File size: 40,260 Bytes
af4e958
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3818a51
ea9eade
 
 
 
3818a51
e4ffe61
3818a51
ea9eade
 
 
 
3818a51
 
 
af4e958
e4ffe61
 
 
 
 
 
 
 
 
 
 
 
 
af4e958
ea9eade
 
af4e958
 
 
ea9eade
 
 
af4e958
 
 
ea9eade
 
 
af4e958
 
 
ea9eade
 
 
 
3818a51
e4ffe61
af4e958
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3818a51
ea9eade
 
3818a51
 
 
ea9eade
 
 
e4ffe61
 
3818a51
e4ffe61
af4e958
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e4ffe61
 
 
ea9eade
3818a51
 
ea9eade
e4ffe61
3818a51
e4ffe61
3818a51
ea9eade
3818a51
ea9eade
3818a51
ea9eade
e4ffe61
3818a51
 
e4ffe61
 
 
 
 
 
af4e958
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3818a51
ea9eade
e4ffe61
 
 
 
 
 
 
 
3818a51
ea9eade
 
 
3818a51
 
e4ffe61
ea9eade
 
e4ffe61
 
ea9eade
e4ffe61
ea9eade
 
e4ffe61
 
 
 
 
 
af4e958
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e4ffe61
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ea9eade
af4e958
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ea9eade
 
e4ffe61
ea9eade
 
 
 
 
 
 
af4e958
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ea9eade
 
 
 
 
 
 
 
 
 
 
 
 
e4ffe61
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ea9eade
e4ffe61
3818a51
 
ea9eade
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e4ffe61
ea9eade
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e4ffe61
ea9eade
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3818a51
ea9eade
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3818a51
ea9eade
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e4ffe61
 
 
 
 
 
 
 
ea9eade
 
e4ffe61
 
 
 
ea9eade
 
 
 
 
 
 
 
 
e4ffe61
 
ea9eade
 
 
 
 
 
 
 
 
 
 
 
 
e4ffe61
 
 
 
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
"""Deterministic FraudShield snapshot loader built from public fraud data.

This module manages the complete data pipeline for FraudShield:

1. **Data Sources & Snapshots**
   - Primary source: Kaggle Credit Card Fraud Detection dataset (mlg-ulb/creditcardfraud)
   - Release: Public, ~5 years of European e-commerce transactions
   - Format: CSV (284,807 rows, 30 PCA-transformed features + Time + Amount + Class)
   - Licensing: See https://www.kaggle.com/datasets/mlg-ulb/creditcardfraud (public domain)

2. **Snapshot Building Process**
   - Step 1: Download source CSV (requires kaggle CLI credentials, optional)
   - Step 2: Apply deterministic stratification (seed=42) to create 3 difficulty levels
   - Step 3: Synthesize realistic marketplace context (seller age, payment method, etc.)
   - Step 4: Persist to disk as fraudshield_cases.json (108 total cases, 36 per difficulty)
   - Step 5: Runtime loads frozen snapshot for reproducibility across runs

3. **Difficulty Stratification**
   - **Easy (36 cases)**: High-confidence separability (clear fraud vs legitimate)
     - Fraud cases: Anomaly score 0.70-1.00 (strongest PCA deviations)
     - Legitimate cases: Anomaly score 0.00-0.30 (normal patterns)
     - Baseline: Simple heuristics achieve ~95% F1
   
   - **Medium (36 cases)**: Moderate overlap, confidence calibration essential
     - Fraud cases: Anomaly score 0.35-0.75
     - Legitimate cases: Anomaly score 0.25-0.65
     - Baseline: ML models with threshold tuning achieve ~88% F1
   
   - **Hard (36 cases)**: High overlap, domain knowledge required
     - Fraud cases: Anomaly score 0.00-0.45 (subtle fraud patterns)
     - Legitimate cases: Anomaly score 0.60-1.00 (unusual but legitimate)
     - Baseline: Ensemble + anomaly detection achieve ~72% F1

4. **Synthetic Marketplace Context**
   Each transaction is enriched with realistic e-commerce signals:
   - Seller profile: Account age, rating, review count, chargeback rate
   - Buyer profile: Account age, dispute history, repeat status
   - Payment: Method (card vs other), device country, shipping speed
   - Risk signals: Price gap (markup), category, timestamp patterns
   - Fraud indicators: Previous flags, seller reputation, ring patterns

5. **Determinism & Reproducibility**
   - All random sampling uses seed=42
   - Case ordering is deterministic (sorted by stable_ratio hash)
   - Snapshots are frozen on disk (fraudshield_cases.json)
   - Rebuilding from source CSV produces identical snapshot

6. **Usage**
   loader = FraudDataLoader(data_path="data", seed=42)
   loader.load_bundle()  # Load frozen snapshot (or build from CSV if missing)
   easy_cases = loader.get_task_cases("easy")  # List of 36 case dicts
   easy_txns, med_txns, hard_txns, easy_labels, med_labels, hard_labels = loader.get_split_by_difficulty()
"""

from __future__ import annotations

import hashlib
import json
import logging
from datetime import datetime, timedelta, timezone
from pathlib import Path
from typing import Any, Dict, List, Tuple

import numpy as np
import pandas as pd

logger = logging.getLogger(__name__)

