File size: 45,222 Bytes
c427f62
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
"""Session Interpretation generator (Task 2).

Generates multi-session behavioral logs and matching structured interpretations.
The log is produced deterministically from a sampled hidden pattern so the
interpretation's gold labels are known by construction.
"""

import random
from datetime import date, timedelta
from pathlib import Path

from .base import (
    load_area,
    load_shared,
    load_yaml,
    strip_trailing_period,
    make_example_envelope,
)

AREA = "session_interpretation"
TASK_TYPE = "session_interpretation"


# Trajectory math — generate per-session accuracy values following a pattern rule


def _clip(x: float, lo: float = 0.0, hi: float = 1.0) -> float:
    return max(lo, min(hi, x))


def generate_accuracy_trajectory(pattern_id: str, rule: dict, n_sessions: int, rng: random.Random) -> list:
    """Return a list of accuracy values for n_sessions based on pattern rule."""
    acc = rule["accuracy"]

    if pattern_id == "generalization_failure":
        alt = acc["alternating"]
        tr_lo, tr_hi = alt["training_range"]
        nov_lo, nov_hi = alt["novel_range"]
        noise = alt["noise"]
        vals = []
        training = rng.uniform(tr_lo, tr_hi)
        novel = rng.uniform(nov_lo, nov_hi)
        for i in range(n_sessions):
            base = training if i % 2 == 0 else novel
            vals.append(_clip(base + rng.uniform(-noise, noise)))
        return vals

    if pattern_id == "skill_loss_after_break":
        base = rng.uniform(*acc["base_range"])
        drop = rng.uniform(*acc["initial_drop"])
        rec = rng.uniform(*acc["recovery_per_session_range"])
        noise = acc["noise"]
        vals = [_clip(base - drop + rng.uniform(-noise, noise))]
        for _ in range(1, n_sessions):
            vals.append(_clip(vals[-1] + rec + rng.uniform(-noise, noise), 0.1, base))
        return vals

    if pattern_id == "motivating_operation_shift":
        base = rng.uniform(*acc["base_range"])
        dip_len = rng.randint(*acc["mo_dip_sessions"])
        dip_mag = rng.uniform(*acc["mo_dip_magnitude"])
        rec = rng.uniform(*acc["recovery_per_session_range"])
        noise = acc["noise"]
        dip_start = max(2, n_sessions // 3)
        vals = []
        for i in range(n_sessions):
            in_dip = dip_start <= i < dip_start + dip_len
            if in_dip:
                v = base - dip_mag + rng.uniform(-noise, noise)
            elif i >= dip_start + dip_len:
                prev = vals[-1]
                v = min(base, prev + rec + rng.uniform(-noise, noise))
            else:
                v = base + rng.uniform(-noise, noise)
            vals.append(_clip(v, 0.1, 0.98))
        return vals

    if pattern_id == "setting_event_trigger":
        base = rng.uniform(*acc["base_range"])
        se_idx = max(1, int(acc["setting_event_session_fraction"] * n_sessions))
        mag = rng.uniform(*acc["setting_event_magnitude"])
        rec = rng.uniform(*acc["recovery_per_session_range"])
        noise = acc["noise"]
        vals = []
        for i in range(n_sessions):
            if i < se_idx:
                v = base + rng.uniform(-noise, noise)
            elif i == se_idx:
                v = base - mag + rng.uniform(-noise, noise)
            else:
                prev = vals[-1]
                v = min(base, prev + rec + rng.uniform(-noise, noise))
            vals.append(_clip(v, 0.1, 0.98))
        return vals

    if pattern_id == "extinction_burst":
        base = rng.uniform(*acc["base_range"])
        delta = rng.uniform(*acc["delta_per_session_range"])
        noise = acc["noise"]
        vals = [_clip(base + i * delta + rng.uniform(-noise, noise)) for i in range(n_sessions)]
        return vals

    # Default: linear delta from base
    base = rng.uniform(*acc["base_range"])
    delta = rng.uniform(*acc["delta_per_session_range"])
    noise = acc["noise"]
    cap = acc.get("cap", 0.98)
    floor_v = acc.get("floor", 0.05)
    return [_clip(base + i * delta + rng.uniform(-noise, noise), floor_v, cap) for i in range(n_sessions)]


def generate_behavior_frequencies(
    traj_name: str, library: dict, n_sessions: int, rule: dict, rng: random.Random
) -> list:
    """Return a list of integer behavior frequencies per session."""
    cfg = library[traj_name]
    if traj_name == "temporary_spike":
        base_lo, base_hi = cfg["base_freq_range"]
        mult_range = cfg["spike_multiplier_range"]
        noise = cfg["noise"]
        burst_start = max(2, int(rule.get("burst_start_fraction", 0.33) * n_sessions))
        burst_len = rule.get("burst_duration_sessions", 2)
        base = rng.uniform(base_lo, base_hi)
        mult = rng.uniform(*mult_range)
        vals = []
        for i in range(n_sessions):
            if burst_start <= i < burst_start + burst_len:
                v = base * mult + rng.uniform(-noise, noise)
            else:
                v = base + rng.uniform(-noise, noise)
            vals.append(max(0, round(v)))
        return vals
    if traj_name == "coincident_spike":
        # Spike coincides with setting event index
        base_lo, base_hi = cfg["base_freq_range"]
        mult_range = cfg["spike_multiplier_range"]
        noise = cfg["noise"]
        se_idx = max(1, int(rule["accuracy"]["setting_event_session_fraction"] * n_sessions))
        base = rng.uniform(base_lo, base_hi)
        mult = rng.uniform(*mult_range)
        vals = []
        for i in range(n_sessions):
            if i in {se_idx, se_idx + 1}:
                v = base * mult + rng.uniform(-noise, noise)
            else:
                v = base + rng.uniform(-noise, noise)
            vals.append(max(0, round(v)))
        return vals
    # default: base_freq_range + cumulative delta + noise
    base_lo, base_hi = cfg["base_freq_range"]
    delta = cfg.get("delta_per_session", 0.0)
    noise = cfg.get("noise", 1.0)
    base = rng.uniform(base_lo, base_hi)
    return [max(0, round(base + i * delta + rng.uniform(-noise, noise))) for i in range(n_sessions)]


