
82
+ ],
83
+ "got_missing": [
84
+ "at least one completed analysis"
85
+ ],
86
+ "correct": true,
87
+ "aligned": false,
88
+ "gap": false,
89
+ "latency_ms": 0.0
90
+ },
91
+ {
92
+ "id": "floor_04",
93
+ "group": "floor",
94
+ "expected_ready": false,
95
+ "got_ready": false,
96
+ "expected_missing": [
97
+ "at least one completed analysis"
98
+ ],
99
+ "got_missing": [
100
+ "at least one completed analysis"
101
+ ],
102
+ "correct": true,
103
+ "aligned": false,
104
+ "gap": false,
105
+ "latency_ms": 0.0
106
+ },
107
+ {
108
+ "id": "floor_05",
109
+ "group": "floor",
110
+ "expected_ready": false,
111
+ "got_ready": false,
112
+ "expected_missing": [
113
+ "at least one completed analysis"
114
+ ],
115
+ "got_missing": [
116
+ "at least one completed analysis"
117
+ ],
118
+ "correct": true,
119
+ "aligned": false,
120
+ "gap": false,
121
+ "latency_ms": 0.0
122
+ },
123
+ {
124
+ "id": "floor_06",
125
+ "group": "floor",
126
+ "expected_ready": true,
127
+ "got_ready": true,
128
+ "expected_missing": [],
129
+ "got_missing": [],
130
+ "correct": true,
131
+ "aligned": true,
132
+ "gap": false,
133
+ "latency_ms": 0.0
134
+ },
135
+ {
136
+ "id": "floor_07",
137
+ "group": "floor",
138
+ "expected_ready": true,
139
+ "got_ready": true,
140
+ "expected_missing": [],
141
+ "got_missing": [],
142
+ "correct": true,
143
+ "aligned": true,
144
+ "gap": false,
145
+ "latency_ms": 0.0
146
+ },
147
+ {
148
+ "id": "floor_08",
149
+ "group": "floor",
150
+ "expected_ready": true,
151
+ "got_ready": true,
152
+ "expected_missing": [],
153
+ "got_missing": [],
154
+ "correct": true,
155
+ "aligned": true,
156
+ "gap": false,
157
+ "latency_ms": 0.0
158
+ },
159
+ {
160
+ "id": "delta_01",
161
+ "group": "delta",
162
+ "expected_ready": false,
163
+ "got_ready": false,
164
+ "expected_missing": [
165
+ "a new analysis since the last report"
166
+ ],
167
+ "got_missing": [
168
+ "a new analysis since the last report"
169
+ ],
170
+ "correct": true,
171
+ "aligned": true,
172
+ "gap": false,
173
+ "latency_ms": 0.0
174
+ },
175
+ {
176
+ "id": "delta_02",
177
+ "group": "delta",
178
+ "expected_ready": true,
179
+ "got_ready": true,
180
+ "expected_missing": [],
181
+ "got_missing": [],
182
+ "correct": true,
183
+ "aligned": true,
184
+ "gap": false,
185
+ "latency_ms": 0.0
186
+ },
187
+ {
188
+ "id": "delta_03",
189
+ "group": "delta",
190
+ "expected_ready": true,
191
+ "got_ready": true,
192
+ "expected_missing": [],
193
+ "got_missing": [],
194
+ "correct": true,
195
+ "aligned": true,
196
+ "gap": false,
197
+ "latency_ms": 0.0
198
+ },
199
+ {
200
+ "id": "delta_04",
201
+ "group": "delta",
202
+ "expected_ready": false,
203
+ "got_ready": false,
204
+ "expected_missing": [
205
+ "a new analysis since the last report"
206
+ ],
207
+ "got_missing": [
208
+ "a new analysis since the last report"
209
+ ],
210
+ "correct": true,
211
+ "aligned": true,
212
+ "gap": false,
213
+ "latency_ms": 0.0
214
+ },
215
+ {
216
+ "id": "delta_05",
217
+ "group": "delta",
218
+ "expected_ready": false,
219
+ "got_ready": false,
220