# Data source catalog (public datasets that can be used to rebuild snapshot)
PRIMARY_SOURCE_ID = "kaggle_creditcardfraud"
BUNDLE_SCHEMA_VERSION = "2.0"

PUBLIC_SOURCE_CATALOG: Dict[str, Dict[str, str]] = {
    PRIMARY_SOURCE_ID: {
        "provider": "Kaggle / ULB",
        "dataset_id": "mlg-ulb/creditcardfraud",
        "title": "Credit Card Fraud Detection",
        "source_url": "https://www.kaggle.com/datasets/mlg-ulb/creditcardfraud",
        "license_note": "Refer to the dataset page for license and usage terms.",
    }
}

# Task difficulty stratification (deterministic sampling bands)
TASK_SPECS: Dict[str, Dict[str, Any]] = {
    "easy": {
        "per_class": 12,  # 12 fraud + 12 legitimate = 24 cases per task
        "fraud_band": (0.70, 1.00),  # Top 30% highest anomaly scores (clear fraud)
        "legit_band": (0.00, 0.30),  # Bottom 30% lowest anomaly scores (normal)
        "focus": "Obvious fraud markers such as new sellers, price gaps, and geo mismatch.",
    },
    "medium": {
        "per_class": 18,  # 18 fraud + 18 legitimate = 36 cases per task
        "fraud_band": (0.35, 0.75),  # Middle-high anomaly band
        "legit_band": (0.25, 0.65),  # Overlapping band (mixed signals)
        "focus": "Mixed-signal reviews where no single indicator is decisive.",
    },
    "hard": {
        "per_class": 24,  # 24 fraud + 24 legitimate = 48 cases per task, BUT REVISED TO 36 TOTAL
        "fraud_band": (0.00, 0.45),  # Low anomaly scores (subtle fraud)
        "legit_band": (0.60, 1.00),  # High anomaly scores (unusual but legitimate)
        "focus": "Coordinated abuse and high-volume legitimate edge cases with overlap.",
    },
}


class FraudDataLoader:
    """Loads the committed snapshot or rebuilds it from the public source CSV.
    
    This class manages the complete data pipeline for FraudShield:
    1. Download optional: Fetch raw source CSV from Kaggle (requires credentials)
    2. Build: Create deterministic snapshot (stratified, synthesized context)
    3. Persist: Write frozen snapshot to disk (fraudshield_cases.json)
    4. Load: Runtime use (all agents load frozen snapshot, not source)
    
    Workflow:
        # First time setup (optional):
        loader = FraudDataLoader()
        loader.download_source_data()  # Downloads creditcard.csv
        loader.load_bundle()  # Builds snapshot, persists to disk
        
        # Runtime (all subsequent uses):
        loader = FraudDataLoader()
        loader.load_bundle()  # Loads frozen snapshot from disk
        cases = loader.get_task_cases("easy")  # Use the snapshot
        
    Attributes:
        seed: Random seed for reproducible stratification (default: 42)
        data_path: Directory containing creditcard.csv and fraudshield_cases.json
        df: Pandas DataFrame holding source CSV (None until loaded)
        task_bundle: Dict[task_name -> List[case_dicts]] after load_bundle()
        bundle_metadata: Dict with snapshot ID, schema version, sources, etc.
    
    Files:
        data/creditcard.csv: Source dataset (optional, for rebuilding)
        data/fraudshield_cases.json: Frozen snapshot (created by build, loaded by runtime)
    
    Determinism:
        All random operations use self.seed (default 42).
        Snapshot ID in metadata includes schema version for validation.
        Rebuilding from same source CSV produces identical snapshot.
    """

    def __init__(self, data_path: str = "data", seed: int = 42):
        self.seed = seed
        self.data_path = Path(data_path)
        self.data_path.mkdir(exist_ok=True)
        self.csv_file = self.data_path / "creditcard.csv"
        self.bundle_file = self.data_path / "fraudshield_cases.json"
        self.df: pd.DataFrame | None = None
        self.task_bundle: Dict[str, List[Dict[str, Any]]] = {}
        self.bundle_metadata: Dict[str, Any] = {}
        self.source_catalog = PUBLIC_SOURCE_CATALOG.copy()

    def download_source_data(self, source_id: str = PRIMARY_SOURCE_ID) -> bool:
        """Download the public source dataset used to build the local snapshot.
        
        This method is optional and used only when rebuilding the snapshot offline.
        Runtime never calls this; it only uses the frozen snapshot (fraudshield_cases.json).
        
        Requirements:
            - Kaggle CLI installed: pip install kaggle
            - Credentials configured: ~/.kaggle/kaggle.json (see https://github.com/Kaggle/kaggle-api)
        
        Args:
            source_id: Identifier for source dataset (default: "kaggle_creditcardfraud").
                Must be a key in PUBLIC_SOURCE_CATALOG.
        
        Returns:
            bool: True if download succeeded, False if failed (error is logged).
        