# Stimulus / profile sampling helpers


SKILL_POOL_BY_CATEGORY = {
    "early_language": [
        ("Mand", "single-word mands for preferred items (e.g., \"cookie\", \"iPad\")", "NET"),
        ("Mand", "mands for missing items needed to complete an activity", "NET"),
        ("Mand", "2-word mand phrases (e.g., \"more cracker\", \"open please\")", "NET"),
        ("Mand", "requesting a break ('break please' or AAC 'I need a break')", "NET"),
        ("Mand", "indicating completion ('all done', 'finished')", "NET"),
        ("Tact", "tacts common objects", "DTT"),
        ("Tact", "tacts colors of objects", "DTT"),
        ("Listener Responding", "follows 1-step motor instructions", "DTT"),
        ("Listener Responding", "selects items by feature", "DTT"),
        ("Echoic", "echoes 2-word phrases", "DTT variant"),
        ("Motor Imitation", "imitates 3-step motor sequences", "DTT variant"),
        ("Intraverbal", "answers WH-questions about familiar topics", "NET"),
        ("Tolerating Denied Access", "wait calmly when access to preferred item is denied", "FCT"),
    ],
    "school_age": [
        ("Reading", "reads 10–20 sight words", "DTT"),
        ("Reading", "sounds out CVC words", "DTT"),
        ("Writing", "copies simple words from model", "DTT"),
        ("Math", "identifies numerals 1–20", "DTT"),
        ("Math", "counts objects 1–20", "DTT"),
        ("Mand", "requesting preferred items using AAC device (e.g., \"I want cookie please\")", "NET"),
        ("Mand", "single-word requests for preferred items", "NET"),
        ("Mand", "requesting preferred items using a full sentence", "NET"),
        ("Mand", "requesting a break during non-preferred tasks", "NET"),
        ("Mand", "indicating 'all done' when an activity is complete", "NET"),
        ("Tolerating Denied Access", "wait calmly when access to preferred item is denied", "FCT"),
        ("Social Behavior", "takes turns during structured activities", "PRT"),
    ],
    "adolescent": [
        ("Self-Care", "washing hands independently", "Task Analysis"),
        ("Self-Care", "dressing (pants, shirt)", "Task Analysis"),
        ("Self-Care", "brushing teeth independently", "Task Analysis"),
        ("Self-Care", "tolerating tooth brushing by caregiver", "Task Analysis"),
        ("Self-Care", "tolerating haircut", "Task Analysis"),
        ("Home Skills", "making the bed", "Task Analysis"),
        ("Home Skills", "meal cleanup after eating", "Task Analysis"),
        ("Community", "identifying community safety signs", "DTT"),
        ("Community", "ordering in a restaurant", "Task Analysis"),
        ("Mand", "requesting preferred items using AAC device", "NET"),
        ("Mand", "requesting help when a task is difficult", "NET"),
        ("Mand", "requesting a break during non-preferred activities", "NET"),
        ("Mand", "indicating 'all done' with a step or activity", "NET"),
        ("Tolerating Denied Access", "wait calmly when access to preferred item is denied", "FCT"),
    ],
    "adult": [
        ("Independent Living", "scheduling medical appointments", "Task Analysis"),
        ("Independent Living", "managing a budget", "Task Analysis"),
        ("Independent Living", "self-advocacy statements", "BST"),
        ("Vocational", "clocking in at start of shift", "Task Analysis"),
        ("Vocational", "completing assigned tasks in order", "Task Analysis"),
        ("Home Skills", "doing laundry", "Task Analysis"),
        ("Home Skills", "simple meal preparation", "Task Analysis"),
        ("Community", "using public transportation", "Task Analysis"),
        ("Mand", "requesting assistance using AAC device", "NET"),
        ("Mand", "requesting a break during extended tasks", "NET"),
        ("Mand", "indicating 'all done' at task completion", "NET"),
        ("Tolerating Denied Access", "wait calmly when a preferred activity is denied", "FCT"),
    ],
}


def sample_programs(learner_profile_id: str, n: int, rng: random.Random) -> list:
    """Sample a realistic set of n acceleration programs for the learner profile."""
    category_map = {
        "early": "early_language",
        "school_age": "school_age",
        "adolescent": "adolescent",
        "adult": "adult",
    }
    primary = category_map.get(learner_profile_id, "school_age")
    pool = list(SKILL_POOL_BY_CATEGORY[primary])
    # Mix in one program from an adjacent category for realism
    adjacent = {"early": "school_age", "school_age": "early_language", "adolescent": "school_age", "adult": "adolescent"}
    pool += SKILL_POOL_BY_CATEGORY[adjacent.get(learner_profile_id, "school_age")][:3]
    rng.shuffle(pool)
    return pool[:n]