+ "expected_missing": [
221
+ "a new analysis since the last report"
222
+ ],
223
+ "got_missing": [
224
+ "a new analysis since the last report"
225
+ ],
226
+ "correct": true,
227
+ "aligned": true,
228
+ "gap": false,
229
+ "latency_ms": 0.0
230
+ },
231
+ {
232
+ "id": "edge_01",
233
+ "group": "edge",
234
+ "expected_ready": false,
235
+ "got_ready": false,
236
+ "expected_missing": [
237
+ "at least one completed analysis"
238
+ ],
239
+ "got_missing": [
240
+ "at least one completed analysis"
241
+ ],
242
+ "correct": true,
243
+ "aligned": false,
244
+ "gap": false,
245
+ "latency_ms": 0.0
246
+ },
247
+ {
248
+ "id": "align_01",
249
+ "group": "alignment",
250
+ "expected_ready": true,
251
+ "got_ready": true,
252
+ "expected_missing": [],
253
+ "got_missing": [],
254
+ "correct": true,
255
+ "aligned": false,
256
+ "gap": true,
257
+ "latency_ms": 0.0
258
+ },
259
+ {
260
+ "id": "align_02",
261
+ "group": "alignment",
262
+ "expected_ready": true,
263
+ "got_ready": true,
264
+ "expected_missing": [],
265
+ "got_missing": [],
266
+ "correct": true,
267
+ "aligned": false,
268
+ "gap": true,
269
+ "latency_ms": 0.0
270
+ },
271
+ {
272
+ "id": "align_03",
273
+ "group": "alignment",
274
+ "expected_ready": true,
275
+ "got_ready": true,
276
+ "expected_missing": [],
277
+ "got_missing": [],
278
+ "correct": true,
279
+ "aligned": true,
280
+ "gap": false,
281
+ "latency_ms": 0.0
282
+ }
283
+ ]
284
+ }
eval/readiness/results/readiness_result_2026-07-02_083335.json ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "run": {
3
+ "timestamp": "2026-07-02T08:33:35",
4
+ "dataset": "readiness_dataset.json",
5
+ "target": "src/agents/report/readiness.is_report_ready",
6
+ "total": 15,
7
+ "passed": 15,
8
+ "accuracy": 1.0,
9
+ "runtime_avg_ms": 0.0
10
+ },
11
+ "alignment_gap": {
12
+ "count": 2,
13
+ "ids": [
14
+ "align_01",
15
+ "align_02"
16
+ ]
17
+ },
18
+ "by_group": {
19
+ "floor": {
20
+ "n": 6,
21
+ "passed": 6,
22
+ "accuracy": 1.0
23
+ },
24
+ "delta": {
25
+ "n": 5,
26
+ "passed": 5,
27
+ "accuracy": 1.0
28
+ },
29
+ "edge": {
30
+ "n": 1,
31
+ "passed": 1,
32
+ "accuracy": 1.0
33
+ },
34
+ "alignment": {
35
+ "n": 3,
36
+ "passed": 3,
37
+ "accuracy": 1.0
38
+ }
39
+ },
40
+ "cases": [
41
+ {
42
+ "id": "floor_01",
43
+ "group": "floor",
44
+ "expected_ready": false,
45
+ "got_ready": false,
46
+ "expected_missing": [
47
+ "at least one completed analysis"
48
+ ],
49
+ "got_missing": [
50
+ "at least one completed analysis"
51
+ ],
52
+ "correct": true,
53
+ "aligned": false,
54
+ "gap": false,
55
+ "latency_ms": 0.0
56
+ },
57
+ {
58
+ "id": "floor_02",
59
+ "group": "floor",
60
+ "expected_ready": false,
61
+ "got_ready": false,
62
+ "expected_missing": [
63
+ "at least one completed analysis"
64
+ ],
65
+ "got_missing": [
66
+ "at least one completed analysis"
67
+ ],
68
+ "correct": true,
69
+ "aligned": false,
70
+ "gap": false,
71
+ "latency_ms": 0.0
72
+ },
73
+ {
74
+ "id": "floor_03",
75
+ "group": "floor",
76
+ "expected_ready": false,
77
+ "got_ready": false,
78
+ "expected_missing": [
79
+ "at least one completed analysis"