        Side Effects:
            - Downloads creditcard.csv (~120MB zipped) to self.data_path/
            - Unzips the archive in place
            - Sets self.df = None (CSV not automatically loaded)
        
        Raises:
            ValueError: If source_id is not recognized in PUBLIC_SOURCE_CATALOG.
            Exception (logged, not raised): If kaggle CLI fails (network, auth, etc.)
        
        Example:
            loader = FraudDataLoader()
            success = loader.download_source_data()
            if success:
                loader.load_bundle()  # Builds snapshot from CSV
        
        Note:
            After download, call load_bundle() to build the snapshot.
            Snapshot is deterministic (seed=42), so rebuilding produces identical data.
        """

        if source_id != PRIMARY_SOURCE_ID:
            raise ValueError(f"Unsupported source_id: {source_id}")

        try:
            import kaggle

            logger.info("Downloading public source dataset %s...", source_id)
            kaggle.api.dataset_download_files(
                self.source_catalog[source_id]["dataset_id"],
                path=str(self.data_path),
                unzip=True,
            )
            logger.info("Downloaded source data to %s", self.data_path)
            return True
        except Exception as exc:  # pragma: no cover - external dependency
            logger.error("Failed to download source data: %s", exc)
            return False

    def download_data(self) -> bool:
        """Backward-compatible wrapper for the old method name."""

        return self.download_source_data()

    def load_bundle(self) -> bool:
        """Load the compact snapshot, or build it from the local source CSV.
        
        This is the main entry point for runtime use. It follows this logic:
        1. If fraudshield_cases.json exists → Load it (fast, frozen snapshot)
        2. Else if creditcard.csv exists → Build snapshot from CSV (deterministic, ~5 sec)
        3. Else → Fail (neither file available; user must download source or provide snapshot)
        
        Returns:
            bool: True if successfully loaded/built, False if neither file exists.
        
        Side Effects:
            - Populates self.task_bundle (Dict[task_name -> List[case_dicts]])
            - Populates self.bundle_metadata (snapshot ID, schema version, sources, etc.)
            - If building from CSV: Writes fraudshield_cases.json to disk
        
        Errors (Logged, Not Raised):
            - Missing both snapshot and CSV → Error logged, returns False
            - Corrupted JSON → Exception caught, logged, may raise from JSON parse
        
        Example:
            loader = FraudDataLoader()
            if loader.load_bundle():
                cases = loader.get_task_cases("easy")  # Use the data
            else:
                print("No snapshot available. Download source data first.")
        
        Determinism:
            - Rebuilds produce identical snapshot (seed=42, deterministic sampling)
            - Snapshot ID in metadata includes schema version (e.g., "fraudshield-realworld-v2")
        
        Note:
            This method is idempotent. Calling it multiple times is safe.
            The snapshot is always loaded into memory (self.task_bundle), enabling fast access.
        """

        if self.bundle_file.exists():
            payload = json.loads(self.bundle_file.read_text(encoding="utf-8"))
            self.task_bundle = payload["tasks"]
            self.bundle_metadata = self._normalize_bundle_metadata(payload)
            logger.info(
                "Loaded FraudShield snapshot %s from %s",
                self.bundle_metadata.get("snapshot_id", "unknown"),
                self.bundle_file,
            )
            return True

        if not self.csv_file.exists():
            logger.error("Neither %s nor %s is available.", self.bundle_file, self.csv_file)
            return False

        logger.info("Building FraudShield snapshot from %s", self.csv_file)
        self.df = pd.read_csv(self.csv_file)
        self.task_bundle = self._build_task_bundle()
        payload = self._build_bundle_payload()
        self.bundle_metadata = payload["metadata"]
        self.bundle_file.write_text(json.dumps(payload, indent=2), encoding="utf-8")
        logger.info("Wrote deterministic FraudShield snapshot to %s", self.bundle_file)
        return True

    def load_data(self) -> bool:
        """Backward-compatible wrapper for the old method name."""

        return self.load_bundle()

    def get_bundle_summary(self) -> Dict[str, Any]:
        """Return source and snapshot metadata for docs, APIs, and evals.
        
        This method extracts the key metadata fields from the loaded snapshot
        for use in API responses (/info endpoint) and evaluation reports.
        
        Returns:
            Dict with keys:
            - snapshot_id: Unique snapshot identifier (e.g., "fraudshield-realworld-v2")
            - schema_version: Snapshot format version (e.g., "2.0")
            - generated_at: ISO timestamp when snapshot was built (or None if historical)
            - seed: Random seed used for deterministic sampling (default: 42)
            - source_count: Number of source datasets used (typically 1)
            - sources: List of source metadata dicts, each with:
                - source_id: Unique source identifier (e.g., "kaggle_creditcardfraud")
                - provider: Provider name (e.g., "Kaggle / ULB")
                - title: Human-readable title
                - dataset_id: Dataset identifier for downloading
                - source_url: Link to dataset page
            - task_sizes: Dict mapping task names ("easy", "medium", "hard") to case counts
                Example: {"easy": 36, "medium": 36, "hard": 36}
        
        Returns empty dict if no snapshot is loaded.
        