# Session log rendering


def _render_measurements_for_program(
    skill_domain: str, accuracy: float, rng: random.Random
) -> str:
    """Return a one-line measurement string for one program in one session."""
    trials = rng.randint(8, 15)
    correct = round(accuracy * trials)
    fp = rng.randint(0, max(0, trials // 3))
    pp = rng.randint(0, max(0, trials // 3))
    g = rng.randint(0, max(0, trials // 3))
    indep = max(0, trials - fp - pp - g)
    prompts = f"FP×{fp}, PP×{pp}, G×{g}, I×{indep}"
    latency = 1.5 + (1.0 - accuracy) * 4.5 + rng.uniform(-0.3, 0.3)
    return f"{correct}/{trials} correct ({round(accuracy*100)}%); latency {latency:.1f}s; prompts {prompts}"


def _render_behavior_measurement(behavior_id: str, freq: int, rng: random.Random) -> str:
    """Return a measurement string for one target behavior in one session."""
    if behavior_id in {"tantrum"}:
        if freq == 0:
            return f"freq 0"
        duration_total = freq * rng.randint(2, 5)
        return f"freq {freq}, duration {duration_total}m total"
    if behavior_id in {"motor_stereotypy", "vocal_stereotypy", "mouthing"}:
        pir_pct = min(60, freq * rng.randint(3, 5) + rng.randint(0, 10))
        return f"freq {freq}; PIR {pir_pct}%"
    if behavior_id == "pica":
        # Pica is measured as attempts (total) and successful ingestions (subset).
        # Most pica events are caught by staff before ingestion — successful ≤ attempts.
        if freq == 0:
            return "attempts 0"
        attempts = freq
        # Successful ingestion is stochastic but bounded; most are intercepted.
        successful = 0
        for _ in range(attempts):
            if rng.random() < 0.25:
                successful += 1
        unsuccessful = attempts - successful
        return (
            f"attempts {attempts} ({unsuccessful} unsuccessful — staff retrieved item before ingestion; "
            f"{successful} successful — item ingested)"
        )
    if behavior_id == "fecal_smearing":
        # Scatolia is measured as attempts (reaching for feces) and completed smearing (subset).
        # Staff often intercept before smearing completes — completed ≤ attempts.
        if freq == 0:
            return "attempts 0"
        attempts = freq
        completed = 0
        for _ in range(attempts):
            if rng.random() < 0.35:
                completed += 1
        intercepted = attempts - completed
        return (
            f"attempts {attempts} ({intercepted} intercepted — staff redirected before smearing; "
            f"{completed} completed — feces transferred to skin, clothing, or surface)"
        )
    if behavior_id == "toileting_accident":
        # Toileting data tracks both successful voids (in toilet) and accidents.
        # BM accidents scale with overall accident load but stay physiologically
        # bounded — typical session window is 0–1 BMs; some learners with GI
        # clustering (IBS, diet, medication-related) can reach 2.
        if freq == 0:
            bm_accidents = 0
        elif freq == 1:
            bm_accidents = 1 if rng.random() < 0.25 else 0
        elif freq <= 4:
            bm_accidents = rng.choices([0, 1], weights=[0.65, 0.35])[0]
        else:
            bm_accidents = rng.choices([0, 1, 2], weights=[0.50, 0.35, 0.15])[0]
        urine_accidents = max(0, freq - bm_accidents)

        # Successful voids scale inversely with accident load — a session full
        # of accidents represents missed toileting opportunities.
        if freq == 0:
            urine_success = rng.randint(3, 6)
        elif freq <= 2:
            urine_success = rng.randint(2, 5)
        elif freq <= 4:
            urine_success = rng.randint(1, 3)
        else:
            urine_success = rng.randint(0, 2)

        # BM successes: typically 0 per session window, sometimes 1. Learners
        # with BM accidents can still have an earlier successful BM in the toilet
        # (e.g., scheduled morning sit worked, later opportunities missed).
        if bm_accidents > 0:
            bm_success = 1 if rng.random() < 0.15 else 0
        else:
            bm_success = rng.choices([0, 1], weights=[0.75, 0.25])[0]

        return (
            f"urine: {urine_success} in-toilet / {urine_accidents} accidents; "
            f"BM: {bm_success} in-toilet / {bm_accidents} accidents"
        )
    if behavior_id in {"elopement"}:
        return f"freq {freq}"
    return f"freq {freq}"


ABC_TEMPLATES = {
    "escape": [
        ("worksheet task presented", "refused + pushed materials away", "task removed for 2 min"),
        ("demand to complete math problem", "threw materials and vocalized 'no'", "staff removed demand, redirected"),
        ("transition to non-preferred activity", "dropped to floor, cried", "transition delayed, preferred item offered"),
        ("asked to finish dressing task", "kicked at materials", "task paused for 3 min"),
    ],
    "attention": [
        ("adult attention diverted to another learner", "vocalized loudly + approached staff", "staff turned to address behavior"),
        ("staff engaged in paperwork nearby", "yelled name + tapped staff", "staff attended and redirected"),
    ],
    "tangible": [
        ("preferred iPad removed at end of break", "screamed + reached for device", "staff redirected to alternative"),
        ("peer took preferred toy", "grabbed toy back + cried", "staff mediated and returned item"),
    ],
    "automatic": [
        ("quiet unstructured moment between trials", "engaged in hand-flapping", "no social consequence; self-terminated after 30s"),
        ("waiting in line transition", "mouthed own hand", "staff redirected to task materials"),
    ],
}