80
+ ],
81
+ "got_missing": [
82
+ "at least one completed analysis"
83
+ ],
84
+ "correct": true,
85
+ "aligned": false,
86
+ "gap": false,
87
+ "latency_ms": 0.0
88
+ },
89
+ {
90
+ "id": "floor_04",
91
+ "group": "floor",
92
+ "expected_ready": true,
93
+ "got_ready": true,
94
+ "expected_missing": [],
95
+ "got_missing": [],
96
+ "correct": true,
97
+ "aligned": true,
98
+ "gap": false,
99
+ "latency_ms": 0.0
100
+ },
101
+ {
102
+ "id": "floor_05",
103
+ "group": "floor",
104
+ "expected_ready": true,
105
+ "got_ready": true,
106
+ "expected_missing": [],
107
+ "got_missing": [],
108
+ "correct": true,
109
+ "aligned": true,
110
+ "gap": false,
111
+ "latency_ms": 0.0
112
+ },
113
+ {
114
+ "id": "floor_06",
115
+ "group": "floor",
116
+ "expected_ready": true,
117
+ "got_ready": true,
118
+ "expected_missing": [],
119
+ "got_missing": [],
120
+ "correct": true,
121
+ "aligned": true,
122
+ "gap": false,
123
+ "latency_ms": 0.0
124
+ },
125
+ {
126
+ "id": "delta_01",
127
+ "group": "delta",
128
+ "expected_ready": false,
129
+ "got_ready": false,
130
+ "expected_missing": [
131
+ "a new analysis since the last report"
132
+ ],
133
+ "got_missing": [
134
+ "a new analysis since the last report"
135
+ ],
136
+ "correct": true,
137
+ "aligned": true,
138
+ "gap": false,
139
+ "latency_ms": 0.0
140
+ },
141
+ {
142
+ "id": "delta_02",
143
+ "group": "delta",
144
+ "expected_ready": true,
145
+ "got_ready": true,
146
+ "expected_missing": [],
147
+ "got_missing": [],
148
+ "correct": true,
149
+ "aligned": true,
150
+ "gap": false,
151
+ "latency_ms": 0.0
152
+ },
153
+ {
154
+ "id": "delta_03",
155
+ "group": "delta",
156
+ "expected_ready": true,
157
+ "got_ready": true,
158
+ "expected_missing": [],
159
+ "got_missing": [],
160
+ "correct": true,
161
+ "aligned": true,
162
+ "gap": false,
163
+ "latency_ms": 0.0
164
+ },
165
+ {
166
+ "id": "delta_04",
167
+ "group": "delta",
168
+ "expected_ready": false,
169
+ "got_ready": false,
170
+ "expected_missing": [
171
+ "a new analysis since the last report"
172
+ ],
173
+ "got_missing": [
174
+ "a new analysis since the last report"
175
+ ],
176
+ "correct": true,
177
+ "aligned": true,
178
+ "gap": false,
179
+ "latency_ms": 0.0
180
+ },
181
+ {
182
+ "id": "delta_05",
183
+ "group": "delta",
184
+ "expected_ready": false,
185
+ "got_ready": false,
186
+ "expected_missing": [
187
+ "a new analysis since the last report"
188
+ ],
189
+ "got_missing": [
190
+ "a new analysis since the last report"
191
+ ],
192
+ "correct": true,
193
+ "aligned": true,
194
+ "gap": false,
195
+ "latency_ms": 0.0
196
+ },
197
+ {
198
+ "id": "edge_01",
199
+ "group": "edge",
200
+ "expected_ready": false,
201
+ "got_ready": false,
202
+ "expected_missing": [
203
+ "at least one completed analysis"
204
+ ],
205
+ "got_missing": [
206
+ "at least one completed analysis"
207
+ ],
208
+ "correct": true,
209
+ "aligned": false,
210
+ "gap": false,
211
+ "latency_ms": 0.0
212
+ },
213
+ {
214
+ "id": "align_01",
215
+ "group": "alignment",
216
+ "expected_ready": true,
217
+ "got_ready": true,