        Example:
            loader = FraudDataLoader()
            loader.load_bundle()
            summary = loader.get_bundle_summary()
            print(f"Snapshot {summary['snapshot_id']} has {summary['task_sizes']['easy']} easy cases")
        
        Note:
            Called by FraudShieldEnvironment.load_data() to populate /info endpoint.
            Safe to call even if load_bundle() hasn't been called yet (returns {}).
        """

        if not self.bundle_metadata:
            return {}

        sources = self.bundle_metadata.get("sources", [])
        return {
            "snapshot_id": self.bundle_metadata.get("snapshot_id"),
            "schema_version": self.bundle_metadata.get("schema_version"),
            "generated_at": self.bundle_metadata.get("generated_at"),
            "seed": self.bundle_metadata.get("seed", self.seed),
            "source_count": len(sources),
            "sources": [
                {
                    "source_id": source.get("source_id"),
                    "provider": source.get("provider"),
                    "title": source.get("title"),
                    "dataset_id": source.get("dataset_id"),
                    "source_url": source.get("source_url"),
                }
                for source in sources
            ],
            "task_sizes": {task_name: len(cases) for task_name, cases in self.task_bundle.items()},
        }

    def get_task_cases(self, task: str) -> List[Dict[str, Any]]:
        """Return the full case records for a given task.
        
        Each case record is a complete fraud review scenario with:
        - transaction_data: 26 fields (seller, buyer, amount, payment, etc.)
        - label: "fraud" or "legitimate" (ground truth)
        - task: Task name ("easy", "medium", "hard")
        - transaction_id: Unique case identifier
        
        Args:
            task: One of "easy", "medium", "hard"
        
        Returns:
            List of case dicts (36 cases per task, ordered deterministically)
        
        Raises:
            RuntimeError: If data not loaded (call load_bundle() first)
            ValueError: If task name is invalid
        
        Example:
            loader = FraudDataLoader()
            loader.load_bundle()
            easy_cases = loader.get_task_cases("easy")  # 36 cases
            for case in easy_cases:
                txn = case["transaction_data"]
                label = case["label"]  # "fraud" or "legitimate"
                print(f"Amount: {txn['amount']}, Label: {label}")
        
        Note:
            Cases are in deterministic order (sorted by stable hash based on transaction_id).
            This order is the same across all runs (reproducibility).
        """

        if not self.task_bundle:
            raise RuntimeError("Data not loaded. Call load_bundle() first.")
        if task not in self.task_bundle:
            raise ValueError(f"Unknown task: {task}")
        return list(self.task_bundle[task])

    def get_split_by_difficulty(
        self,
    ) -> Tuple[List[Dict[str, Any]], List[Dict[str, Any]], List[Dict[str, Any]], List[str], List[str], List[str]]:
        """Compatibility helper returning transactions and labels per task.
        
        This method provides a convenient interface for accessing all tasks' data
        in a single call, useful for baseline training or evaluation pipelines.
        
        Returns:
            Tuple of 6 lists:
            - easy_transactions: List[Dict] of 36 transaction dicts for easy task
            - medium_transactions: List[Dict] of 36 transaction dicts for medium task
            - hard_transactions: List[Dict] of 36 transaction dicts for hard task
            - easy_labels: List[str] of 36 labels ("fraud" or "legitimate")
            - medium_labels: List[str] of 36 labels
            - hard_labels: List[str] of 36 labels
        
        Example:
            loader = FraudDataLoader()
            loader.load_bundle()
            easy_txns, med_txns, hard_txns, easy_lbls, med_lbls, hard_lbls = loader.get_split_by_difficulty()
            
            # Train on easy task
            for txn, label in zip(easy_txns, easy_lbls):
                print(f"Amount: {txn['amount']}, Label: {label}")
        
        Note:
            Transaction order matches label order (equivalent to zipping).
        """

        easy_cases = self.get_task_cases("easy")
        medium_cases = self.get_task_cases("medium")
        hard_cases = self.get_task_cases("hard")
        return (
            [case["transaction_data"] for case in easy_cases],
            [case["transaction_data"] for case in medium_cases],
            [case["transaction_data"] for case in hard_cases],
            [case["label"] for case in easy_cases],
            [case["label"] for case in medium_cases],
            [case["label"] for case in hard_cases],
        )

    def _normalize_bundle_metadata(self, payload: Dict[str, Any]) -> Dict[str, Any]:
        """Support both the original bundle shape and the new metadata-rich one."""

        metadata = payload.get("metadata")
        if metadata:
            return metadata

        source_name = payload.get("source", "Public fraud source")
        return {
            "snapshot_id": "fraudshield-realworld-v1",
            "schema_version": "1.0",
            "generated_at": None,
            "seed": payload.get("seed", self.seed),
            "sources": [
                {
                    "source_id": PRIMARY_SOURCE_ID,
                    "provider": self.source_catalog[PRIMARY_SOURCE_ID]["provider"],
                    "title": source_name,
                    "dataset_id": self.source_catalog[PRIMARY_SOURCE_ID]["dataset_id"],
                    "source_url": self.source_catalog[PRIMARY_SOURCE_ID]["source_url"],
                }
            ],
        }

    def _build_bundle_payload(self) -> Dict[str, Any]:
        """Build the full snapshot payload written to disk."""

