AML Model Gym — Top-K Alerts

Shortlist Composition

K=10 per model × 6 models → up to 60, but with overlap:

CWAE top-10 10

IF top-10 10

Ext IF top-10 10

Deep IF top-10 10

AE top-10 10

CAVE top-10 10

60 total selections

− 37 duplicates (overlap)

= 23 unique entities in shortlist

Shortlist Size

Union of 6 models' top-10

62%

Overlap Rate

37 of 60 selections shared

In All 6 Top-K

Unanimous across models

Unique Finds

In only 1 model's top-K

Shortlist — Union of All Models' Top-10

23 Unique Entities

#	Entity	Avg. Score	In Top-K Of	Top Features
1	ENT-4821 Margaret Winslow · Brampton, ON	0.95	All 6 top-K	rapid in/out intl wires structuring	Investigate →
2	ENT-3097 Darren Kowalchuk · Calgary, AB	0.93	All 6 top-K	structured cash China wires coded msgs	Investigate →
3	ENT-7214 Golden Dragon Restaurant · Richmond, BC	0.91	All 6 top-K	cash excess 68% low payroll flat revenue	Investigate →
4	ENT-5583 Jordan MacIntyre · St. Catharines, ON	0.89	All 6 top-K	cannabis EMTs high utility 45+ senders	Investigate →
5	ENT-1938 Chen Wei Holdings · Vancouver, BC	0.88	5 of 6 top-K	HK wires real estate round amounts	Investigate →
6	ENT-6641 Darren Kolychev · Mississauga, ON	0.86	3 of 6 top-K	late-night hotels multi-city EMT women	Investigate →
7	ENT-2205 Apex Capital Corp · Toronto, ON	0.84	5 of 6 top-K	layering shell txns	Investigate →
8	ENT-9102 Priya Sharma · Surrey, BC	0.82	All 6 top-K	India hawala cash deposit	Investigate →
9	ENT-4417 Mike's Auto Detailing · Brampton, ON	0.80	5 of 6 top-K	cash-intensive low payroll	Investigate →
10	ENT-8830 Viktor Petrov · Edmonton, AB	0.78	4 of 6 top-K	crypto OTC EMT velocity	Investigate →
11	ENT-3341 Nadia Okafor · Ottawa, ON	0.72	Unique to CAVE	behavioral shift context anomaly	Investigate →
12	ENT-5190 R&K Import/Export · Markham, ON	0.70	Unique to IF	outlier amounts high variance	Investigate →
13	ENT-8477 Luis Fernandez · Montreal, QC	0.68	2 of 6 top-K	reconstruction err distributional	Investigate →

Showing 13 of 23 · Scroll for more · Blue rows = unique finds (in only 1-2 models' top-K)

Score Distribution (23 shortlisted)

Histogram

0.500.600.700.800.901.00

Bimodal distribution: The 7 entities in all 6 models' top-K cluster at 0.85-0.95 (strong consensus). The 4 unique finds sit at 0.65-0.75 (caught by only 1 model). The gap between them is the interesting zone.

Top-K Overlap Matrix

K=10 per model

Cell = how many of Model A's top-10 also appear in Model B's top-10

	CWAE	IF	EIF	DIF	AE	CAVE
CWAE	10	6	7	8	4	9
IF	6	10	9	6	5	4
EIF	7	9	10	7	5	6
DIF	8	6	7	10	4	8
AE	4	5	5	4	10	3
CAVE	9	4	6	8	3	10

CWAE-CAVE share 9/10 — almost identical top-10 lists. IF-EIF share 9/10 — tree models agree. AE is the outlier (3-5 overlap with others) — it surfaces a unique population that other models miss. This is why the union approach matters: AE contributes 4 entities that no other model puts in their top-10.

Dominant Features

Cohort

Cash deposit freq.

17/20

EMT velocity

14/20

Intl wire amount

11/20

Rapid in/out

10/20

Counterparty count

9/20

Structuring pattern

8/20

Channel Concentration

Shortlist

Cash

18/20 entities use Cash

EMT

16/20 entities use EMT

Wire

11/20 entities use Wire

EFT

9/20 entities use EFT

Connex

14/20 entities use Connex

Cheque

5/20 entities use Cheque

ABM

4/20 entities use ABM

Typology Breakdown

Shortlist

Placement

40%

Layering

25%

Integration

15%

Structuring

30%

Network

20%

Entities may map to multiple typologies. Percentages reflect primary typology.