218
+ "expected_missing": [],
219
+ "got_missing": [],
220
+ "correct": true,
221
+ "aligned": false,
222
+ "gap": true,
223
+ "latency_ms": 0.0
224
+ },
225
+ {
226
+ "id": "align_02",
227
+ "group": "alignment",
228
+ "expected_ready": true,
229
+ "got_ready": true,
230
+ "expected_missing": [],
231
+ "got_missing": [],
232
+ "correct": true,
233
+ "aligned": false,
234
+ "gap": true,
235
+ "latency_ms": 0.0
236
+ },
237
+ {
238
+ "id": "align_03",
239
+ "group": "alignment",
240
+ "expected_ready": true,
241
+ "got_ready": true,
242
+ "expected_missing": [],
243
+ "got_missing": [],
244
+ "correct": true,
245
+ "aligned": true,
246
+ "gap": false,
247
+ "latency_ms": 0.0
248
+ }
249
+ ]
250
+ }
eval/readiness/results/readiness_result_2026-07-02_093621.json ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "run": {
3
+ "timestamp": "2026-07-02T09:36:21",
4
+ "dataset": "readiness_dataset.json",
5
+ "target": "src/agents/report/readiness.is_report_ready",
6
+ "total": 15,
7
+ "passed": 15,
8
+ "accuracy": 1.0,
9
+ "runtime_avg_ms": 0.0
10
+ },
11
+ "alignment_gap": {
12
+ "count": 2,
13
+ "ids": [
14
+ "align_01",
15
+ "align_02"
16
+ ]
17
+ },
18
+ "by_group": {
19
+ "floor": {
20
+ "n": 6,
21
+ "passed": 6,
22
+ "accuracy": 1.0
23
+ },
24
+ "delta": {
25
+ "n": 5,
26
+ "passed": 5,
27
+ "accuracy": 1.0
28
+ },
29
+ "edge": {
30
+ "n": 1,
31
+ "passed": 1,
32
+ "accuracy": 1.0
33
+ },
34
+ "alignment": {
35
+ "n": 3,
36
+ "passed": 3,
37
+ "accuracy": 1.0
38
+ }
39
+ },
40
+ "cases": [
41
+ {
42
+ "id": "floor_01",
43
+ "group": "floor",
44
+ "expected_ready": false,
45
+ "got_ready": false,
46
+ "expected_missing": [
47
+ "at least one completed analysis"
48
+ ],
49
+ "got_missing": [
50
+ "at least one completed analysis"
51
+ ],
52
+ "correct": true,
53
+ "aligned": false,
54
+ "gap": false,
55
+ "latency_ms": 0.0
56
+ },
57
+ {
58
+ "id": "floor_02",
59
+ "group": "floor",
60
+ "expected_ready": false,
61
+ "got_ready": false,
62
+ "expected_missing": [
63
+ "at least one completed analysis"
64
+ ],
65
+ "got_missing": [
66
+ "at least one completed analysis"
67
+ ],
68
+ "correct": true,
69
+ "aligned": false,
70
+ "gap": false,
71
+ "latency_ms": 0.0
72
+ },
73
+ {
74
+ "id": "floor_03",
75
+ "group": "floor",
76
+ "expected_ready": false,
77
+ "got_ready": false,
78
+ "expected_missing": [
79
+ "at least one completed analysis"
80
+ ],
81
+ "got_missing": [
82
+ "at least one completed analysis"
83
+ ],
84
+ "correct": true,
85
+ "aligned": false,
86
+ "gap": false,
87
+ "latency_ms": 0.0
88
+ },
89
+ {
90
+ "id": "floor_04",
91
+ "group": "floor",
92
+ "expected_ready": true,
93
+ "got_ready": true,
94
+ "expected_missing": [],
95
+ "got_missing": [],
96
+ "correct": true,
97
+ "aligned": true,
98
+ "gap": false,
99
+ "latency_ms": 0.0
100
+ },
101
+ {
102
+ "id": "floor_05",
103
+ "group": "floor",
104
+ "expected_ready": true,
105
+ "got_ready": true,
106
+ "expected_missing": [],
107
+ "got_missing": [],
108
+ "correct": true,
109
+ "aligned": true,
110
+ "gap": false,
111
+ "latency_ms": 0.0
112
+ },
113