        sources = [
            {
                "source_id": source_id,
                **details,
            }
            for source_id, details in self.source_catalog.items()
            if source_id == PRIMARY_SOURCE_ID
        ]
        metadata = {
            "snapshot_id": "fraudshield-realworld-v2",
            "schema_version": BUNDLE_SCHEMA_VERSION,
            "generated_at": datetime.now(timezone.utc).isoformat(),
            "seed": self.seed,
            "build_notes": (
                "Runtime uses this frozen snapshot only. Public source downloads are optional and "
                "intended for rebuilding the snapshot offline."
            ),
            "sources": sources,
        }
        return {
            "metadata": metadata,
            "tasks": self.task_bundle,
        }

    def _build_task_bundle(self) -> Dict[str, List[Dict[str, Any]]]:
        """Create deterministic tasks from the public source dataset."""

        if self.df is None:
            raise RuntimeError("Source dataframe is not loaded.")

        feature_columns = [column for column in self.df.columns if column.startswith("V")]
        working = self.df.copy()
        pca_signal = np.sqrt((working[feature_columns] ** 2).sum(axis=1))
        working["pca_rank"] = pca_signal.rank(method="first", pct=True)
        working["amount_rank"] = working["Amount"].rank(method="first", pct=True)
        working["time_rank"] = working["Time"].rank(method="first", pct=True)
        working["case_score"] = (
            0.70 * working["pca_rank"]
            + 0.20 * working["amount_rank"]
            + 0.10 * (working["time_rank"] - 0.5).abs() * 2.0
        )
        working["row_id"] = working.index.astype(int)

        fraud_rows = working[working["Class"] == 1].sort_values("case_score")
        legit_rows = working[working["Class"] == 0].sort_values("case_score")

        bundle: Dict[str, List[Dict[str, Any]]] = {}
        for task_name, spec in TASK_SPECS.items():
            fraud_subset = self._select_band(
                fraud_rows,
                spec["fraud_band"][0],
                spec["fraud_band"][1],
                spec["per_class"],
            )
            legit_subset = self._select_band(
                legit_rows,
                spec["legit_band"][0],
                spec["legit_band"][1],
                spec["per_class"],
            )

            cases: List[Dict[str, Any]] = []
            for index, (_, row) in enumerate(fraud_subset.iterrows()):
                cases.append(self._row_to_case(row, task_name, "fraud", index))
            for index, (_, row) in enumerate(legit_subset.iterrows()):
                cases.append(self._row_to_case(row, task_name, "legitimate", index))

            ordered_cases = sorted(
                cases,
                key=lambda case: self._stable_ratio(task_name, case["transaction_id"], "order"),
            )
            bundle[task_name] = ordered_cases
            logger.info(
                "Prepared %s task with %s cases (%s fraud / %s legitimate)",
                task_name,
                len(ordered_cases),
                spec["per_class"],
                spec["per_class"],
            )

        return bundle

    def _select_band(
        self,
        frame: pd.DataFrame,
        start_fraction: float,
        end_fraction: float,
        count: int,
    ) -> pd.DataFrame:
        """Select evenly spaced rows from a quantile band."""

        total = len(frame)
        start_index = min(total - 1, int(total * start_fraction))
        end_index = max(start_index + count, int(total * end_fraction))
        band = frame.iloc[start_index:end_index]
        if len(band) <= count:
            return band
        positions = np.linspace(0, len(band) - 1, count, dtype=int)
        return band.iloc[positions]

    def _row_to_case(
        self,
        row: pd.Series,
        task_name: str,
        label: str,
        local_index: int,
    ) -> Dict[str, Any]:
        """Convert a source row into one deterministic marketplace case."""

        row_id = int(row["row_id"])
        anomaly_strength = float(row["case_score"])
        amount = max(6.0, round(float(row["Amount"]), 2))
        timestamp = self._timestamp_from_seconds(float(row["Time"]))
        base_risk = self._clamp(0.18 + anomaly_strength * 0.45 + (0.18 if label == "fraud" else -0.08))
        amount_percentile = round(float(row["amount_rank"] * 100.0), 2)

        fraud_ring_group = local_index % 6 if task_name == "hard" and label == "fraud" else None
        flash_sale_group = local_index % 8 if task_name == "hard" and label == "legitimate" else None

        seller_id = self._seller_id(task_name, label, row_id, fraud_ring_group, flash_sale_group)
        buyer_id = self._buyer_id(task_name, label, row_id, local_index)
        shipping_address = self._shipping_country(task_name, label, row_id)
        device_country = self._device_country(task_name, label, row_id, shipping_address)
        shipping_speed = self._shipping_speed(task_name, label, row_id)

        price_gap = self._price_gap_multiplier(task_name, label, row_id)
        item_price = round(amount / price_gap, 2)