# Behavior-specific ABC overrides — used in preference to the function-keyed
# templates above when the behavior has distinctive antecedents / topography.
BEHAVIOR_ABC_TEMPLATES = {
    "fecal_smearing": [
        ("quiet moment in bathroom after bowel movement", "reached into diaper, retrieved feces", "staff redirected hands and initiated clean-up"),
        ("unstructured downtime, staff out of direct line of sight", "reached behind waistband of pants", "staff intercepted and guided to bathroom"),
        ("post-toileting, before hygiene steps completed", "transferred feces to arm and wall surface", "staff blocked further smearing, cleaned skin and surface"),
        ("lying down for rest period", "reached into pull-up, brought hand to face", "staff intercepted before oral contact; hygiene routine initiated"),
    ],
    "toileting_accident": [
        ("during demand sequence at the work table", "voided (urine) without signaling need", "staff paused programming, changed clothing, resumed after 10 min"),
        ("transition from preferred activity to non-preferred task", "voided (urine) en route to the work area", "staff redirected to bathroom, changed clothing"),
        ("end of 90-minute block without scheduled toileting", "soiled (bowel) at the work table", "staff initiated clean-up and hygiene routine; program paused"),
        ("during group activity without requesting bathroom", "voided (urine) in seat", "staff removed learner from group, changed clothing, re-entered after clean-up"),
    ],
}


def _render_abc_entry(behavior_id: str, behavior_name: str, function: str, rng: random.Random) -> str:
    templates = BEHAVIOR_ABC_TEMPLATES.get(behavior_id) or ABC_TEMPLATES.get(function) or ABC_TEMPLATES["escape"]
    a, b, c = rng.choice(templates)
    return f"ABC ({behavior_name.lower()}): A = {a}; B = {b}; C = {c}"


def render_session_log(
    *,
    synthetic_id: int,
    learner_profile: dict,
    curricula_line: str,
    programs: list,
    accuracies_per_program: list,
    behaviors: list,
    behavior_frequencies_per_behavior: list,
    behavior_functions: list,
    n_sessions: int,
    start_date: date,
    duration_per_session: int,
    ioa_session_idx: int | None,
    ioa_agreement: float | None,
    include_abc: bool,
    abc_sessions: list,
    behavioral_indicators_cluster: list,
    rng: random.Random,
) -> str:
    """Render the full multi-session log as text."""
    lines = []
    lines.append("LEARNER PROFILE")
    lines.append(f"Synthetic ID: SYN-{synthetic_id}")
    profile_text = f"{learner_profile['name']}"
    if "chronological_age" in learner_profile:
        profile_text += f" ({learner_profile['chronological_age']})"
    elif "developmental_age" in learner_profile:
        profile_text += f" (dev age {learner_profile['developmental_age']})"
    lines.append(f"Profile: {profile_text}")
    lines.append(f"Curricula: {curricula_line}")
    end_date = start_date + timedelta(days=n_sessions * 2)
    lines.append(f"Date range: Sessions 1–{n_sessions} across {(end_date - start_date).days} days ({start_date.isoformat()} to {end_date.isoformat()})")
    lines.append("")

    lines.append("ACCELERATION PROGRAMS")
    for i, (domain, skill, method) in enumerate(programs, 1):
        lines.append(f"{i}. {skill} ({domain})")
        lines.append(f"   Method: {method}")
    lines.append("")

    if behaviors:
        lines.append("DECELERATION TARGETS")
        for i, (beh, fn) in enumerate(zip(behaviors, behavior_functions), 1):
            lines.append(f"{i}. {beh['name']} — function hypothesized: {fn}")
        lines.append("")

    lines.append("SESSION DATA")
    for s_idx in range(n_sessions):
        session_num = s_idx + 1
        d = (start_date + timedelta(days=s_idx * 2)).isoformat()
        if ioa_session_idx == s_idx:
            lines.append(f"")
            lines.append(f"Session {session_num}{d} — IOA SESSION — 2 observers")
            for i, (domain, skill, method) in enumerate(programs):
                acc = accuracies_per_program[i][s_idx]
                lines.append(f"  {skill}: {round(acc*100)}% (primary); IOA {round(ioa_agreement*100)}% agreement")
            for i, beh in enumerate(behaviors):
                freq = behavior_frequencies_per_behavior[i][s_idx]
                ms = _render_behavior_measurement(beh["id"], freq, rng)
                lines.append(f"  {beh['name']}: {ms}; IOA {round(ioa_agreement*100)}% agreement")
        else:
            lines.append(f"")
            lines.append(f"Session {session_num}{d}{duration_per_session} min — 1 observer")
            for i, (domain, skill, method) in enumerate(programs):
                acc = accuracies_per_program[i][s_idx]
                ms = _render_measurements_for_program(domain, acc, rng)
                lines.append(f"  {skill}: {ms}")
            for i, beh in enumerate(behaviors):
                freq = behavior_frequencies_per_behavior[i][s_idx]
                ms = _render_behavior_measurement(beh["id"], freq, rng)
                lines.append(f"  {beh['name']}: {ms}")
            # ABC entry for this session?
            if include_abc and s_idx in abc_sessions and behaviors:
                beh_idx = rng.randrange(len(behaviors))
                fn = behavior_functions[beh_idx]
                if behavior_frequencies_per_behavior[beh_idx][s_idx] > 0:
                    beh = behaviors[beh_idx]
                    lines.append(f"  {_render_abc_entry(beh['id'], beh['name'], fn, rng)}")

    if behavioral_indicators_cluster:
        lines.append("")
        lines.append("BEHAVIORAL OBSERVATIONS (across sessions)")
        for ind in behavioral_indicators_cluster:
            lines.append(f"- {ind}")

    return "\n".join(lines)


