← Back to the overview
Ring case files
Week-2 graph, 35,701,750 edges. Score cut-off τ = 0.5, λ = 5.0. Ring precision 0.7292 against a base rate of 0.2242.
25rings
1,638accounts
245fraud found
0.371real customers per catch
₹74,400cost of being wrong
Each card carries the action the review policy would take if these 10 rings were tonight's queue and an analyst had 120 minutes, with a wrongly held customer assumed to lose 10% of their value. The two figures behind it are shown so a reviewer can see which one the recommendation turned on.
These are leads for a human to review, not verdicts. Rupee figures are counts multiplied by an assumed value — this dataset ships no monetary amounts — so they rank rings against each other and mean nothing in absolute terms.
Ring #15 — 14 accounts
Case #279 · first seen night 2 of 4 · continued · matches the tracked ring at Jaccard 0.60
mean member score 0.906 against 0.192 across every scored account — 4.7× the population
density 7.0657 · 60 orders over 8 days · 23% of them on the busiest single day · ₹3,100 at stake
5 known fraud0 known good9 unreviewed
Auto-holdreviewing it is worth ₹1,632 and costs 6 analyst minutesauto-holding it is worth ₹1,619 and costs none
| what they share | members | coverage | accounts with it, platform-wide |
|---|
order location R1-001 | 9 | 64% | 11 |
order location R1-002 | 7 | 50% | 8 |
order location R1-003 | 4 | 29% | 4 |
order location R1-004 | 3 | 21% | 3 |
order location R1-005 | 3 | 21% | 5 |
promotion R6-001 | 9 | 64% | 392 |
order location R1-006 | 3 | 21% | 8 |
order location R1-007 | 3 | 21% | 8 |
Coverage only means something next to the last column. An entity that a handful of accounts on the whole platform have ever used is evidence; one that millions share is not.
Ring #16 — 40 accounts
mean member score 0.819 against 0.192 across every scored account — 4.3× the population
density 7.0503 · 250 orders over 8 days · 14% of them on the busiest single day · ₹12,300 at stake
12 known fraud7 known good21 unreviewed
Auto-holdreviewing it is worth ₹4,833 and costs 13 analyst minutesauto-holding it is worth ₹4,669 and costs none
| what they share | members | coverage | accounts with it, platform-wide |
|---|
order location R1-008 | 11 | 28% | 22 |
promotion R6-003 | 11 | 28% | 96 |
promotion R6-004 | 8 | 20% | 106 |
Coverage only means something next to the last column. An entity that a handful of accounts on the whole platform have ever used is evidence; one that millions share is not.
Ring #20 — 7 accounts
mean member score 0.785 against 0.192 across every scored account — 4.1× the population
density 6.5859 · 163 orders over 7 days · 42% of them on the busiest single day · ₹14,400 at stake
4 known fraud0 known good3 unreviewed
Reviewreviewing it is worth ₹5,420 and costs 5 analyst minutesauto-holding it is worth ₹5,210 and costs none
| what they share | members | coverage | accounts with it, platform-wide |
|---|
order location R1-009 | 4 | 57% | 4 |
order location R1-010 | 4 | 57% | 4 |
sales stimulation R8-001 | 3 | 43% | 3 |
sales stimulation R8-002 | 2 | 29% | 2 |
sales stimulation R8-003 | 2 | 29% | 2 |
sales stimulation R8-004 | 2 | 29% | 2 |
sales stimulation R8-005 | 2 | 29% | 2 |
sales stimulation R8-006 | 2 | 29% | 2 |
Coverage only means something next to the last column. An entity that a handful of accounts on the whole platform have ever used is evidence; one that millions share is not.
Ring #5 — 48 accounts
mean member score 0.783 against 0.192 across every scored account — 4.1× the population
density 8.9545 · 117 orders over 8 days · 27% of them on the busiest single day · ₹2,300 at stake
3 known fraud0 known good45 unreviewed
Auto-holdreviewing it is worth ₹1,096 and costs 15 analyst minutesauto-holding it is worth ₹1,009 and costs none
| what they share | members | coverage | accounts with it, platform-wide |
|---|
order location R1-011 | 24 | 50% | 39 |
order location R1-012 | 24 | 50% | 43 |
Coverage only means something next to the last column. An entity that a handful of accounts on the whole platform have ever used is evidence; one that millions share is not.
Ring #6 — 43 accounts
Case #565 · first seen night 4 of 4 · born · matches the tracked ring at Jaccard 0.98
mean member score 0.749 against 0.192 across every scored account — 3.9× the population
density 8.7761 · 275 orders over 8 days · 25% of them on the busiest single day · ₹12,100 at stake
2 known fraud0 known good41 unreviewed
Reviewreviewing it is worth ₹4,491 and costs 14 analyst minutesauto-holding it is worth ₹4,125 and costs none
| what they share | members | coverage | accounts with it, platform-wide |
|---|
sales stimulation R8-011 | 43 | 100% | 93 |
Coverage only means something next to the last column. An entity that a handful of accounts on the whole platform have ever used is evidence; one that millions share is not.