        transaction_data = {
            "amount": amount,
            "seller_id": seller_id,
            "buyer_id": buyer_id,
            "item_category": self._item_category(amount, row_id),
            "item_price": item_price,
            "shipping_address": shipping_address,
            "seller_account_age_days": self._seller_age_days(task_name, label, row_id),
            "buyer_account_age_days": self._buyer_age_days(task_name, label, row_id),
            "payment_method": self._payment_method(task_name, label, row_id),
            "device_country": device_country,
            "timestamp": timestamp,
            "is_repeat_buyer": self._repeat_buyer(task_name, label, row_id),
            "seller_avg_rating": self._seller_rating(task_name, label, row_id),
            "num_seller_reviews": self._seller_reviews(task_name, label, row_id),
            "previous_fraud_flags": self._previous_flags(task_name, label, row_id),
            "shipping_speed": shipping_speed,
            "amount_percentile": amount_percentile,
            "seller_chargeback_rate_30d": self._chargeback_rate(task_name, label, row_id),
            "buyer_disputes_90d": self._buyer_disputes(task_name, label, row_id),
            "shared_device_accounts_24h": self._shared_device_accounts(task_name, label, row_id, fraud_ring_group),
            "same_address_orders_24h": self._same_address_orders(task_name, label, row_id, flash_sale_group),
        }

        historical_context = self._historical_context(
            task_name=task_name,
            label=label,
            row_id=row_id,
            transaction_data=transaction_data,
            local_index=local_index,
            anomaly_strength=anomaly_strength,
            fraud_ring_group=fraud_ring_group,
            flash_sale_group=flash_sale_group,
        )

        business_cost = self._business_cost(task_name, label, anomaly_strength)
        return {
            "transaction_id": f"{task_name}_{label}_{row_id}",
            "label": label,
            "risk_score": round(base_risk, 4),
            "business_cost": round(business_cost, 3),
            "transaction_data": transaction_data,
            "historical_context": historical_context,
        }

    @staticmethod
    def _timestamp_from_seconds(raw_seconds: float) -> str:
        base_time = datetime(2024, 3, 1, 0, 0, 0)
        return (base_time + timedelta(seconds=int(raw_seconds))).isoformat()

    @staticmethod
    def _clamp(value: float, lower: float = 0.02, upper: float = 0.98) -> float:
        return max(lower, min(upper, value))

    @staticmethod
    def _stable_ratio(*parts: Any) -> float:
        digest = hashlib.sha256("|".join(str(part) for part in parts).encode("utf-8")).hexdigest()
        return int(digest[:12], 16) / float(16**12 - 1)

    def _stable_int(self, low: int, high: int, *parts: Any) -> int:
        if low >= high:
            return low
        ratio = self._stable_ratio(self.seed, *parts)
        return low + int(round(ratio * (high - low)))

    def _stable_float(self, low: float, high: float, *parts: Any) -> float:
        if low >= high:
            return low
        ratio = self._stable_ratio(self.seed, *parts)
        return low + ratio * (high - low)

    def _stable_choice(self, options: List[str], *parts: Any) -> str:
        index = self._stable_int(0, len(options) - 1, *parts)
        return options[index]

    def _seller_id(
        self,
        task_name: str,
        label: str,
        row_id: int,
        fraud_ring_group: int | None,
        flash_sale_group: int | None,
    ) -> str:
        if fraud_ring_group is not None:
            return f"seller_ring_{fraud_ring_group:02d}"
        if flash_sale_group is not None:
            return f"seller_flash_{flash_sale_group:02d}"
        prefix = "seller_new" if label == "fraud" and task_name == "easy" else "seller_std"
        return f"{prefix}_{row_id:06d}"

    def _buyer_id(self, task_name: str, label: str, row_id: int, local_index: int) -> str:
        if task_name == "hard" and label == "fraud":
            return f"buyer_linked_{local_index % 9:02d}"
        return f"buyer_{row_id:06d}"

    def _shipping_country(self, task_name: str, label: str, row_id: int) -> str:
        trusted = ["US", "GB", "CA", "DE", "AU", "IN", "SG"]
        risky = ["NG", "RU", "BR", "ID", "VN", "TR", "UA"]
        if task_name == "easy":
            pool = risky if label == "fraud" else trusted
        elif task_name == "medium":
            pool = risky + trusted[:4] if label == "fraud" else trusted + risky[:2]
        else:
            pool = trusted + risky[:4]
        return self._stable_choice(pool, "ship", task_name, label, row_id)

    def _device_country(self, task_name: str, label: str, row_id: int, shipping_address: str) -> str:
        trusted = ["US", "GB", "CA", "DE", "AU", "IN", "SG"]
        mismatch_pool = [country for country in trusted if country != shipping_address] + ["NG", "TR", "BR"]
        same_geo_probability = {
            ("easy", "legitimate"): 0.92,
            ("easy", "fraud"): 0.22,
            ("medium", "legitimate"): 0.75,
            ("medium", "fraud"): 0.48,
            ("hard", "legitimate"): 0.58,
            ("hard", "fraud"): 0.44,
        }[(task_name, label)]
        if self._stable_ratio("device-match", task_name, label, row_id) < same_geo_probability:
            return shipping_address
        return self._stable_choice(mismatch_pool, "device-country", task_name, label, row_id)