# Interpretation section assembly


def assemble_clinical_concerns(
    pattern_id: str,
    pattern: dict,
    accuracies_first_last: tuple,
    behaviors: list,
    behavior_frequencies_per_behavior: list,
    n_sessions: int,
) -> str:
    """Produce the clinical-concerns bullets."""
    first, last = accuracies_first_last
    bullets = []
    if pattern_id == "mastery_progression":
        bullets.append(f"- Steady improvement: accuracy rose from {round(first*100)}% to {round(last*100)}% across {n_sessions} sessions.")
        bullets.append("- No clinical concerns at this time; mastery criteria should be reachable within a few additional sessions.")
    elif pattern_id == "regression":
        bullets.append(f"- **Regression identified**: accuracy has declined from {round(first*100)}% to {round(last*100)}% across {n_sessions} sessions.")
        if behaviors:
            bullets.append("- Accompanying behavior data suggests the decline may be function-related.")
    elif pattern_id == "plateau":
        bullets.append(f"- **Plateau**: accuracy has remained near {round(((first+last)/2)*100)}% for {n_sessions} sessions without meaningful improvement.")
        bullets.append("- Performance is below mastery criteria; current teaching procedure may not be effective for this learner.")
    elif pattern_id == "frustration_pattern":
        bullets.append(f"- **Frustration pattern**: declining accuracy ({round(first*100)}% -> {round(last*100)}%) with escape-indicator behaviors.")
        bullets.append("- Risk of learned helplessness and escape-maintained behavior if not addressed.")
    elif pattern_id == "variable_performance":
        avg = (first + last) / 2
        bullets.append(f"- **Variable performance**: accuracy fluctuating around {round(avg*100)}% with no clear trend.")
        bullets.append("- Inconsistency may indicate stimulus-control issues or motivational variability across sessions.")
    elif pattern_id == "prompt_dependency":
        bullets.append(f"- **Prompt dependency**: high accuracy maintained ({round(first*100)}%–{round(last*100)}%) but primarily at prompted levels.")
        bullets.append("- Independent responding remains low despite extended training.")
    elif pattern_id == "rapid_acquisition":
        bullets.append(f"- **Rapid acquisition**: accelerated mastery from {round(first*100)}% to {round(last*100)}% across {n_sessions} sessions.")
        bullets.append("- Acquisition has exceeded expected timeline.")
    elif pattern_id == "generalization_failure":
        bullets.append("- **Generalization failure**: strong performance in training context but failure to generalize to novel conditions.")
        bullets.append("- Skill may be under restricted stimulus control.")
    elif pattern_id == "extinction_burst":
        bullets.append("- **Extinction burst** observed: temporary increase in problem behavior consistent with an extinction procedure.")
        bullets.append("- This is an expected phase of behavior change and typically indicates the intervention is working.")
    elif pattern_id == "skill_loss_after_break":
        bullets.append(f"- **Skill loss following break**: performance dropped to {round(first*100)}% at return, recovering to {round(last*100)}%.")
        bullets.append("- Maintenance schedule should be reviewed and strengthened.")
    elif pattern_id == "motivating_operation_shift":
        bullets.append("- **Motivating-operation shift**: mid-log dip in responding followed by recovery suggests reinforcer-value change.")
        bullets.append("- Possible satiation with current reinforcer; preference reassessment indicated.")
    elif pattern_id == "setting_event_trigger":
        bullets.append("- **Setting event pattern**: observable change in performance correlated with an external event (illness, sleep, schedule).")
        bullets.append("- Correlation between setting event and performance should be tracked explicitly.")
    return "\n".join(bullets)


def assemble_function_hypothesis_section(
    behaviors: list, behavior_functions: list, abc_events_summary: str
) -> str:
    """Produce the function hypothesis section, or empty string if no behaviors."""
    if not behaviors:
        return ""
    lines = ["", "## Behavior Function Hypothesis"]
    for beh, fn in zip(behaviors, behavior_functions):
        lines.append(f"{beh['name']}: {fn}")
        if fn != "unknown":
            ev_bits = []
            if abc_events_summary:
                ev_bits.append(f"ABC events in the log support the {fn} function")
            # Function-specific evidence without showing the operational definition
            func_evidence = {
                "escape": "behavior occurs in the context of demand presentation and terminates contingent on task removal",
                "attention": "behavior occurs during low-attention periods and results in social attention contingently",
                "tangible": "behavior occurs when preferred items are restricted and is followed by access to the item",
                "automatic": "behavior persists across contexts without clear social mediation",
            }
            if fn in func_evidence:
                ev_bits.append(func_evidence[fn])
            lines.append(f"  Evidence: " + "; ".join(ev_bits) + ".")
        else:
            lines.append("  Evidence: insufficient data to disambiguate functions; recommend FBA / preference assessment before programming changes.")
    lines.append("")
    return "\n".join(lines)