Ring #12 — 171 accounts
mean member score 0.707 against 0.192 across every scored account — 3.7× the population
density 7.1129 · 639 orders over 8 days · 20% of them on the busiest single day · ₹24,200 at stake
50 known fraud22 known good99 unreviewed
Reviewreviewing it is worth ₹9,403 and costs 46 analyst minutesauto-holding it is worth ₹8,578 and costs none
| what they share | members | coverage | accounts with it, platform-wide |
|---|
promotion R6-005 | 47 | 27% | 79 |
Coverage only means something next to the last column. An entity that a handful of accounts on the whole platform have ever used is evidence; one that millions share is not.
Ring #9 — 68 accounts
mean member score 0.699 against 0.192 across every scored account — 3.6× the population
density 7.8662 · 418 orders over 8 days · 24% of them on the busiest single day · ₹32,700 at stake
11 known fraud1 known good56 unreviewed
Reviewreviewing it is worth ₹13,201 and costs 20 analyst minutesauto-holding it is worth ₹12,642 and costs none
| what they share | members | coverage | accounts with it, platform-wide |
|---|
promotion R6-006 | 49 | 72% | 70 |
Coverage only means something next to the last column. An entity that a handful of accounts on the whole platform have ever used is evidence; one that millions share is not.
Ring #13 — 41 accounts
mean member score 0.676 against 0.192 across every scored account — 3.5× the population
density 6.9075 · 289 orders over 8 days · 20% of them on the busiest single day · ₹16,800 at stake
1 known fraud6 known good34 unreviewed
Reviewreviewing it is worth ₹7,099 and costs 13 analyst minutesauto-holding it is worth ₹6,640 and costs none
| what they share | members | coverage | accounts with it, platform-wide |
|---|
promotion R6-007 | 11 | 27% | 38 |
promotion R6-008 | 11 | 27% | 39 |
promotion R6-009 | 7 | 17% | 475 |
Coverage only means something next to the last column. An entity that a handful of accounts on the whole platform have ever used is evidence; one that millions share is not.
Ring #18 — 34 accounts
mean member score 0.655 against 0.192 across every scored account — 3.4× the population
density 6.7647 · 211 orders over 8 days · 19% of them on the busiest single day · ₹9,100 at stake
6 known fraud2 known good26 unreviewed
Reviewreviewing it is worth ₹3,077 and costs 12 analyst minutesauto-holding it is worth ₹2,602 and costs none
| what they share | members | coverage | accounts with it, platform-wide |
|---|
sales stimulation R8-012 | 28 | 82% | 79 |
Coverage only means something next to the last column. An entity that a handful of accounts on the whole platform have ever used is evidence; one that millions share is not.
Ring #11 — 28 accounts
mean member score 0.632 against 0.192 across every scored account — 3.3× the population
density 7.0904 · 111 orders over 8 days · 24% of them on the busiest single day · ₹7,800 at stake
5 known fraud6 known good17 unreviewed
Reviewreviewing it is worth ₹3,349 and costs 10 analyst minutesauto-holding it is worth ₹3,163 and costs none
| what they share | members | coverage | accounts with it, platform-wide |
|---|
order location R1-013 | 13 | 46% | 24 |
promotion R6-010 | 8 | 29% | 10 |
order location R1-014 | 7 | 25% | 7 |
order location R1-015 | 7 | 25% | 14 |
promotion R6-011 | 5 | 18% | 7 |
promotion R6-012 | 7 | 25% | 3,552 |
promotion R6-013 | 5 | 18% | 1,224 |
Coverage only means something next to the last column. An entity that a handful of accounts on the whole platform have ever used is evidence; one that millions share is not.
The clusters it left alone, and why
2,444 of 2,446 co-located groups whose labelled members are overwhelmingly normal — hostels, shared offices, joint families — have no member in any ring (99.92% of them)
| the groups it left alone | the ones it touched |
|---|
| mean account score | 0.15 | 0.29 |
| kinds of shared entity | 1.76 | 1.50 |
| edges inside the group | 147.65 | 112.50 |
The separation is the account score, not the structure. Groups that were left alone are actually more densely connected and share more kinds of entity — which is exactly what a hostel looks like. What keeps them out of a ring is that their members do not behave like fraudsters, and that is the whole argument for scoring accounts before looking for dense structure rather than after.
Toucheda group of 23 sharing one entity, 93% of its labelled members good9 of them ended up in a ring (39%), mean score 0.318
Toucheda group of 19 sharing one entity, 92% of its labelled members good1 of them ended up in a ring (5%), mean score 0.258
These two are the failures, shown because a page that only listed the successes would be worth less. Both are groups where the members really did score high, so the score cut-off let them through — the cost of the ordering that makes everything else work.
Generated by make report.