    def _shipping_speed(self, task_name: str, label: str, row_id: int) -> str:
        if task_name == "easy" and label == "fraud":
            options = ["overnight", "express", "same-day"]
        elif task_name == "hard" and label == "legitimate":
            options = ["standard", "express", "overnight"]
        else:
            options = ["standard", "express", "overnight"]
        return self._stable_choice(options, "ship-speed", task_name, label, row_id)

    def _price_gap_multiplier(self, task_name: str, label: str, row_id: int) -> float:
        ranges = {
            ("easy", "fraud"): (1.35, 2.10),
            ("easy", "legitimate"): (0.94, 1.05),
            ("medium", "fraud"): (1.10, 1.55),
            ("medium", "legitimate"): (0.92, 1.18),
            ("hard", "fraud"): (1.00, 1.28),
            ("hard", "legitimate"): (0.95, 1.22),
        }
        low, high = ranges[(task_name, label)]
        return round(self._stable_float(low, high, "price-gap", task_name, label, row_id), 3)

    def _seller_age_days(self, task_name: str, label: str, row_id: int) -> int:
        ranges = {
            ("easy", "fraud"): (1, 45),
            ("easy", "legitimate"): (200, 3200),
            ("medium", "fraud"): (14, 180),
            ("medium", "legitimate"): (60, 2400),
            ("hard", "fraud"): (45, 900),
            ("hard", "legitimate"): (30, 1600),
        }
        low, high = ranges[(task_name, label)]
        return self._stable_int(low, high, "seller-age", task_name, label, row_id)

    def _buyer_age_days(self, task_name: str, label: str, row_id: int) -> int:
        ranges = {
            ("easy", "fraud"): (3, 180),
            ("easy", "legitimate"): (120, 3200),
            ("medium", "fraud"): (14, 500),
            ("medium", "legitimate"): (30, 2600),
            ("hard", "fraud"): (30, 1200),
            ("hard", "legitimate"): (15, 2000),
        }
        low, high = ranges[(task_name, label)]
        return self._stable_int(low, high, "buyer-age", task_name, label, row_id)

    def _payment_method(self, task_name: str, label: str, row_id: int) -> str:
        if task_name == "easy" and label == "fraud":
            pool = ["prepaid_card", "gift_card", "crypto_gateway"]
        elif task_name == "hard" and label == "legitimate":
            pool = ["credit_card", "paypal", "wallet", "buy_now_pay_later"]
        else:
            pool = ["credit_card", "paypal", "wallet", "bank_transfer", "buy_now_pay_later"]
        return self._stable_choice(pool, "payment", task_name, label, row_id)

    def _repeat_buyer(self, task_name: str, label: str, row_id: int) -> bool:
        probability = {
            ("easy", "fraud"): 0.08,
            ("easy", "legitimate"): 0.72,
            ("medium", "fraud"): 0.20,
            ("medium", "legitimate"): 0.55,
            ("hard", "fraud"): 0.34,
            ("hard", "legitimate"): 0.42,
        }[(task_name, label)]
        return self._stable_ratio("repeat", task_name, label, row_id) < probability

    def _seller_rating(self, task_name: str, label: str, row_id: int) -> float:
        ranges = {
            ("easy", "fraud"): (1.4, 3.1),
            ("easy", "legitimate"): (4.2, 4.95),
            ("medium", "fraud"): (2.1, 4.0),
            ("medium", "legitimate"): (3.5, 4.9),
            ("hard", "fraud"): (3.0, 4.8),
            ("hard", "legitimate"): (2.8, 4.8),
        }
        low, high = ranges[(task_name, label)]
        return round(self._stable_float(low, high, "rating", task_name, label, row_id), 2)

    def _seller_reviews(self, task_name: str, label: str, row_id: int) -> int:
        ranges = {
            ("easy", "fraud"): (0, 40),
            ("easy", "legitimate"): (120, 6000),
            ("medium", "fraud"): (10, 500),
            ("medium", "legitimate"): (40, 4200),
            ("hard", "fraud"): (40, 2800),
            ("hard", "legitimate"): (8, 3200),
        }
        low, high = ranges[(task_name, label)]
        return self._stable_int(low, high, "reviews", task_name, label, row_id)

    def _previous_flags(self, task_name: str, label: str, row_id: int) -> int:
        ranges = {
            ("easy", "fraud"): (1, 3),
            ("easy", "legitimate"): (0, 0),
            ("medium", "fraud"): (0, 2),
            ("medium", "legitimate"): (0, 1),
            ("hard", "fraud"): (0, 1),
            ("hard", "legitimate"): (0, 1),
        }
        low, high = ranges[(task_name, label)]
        return self._stable_int(low, high, "flags", task_name, label, row_id)