+ {
114
+ "id": "floor_06",
115
+ "group": "floor",
116
+ "expected_ready": true,
117
+ "got_ready": true,
118
+ "expected_missing": [],
119
+ "got_missing": [],
120
+ "correct": true,
121
+ "aligned": true,
122
+ "gap": false,
123
+ "latency_ms": 0.0
124
+ },
125
+ {
126
+ "id": "delta_01",
127
+ "group": "delta",
128
+ "expected_ready": false,
129
+ "got_ready": false,
130
+ "expected_missing": [
131
+ "a new analysis since the last report"
132
+ ],
133
+ "got_missing": [
134
+ "a new analysis since the last report"
135
+ ],
136
+ "correct": true,
137
+ "aligned": true,
138
+ "gap": false,
139
+ "latency_ms": 0.0
140
+ },
141
+ {
142
+ "id": "delta_02",
143
+ "group": "delta",
144
+ "expected_ready": true,
145
+ "got_ready": true,
146
+ "expected_missing": [],
147
+ "got_missing": [],
148
+ "correct": true,
149
+ "aligned": true,
150
+ "gap": false,
151
+ "latency_ms": 0.0
152
+ },
153
+ {
154
+ "id": "delta_03",
155
+ "group": "delta",
156
+ "expected_ready": true,
157
+ "got_ready": true,
158
+ "expected_missing": [],
159
+ "got_missing": [],
160
+ "correct": true,
161
+ "aligned": true,
162
+ "gap": false,
163
+ "latency_ms": 0.0
164
+ },
165
+ {
166
+ "id": "delta_04",
167
+ "group": "delta",
168
+ "expected_ready": false,
169
+ "got_ready": false,
170
+ "expected_missing": [
171
+ "a new analysis since the last report"
172
+ ],
173
+ "got_missing": [
174
+ "a new analysis since the last report"
175
+ ],
176
+ "correct": true,
177
+ "aligned": true,
178
+ "gap": false,
179
+ "latency_ms": 0.0
180
+ },
181
+ {
182
+ "id": "delta_05",
183
+ "group": "delta",
184
+ "expected_ready": false,
185
+ "got_ready": false,
186
+ "expected_missing": [
187
+ "a new analysis since the last report"
188
+ ],
189
+ "got_missing": [
190
+ "a new analysis since the last report"
191
+ ],
192
+ "correct": true,
193
+ "aligned": true,
194
+ "gap": false,
195
+ "latency_ms": 0.0
196
+ },
197
+ {
198
+ "id": "edge_01",
199
+ "group": "edge",
200
+ "expected_ready": false,
201
+ "got_ready": false,
202
+ "expected_missing": [
203
+ "at least one completed analysis"
204
+ ],
205
+ "got_missing": [
206
+ "at least one completed analysis"
207
+ ],
208
+ "correct": true,
209
+ "aligned": false,
210
+ "gap": false,
211
+ "latency_ms": 0.0
212
+ },
213
+ {
214
+ "id": "align_01",
215
+ "group": "alignment",
216
+ "expected_ready": true,
217
+ "got_ready": true,
218
+ "expected_missing": [],
219
+ "got_missing": [],
220
+ "correct": true,
221
+ "aligned": false,
222
+ "gap": true,
223
+ "latency_ms": 0.0
224
+ },
225
+ {
226
+ "id": "align_02",
227
+ "group": "alignment",
228
+ "expected_ready": true,
229
+ "got_ready": true,
230
+ "expected_missing": [],
231
+ "got_missing": [],
232
+ "correct": true,
233
+ "aligned": false,
234
+ "gap": true,
235
+ "latency_ms": 0.0
236
+ },
237
+ {
238
+ "id": "align_03",
239
+ "group": "alignment",
240
+ "expected_ready": true,
241
+ "got_ready": true,
242
+ "expected_missing": [],
243
+ "got_missing": [],
244
+ "correct": true,
245
+ "aligned": true,
246
+ "gap": false,
247
+ "latency_ms": 0.0
248
+ }
249
+ ]
250
+ }