def assemble_replacement_section(behaviors: list, behavior_functions: list, recs_pool: list) -> str:
    """Produce the replacement behavior (FCT) section — only if behaviors present."""
    if not behaviors or not recs_pool:
        return ""
    lines = ["", "### Replacement behavior (FCT)"]
    for r in recs_pool[:3]:
        lines.append(f"- {r}")
    return "\n".join(lines) + "\n"


def assemble_crisis_section(recs_pool: list, escalation_level: int) -> str:
    """Produce the crisis plan section — only if escalation >= 3 and pool has items."""
    if escalation_level < 3 or not recs_pool:
        return ""
    lines = ["", "### Crisis plan"]
    for r in recs_pool[:3]:
        lines.append(f"- {r}")
    return "\n".join(lines) + "\n"


def assemble_rationale(
    accuracies_first_last: tuple,
    mean_accuracy: float,
    n_sessions: int,
    behaviors: list,
    behavior_frequencies_per_behavior: list,
    ioa_present: bool,
    ioa_agreement: float | None,
) -> str:
    """Produce the data-supported rationale bullets."""
    first, last = accuracies_first_last
    bullets = []
    bullets.append(f"- Mean accuracy across {n_sessions} sessions: {round(mean_accuracy*100, 1)}%.")
    direction = "upward" if last > first else "downward" if last < first else "flat"
    bullets.append(f"- Trend direction: {direction} (first session: {round(first*100)}%, last session: {round(last*100)}%).")
    if behaviors:
        for beh, freqs in zip(behaviors, behavior_frequencies_per_behavior):
            mean_f = sum(freqs) / len(freqs)
            first_third = sum(freqs[:max(1, len(freqs)//3)]) / max(1, len(freqs)//3)
            last_third = sum(freqs[-max(1, len(freqs)//3):]) / max(1, len(freqs)//3)
            bullets.append(f"- {beh['name']}: mean {mean_f:.1f}/session (early-window mean {first_third:.1f} -> late-window mean {last_third:.1f}).")
    if ioa_present and ioa_agreement is not None:
        bullets.append(f"- IOA data: {round(ioa_agreement*100)}% agreement on IOA session — {'adequate' if ioa_agreement >= 0.80 else 'below threshold'}.")
    else:
        bullets.append("- No IOA session in this log; reliability of measurements cannot be independently verified.")
    return "\n".join(bullets)


# Escalation + confidence inference


def infer_escalation(
    pattern: dict,
    behaviors: list,
    behavior_frequencies_per_behavior: list,
    rules: dict,
) -> int:
    """Compute escalation level from pattern + behavior severity."""
    level = int(pattern["default_escalation"])
    severe_ids = set(rules.get("severe_behavior_ids", []))
    if any(b["id"] in severe_ids for b in behaviors):
        level = min(4, level + rules.get("severe_behavior_escalation_bump", 1))
    # Safety-immediate threshold
    safety_threshold_ids = set(rules.get("safety_immediate_threshold_behaviors", []))
    threshold_rate = rules.get("safety_immediate_sib_rate_per_session", 8)
    for beh, freqs in zip(behaviors, behavior_frequencies_per_behavior):
        if beh["id"] in safety_threshold_ids:
            if len(freqs) >= 2 and all(f >= threshold_rate for f in freqs[-2:]):
                level = 4
    return level


def infer_confidence(
    n_sessions: int,
    ioa_present: bool,
    accuracies: list,
    rules: dict,
) -> str:
    """Compute confidence level from data-quality features.

    Rules:
      - low: very short log, OR short log with no IOA and non-trivial variance
      - moderate: default when criteria for high are not met
      - high: long log with IOA evidence and low variance
    """
    mean = sum(accuracies) / len(accuracies)
    sd = (sum((a - mean) ** 2 for a in accuracies) / len(accuracies)) ** 0.5

    # Explicit low triggers
    if n_sessions < rules["min_sessions_for_moderate"]:
        return "low"
    if n_sessions < 7 and not ioa_present and sd > 0.15:
        return "low"
    if sd > 0.35:
        return "low"

    # High requires long log + IOA + modest variance
    if (
        n_sessions >= rules["min_sessions_for_high"]
        and (not rules["ioa_required_for_high"] or ioa_present)
        and sd <= rules["variance_threshold_low"]
    ):
        return "high"

    return "moderate"


# Main generator class


class SessionInterpretationGenerator:
    def __init__(self, config_dir: Path):
        self.config_dir = Path(config_dir)
        self.shared = load_shared(self.config_dir)
        self.area = load_area(self.config_dir, AREA)
        self.template = self.area["template"]
        self.taxonomy = self.area["taxonomy"]
        self.compat = self.area["compatibility"]
        self.trajectory_rules = load_yaml(self.config_dir / AREA / "trajectory_rules.yaml")
        self.recommendations = load_yaml(self.config_dir / AREA / "recommendations.yaml")

    def _pattern_by_id(self, pid: str) -> dict:
        return next(p for p in self.taxonomy["patterns"] if p["id"] == pid)

    def _behavior_by_id(self, bid: str) -> dict:
        return next(b for b in self.taxonomy["target_behaviors"] if b["id"] == bid)

    def _escalation_label(self, level: int) -> str:
        return next(e["label"] for e in self.taxonomy["escalation_levels"] if e["id"] == level)

    def render_example(self, rng: random.Random) -> dict:
        # Sample hidden pattern
        pattern = rng.choice(self.taxonomy["patterns"])
        pattern_id = pattern["id"]