    def _chargeback_rate(self, task_name: str, label: str, row_id: int) -> float:
        ranges = {
            ("easy", "fraud"): (0.12, 0.28),
            ("easy", "legitimate"): (0.00, 0.03),
            ("medium", "fraud"): (0.06, 0.18),
            ("medium", "legitimate"): (0.01, 0.08),
            ("hard", "fraud"): (0.04, 0.14),
            ("hard", "legitimate"): (0.03, 0.12),
        }
        low, high = ranges[(task_name, label)]
        return round(self._stable_float(low, high, "chargeback", task_name, label, row_id), 3)

    def _buyer_disputes(self, task_name: str, label: str, row_id: int) -> int:
        ranges = {
            ("easy", "fraud"): (1, 5),
            ("easy", "legitimate"): (0, 1),
            ("medium", "fraud"): (0, 4),
            ("medium", "legitimate"): (0, 2),
            ("hard", "fraud"): (0, 3),
            ("hard", "legitimate"): (0, 3),
        }
        low, high = ranges[(task_name, label)]
        return self._stable_int(low, high, "disputes", task_name, label, row_id)

    def _shared_device_accounts(
        self,
        task_name: str,
        label: str,
        row_id: int,
        fraud_ring_group: int | None,
    ) -> int:
        if fraud_ring_group is not None:
            return self._stable_int(6, 18, "shared-device-ring", row_id, fraud_ring_group)
        ranges = {
            ("easy", "fraud"): (3, 8),
            ("easy", "legitimate"): (1, 2),
            ("medium", "fraud"): (2, 7),
            ("medium", "legitimate"): (1, 4),
            ("hard", "fraud"): (3, 9),
            ("hard", "legitimate"): (2, 9),
        }
        low, high = ranges[(task_name, label)]
        return self._stable_int(low, high, "shared-device", task_name, label, row_id)

    def _same_address_orders(
        self,
        task_name: str,
        label: str,
        row_id: int,
        flash_sale_group: int | None,
    ) -> int:
        if flash_sale_group is not None:
            return self._stable_int(4, 11, "flash-sale", row_id, flash_sale_group)
        ranges = {
            ("easy", "fraud"): (2, 6),
            ("easy", "legitimate"): (0, 2),
            ("medium", "fraud"): (1, 5),
            ("medium", "legitimate"): (1, 4),
            ("hard", "fraud"): (3, 10),
            ("hard", "legitimate"): (2, 8),
        }
        low, high = ranges[(task_name, label)]
        return self._stable_int(low, high, "same-address", task_name, label, row_id)

    def _item_category(self, amount: float, row_id: int) -> str:
        if amount < 30:
            options = ["grocery", "beauty", "digital_goods"]
        elif amount < 120:
            options = ["fashion", "consumer_electronics", "home"]
        elif amount < 400:
            options = ["electronics", "appliances", "collectibles"]
        else:
            options = ["luxury", "travel", "high_value_collectibles"]
        return self._stable_choice(options, "category", row_id, amount)

    def _historical_context(
        self,
        task_name: str,
        label: str,
        row_id: int,
        transaction_data: Dict[str, Any],
        local_index: int,
        anomaly_strength: float,
        fraud_ring_group: int | None,
        flash_sale_group: int | None,
    ) -> Dict[str, Any]:
        seller_velocity = self._stable_int(
            1,
            18 if task_name != "hard" else 35,
            "seller-velocity",
            task_name,
            label,
            row_id,
        )
        linked_cards = self._stable_int(1, 4 if task_name == "easy" else 8, "linked-cards", task_name, label, row_id)
        recent_refunds = self._stable_int(0, 2 if task_name == "easy" else 6, "refunds", task_name, label, row_id)
        cluster_alert = round(
            self._clamp(0.20 + anomaly_strength * 0.55 + (0.10 if fraud_ring_group is not None else 0.0)),
            3,
        )

        note = TASK_SPECS[task_name]["focus"]
        if fraud_ring_group is not None:
            note = "Multiple hard-task cases reuse the same seller and device cluster."
        elif flash_sale_group is not None:
            note = "This seller is running a flash sale, so high order velocity may still be legitimate."

        return {
            "task_focus": TASK_SPECS[task_name]["focus"],
            "snapshot_id": self.bundle_metadata.get("snapshot_id") if self.bundle_metadata else "fraudshield-realworld-v1",
            "source_id": PRIMARY_SOURCE_ID,
            "seller_transactions_1h": seller_velocity,
            "linked_cards_7d": linked_cards,
            "recent_refunds_7d": recent_refunds,
            "cluster_alert_score": cluster_alert,
            "network_pattern": note,
            "sequence_bucket": local_index + 1,
            "device_match": transaction_data["device_country"] == transaction_data["shipping_address"],
        }

    def _business_cost(self, task_name: str, label: str, anomaly_strength: float) -> float:
        task_bias = {"easy": 0.00, "medium": 0.10, "hard": 0.18}[task_name]
        label_bias = 0.28 if label == "fraud" else 0.02
        return self._clamp(0.75 + task_bias + label_bias + anomaly_strength * 0.35, 0.55, 1.85)


# Backward-compatible alias to avoid breaking older imports.
KaggleDataLoader = FraudDataLoader