Hypotheses
Detection workbench
A judge should be able to see what signals were tried, which ones are still active, and where each hypothesis shows up in the reviewer workflow.
Hypotheses kept
5
Focused on signals that improve reviewer confidence within the challenge scope.
Active fraud patterns
5
Primary detector patterns currently visible in scored transactions.
Review rate
8%
75 rows are above the current risk threshold.
Feedback events
17
Session decisions available for audit, undo, and retraining context.
Hypothesis board
These are the product-facing hypotheses the UI should make easy to defend.
| Hypothesis | Status | Evidence | Open |
|---|---|---|---|
| Per-card baseline | kept | Flags compare amount, country, merchant category, device, and IP against each card's own history. | Inspect |
| Cross-card graph reuse | kept | Shared devices, IPs, and merchants are ranked as clusters because one fraud pattern is invisible per card. | Inspect |
| Explainable queue | kept |
Signal strength
Average component score across the current month.
Model consensus
500 active flags
50
Category surprise
94 active flags
28
Card baseline
16 active flags
11
Shared entity graph
0 active flags
7
Velocity
39 active flags
4
Top pattern evidence
The strongest hypotheses should map directly to queue filters.