        # Log length
        lo, hi = self.compat["log_length_by_pattern"][pattern_id]
        n_sessions = rng.randint(lo, hi)

        # Learner profile (any — patterns are profile-agnostic)
        profile = rng.choice(self.shared["learner_profiles"]["profiles"])

        # Acceleration programs
        n_programs = rng.randint(*self.compat["programs_per_log_range"])
        programs = sample_programs(profile["id"], n_programs, rng)
        curricula_names = sorted(set(p[0] for p in programs))
        curricula_line = " + ".join(curricula_names)

        # Trajectory per program (all programs follow the pattern's trajectory loosely)
        rule = self.trajectory_rules["rules"][pattern_id]
        accuracies_per_program = [
            generate_accuracy_trajectory(pattern_id, rule, n_sessions, rng)
            for _ in programs
        ]

        # Target behaviors (if any)
        n_beh_lo, n_beh_hi = self.compat["pattern_behavior_count_ranges"][pattern_id]
        n_behaviors = rng.randint(n_beh_lo, n_beh_hi)
        behaviors = []
        behavior_functions = []
        if n_behaviors > 0:
            function_bias = self.compat["pattern_function_bias"].get(pattern_id, [])
            candidate_behs = [
                b for b in self.taxonomy["target_behaviors"]
                if any(f in b["plausible_functions"] for f in function_bias) or not function_bias
            ]
            behaviors = rng.sample(candidate_behs, min(n_behaviors, len(candidate_behs)))
            for beh in behaviors:
                if function_bias:
                    valid = [f for f in function_bias if f in beh["plausible_functions"]]
                    behavior_functions.append(rng.choice(valid) if valid else rng.choice(beh["plausible_functions"]))
                else:
                    behavior_functions.append(rng.choice(beh["plausible_functions"]))

        # Behavior frequencies per session
        behavior_freqs = []
        if behaviors:
            traj_name = self.trajectory_rules["rules"][pattern_id]["behavior_trajectory"]
            for _ in behaviors:
                freqs = generate_behavior_frequencies(
                    traj_name, self.trajectory_rules["behavior_trajectory_library"],
                    n_sessions, rule, rng,
                )
                behavior_freqs.append(freqs)

        # IOA decision
        include_ioa = rng.random() < self.compat["ioa_inclusion_probability"]
        ioa_session_idx = None
        ioa_agreement = None
        if include_ioa and n_sessions >= 4:
            ioa_session_idx = rng.randint(n_sessions // 3, 2 * n_sessions // 3)
            if rng.random() < self.compat["ioa_low_probability"]:
                ioa_agreement = rng.uniform(*self.compat["ioa_low_range"])
            else:
                ioa_agreement = rng.uniform(*self.compat["ioa_agreement_range"])

        # ABC decision
        include_abc = bool(behaviors) and rng.random() < self.compat["abc_inclusion_probability"]
        abc_sessions = []
        if include_abc:
            candidate_sessions = [i for i in range(n_sessions) if any(bf[i] > 0 for bf in behavior_freqs)]
            n_abc = min(rng.randint(1, 3), len(candidate_sessions))
            abc_sessions = rng.sample(candidate_sessions, n_abc) if candidate_sessions else []

        # Behavioral indicators
        cluster = self.compat["pattern_behavioral_indicator_cluster"].get(pattern_id)
        indicators = []
        if cluster:
            pool = self.taxonomy["behavioral_indicators"][cluster]
            indicators = rng.sample(pool, min(rng.randint(2, 4), len(pool)))

        # Log dates
        start_date = date(2026, rng.randint(1, 10), rng.randint(1, 28))
        duration_per_session = rng.randint(*self.compat["session_duration_minutes_range"])

        # Synthetic ID
        synthetic_id = rng.randint(1000, 9999)

        # Render the user's session log
        session_log = render_session_log(
            synthetic_id=synthetic_id,
            learner_profile=profile,
            curricula_line=curricula_line,
            programs=programs,
            accuracies_per_program=accuracies_per_program,
            behaviors=behaviors,
            behavior_frequencies_per_behavior=behavior_freqs,
            behavior_functions=behavior_functions,
            n_sessions=n_sessions,
            start_date=start_date,
            duration_per_session=duration_per_session,
            ioa_session_idx=ioa_session_idx,
            ioa_agreement=ioa_agreement,
            include_abc=include_abc,
            abc_sessions=abc_sessions,
            behavioral_indicators_cluster=indicators,
            rng=rng,
        )

        # Compute gold labels
        escalation_level = infer_escalation(
            pattern, behaviors, behavior_freqs, self.compat["escalation_rules"]
        )
        # Flatten accuracies across all programs for variance check
        flat_accs = [v for traj in accuracies_per_program for v in traj]
        confidence_level = infer_confidence(
            n_sessions, ioa_session_idx is not None, flat_accs, self.compat["confidence_rules"]
        )

        # Primary-program accuracies for concerns/rationale (take first program)
        primary_accs = accuracies_per_program[0]
        mean_primary = sum(primary_accs) / len(primary_accs)

        # Assemble the assistant message sections
        clinical_concerns_bullets = assemble_clinical_concerns(
            pattern_id, pattern, (primary_accs[0], primary_accs[-1]),
            behaviors, behavior_freqs, n_sessions,
        )

        pattern_explanations = {
            "mastery_progression": "Accuracy is ascending cleanly across sessions with no interfering behaviors, consistent with a mastery-progression trajectory.",
            "regression": "Previously-established responding has declined across recent sessions, warranting review of environmental and reinforcement variables.",
            "plateau": "Responding has stabilized below mastery criterion without meaningful movement, indicating the current procedure is limiting further acquisition.",
            "frustration_pattern": "Declining accuracy co-occurs with escape-indicator behaviors, consistent with frustration and escape-maintained responding to demand.",
            "variable_performance": "High session-to-session variability with no clear trend suggests environmental or motivational inconsistency rather than a single programmatic issue.",
            "prompt_dependency": "Accuracy stays high only at prompted levels; independent responding remains low despite extended training.",
            "rapid_acquisition": "Accuracy ascended steeply beyond expected timeline for this learner's profile.",
            "generalization_failure": "Alternation between strong training performance and poor novel-condition performance indicates restricted stimulus control.",
            "extinction_burst": "Mid-log spike in problem-behavior frequency followed by recovery is consistent with a planned extinction procedure.",
            "skill_loss_after_break": "Sharp drop at the start of the log followed by gradual recovery indicates skill loss following an absence.",
            "motivating_operation_shift": "A mid-log dip followed by recovery maps onto a motivating-operation change rather than a skill issue.",
            "setting_event_trigger": "A discontinuous performance change at a specific session correlates with an external setting event.",
        }

        function_section = assemble_function_hypothesis_section(
            behaviors, behavior_functions,
            "ABC events present" if include_abc and abc_sessions else ""
        )

        pattern_recs = self.recommendations["patterns"][pattern_id]
        antecedent_pool = pattern_recs["antecedent"] or ["Continue monitoring; reassess in 2 weeks."]
        consequence_pool = pattern_recs["consequence"] or ["Maintain current consequences."]
        ant_bullets = "\n".join(f"- {r}" for r in antecedent_pool[:4])
        con_bullets = "\n".join(f"- {r}" for r in consequence_pool[:3])

        rep_section = assemble_replacement_section(
            behaviors, behavior_functions, pattern_recs.get("replacement", [])
        )
        crisis_section = assemble_crisis_section(
            pattern_recs.get("crisis", []), escalation_level
        )

        escalation_justifications = {
            1: "Current programming is effective; no change needed.",
            2: "Programming adjustments indicated at next session; no immediate safety concern.",
            3: "Combination of performance decline and/or behavior data warrants supervising BCBA review within 24–48 hours.",
            4: "Safety-critical signals require immediate cessation of current program and direct BCBA contact; consider crisis-plan activation.",
        }

        confidence_justifications = {
            "high": f"Pattern is well-supported: {n_sessions} sessions of data, low variance, and IOA evidence confirm reliability.",
            "moderate": f"Pattern is supported but alternatives cannot be fully excluded: {n_sessions} sessions of data" + (" with IOA evidence." if ioa_session_idx is not None else "; no IOA session included."),
            "low": f"Insufficient data: only {n_sessions} sessions, high variance, and/or no IOA evidence. Recommend additional data collection before programming changes.",
        }

        rationale_bullets = assemble_rationale(
            (primary_accs[0], primary_accs[-1]),
            mean_primary,
            n_sessions,
            behaviors,
            behavior_freqs,
            ioa_session_idx is not None,
            ioa_agreement,
        )

        # Fill template
        user_variant = rng.choice(self.template["user_variants"])
        user_content = user_variant.format(session_log=session_log)

        assistant_content = self.template["assistant_template"].format(
            clinical_concerns_bullets=clinical_concerns_bullets,
            pattern_class=pattern_id,
            pattern_explanation=pattern_explanations[pattern_id],
            function_hypothesis_section=function_section,
            antecedent_bullets=ant_bullets,
            replacement_section=rep_section,
            consequence_bullets=con_bullets,
            crisis_section=crisis_section,
            escalation_level=escalation_level,
            escalation_label=self._escalation_label(escalation_level),
            escalation_justification=escalation_justifications[escalation_level],
            confidence_level=confidence_level,
            confidence_justification=confidence_justifications[confidence_level],
            rationale_bullets=rationale_bullets,
        )

        # Metadata
        gold_labels = {
            "pattern_class": pattern_id,
            "behavior_functions": {b["name"]: f for b, f in zip(behaviors, behavior_functions)},
            "escalation_level": escalation_level,
            "confidence": confidence_level,
            "crisis_plan_required": escalation_level >= 3,
        }
        provenance = {
            "layer": 1,
            "area": AREA,
            "template_id": self.template["template_id"],
            "taxonomy_cells": {
                "hidden_pattern_id": pattern_id,
                "learner_profile": profile["id"],
                "n_sessions": n_sessions,
                "n_programs": len(programs),
                "n_behaviors": len(behaviors),
                "behavior_ids": [b["id"] for b in behaviors],
                "behavior_functions": list(behavior_functions),
                "has_abc_data": include_abc and bool(abc_sessions),
                "has_ioa_session": ioa_session_idx is not None,
            },
            "teacher_model": None,
            "seed_tag": str(rng.getstate()[1][0] % 100000),
        }

        return make_example_envelope(
            system_content=self.template["system_prompt"],
            user_content=user_content,
            assistant_content=assistant_content,
            task_type=TASK_TYPE,
            gold_labels=gold_labels,
            provenance=provenance,
        )