What this found, and what it did not
Every figure is regenerated by make reproduce; none of
the numbers are typed in by hand. Four of the thirteen investigations returned
negative results and they are reported alongside the rest.
44% of the rings found on the last night had a case open the night before, and the median case was first seen on night 3 of 4.
With one analyst for an hour a night, the review policy stops ₹67,900 of promotion value and harms ₹16,040 of legitimate value, on stated assumptions.
/check serves stored neighbour counts rather than computing a ring around an account live. Everything is reproducible from the raw data with make reproduce.headline precision vs cost
Ring precision against the cost of being wrong, across operating points.
ring persistence
Whether a ring found tonight existed last night, and when each was first seen.
policy frontier
What each review policy stops against what it harms, as the budget grows.
time to detection
Precision by how many nights of data have accumulated.
merchant vs platform
What one business sees against what the platform sees, at equal review capacity.
adversarial
Precision as an attacker fragments the ring into smaller cells.
relation lift
How much each kind of shared entity predicts fraud.
ieee relation lift
The same method on a payment processor's transactions.
queue by ranking
Three ways of ordering the review queue.
ring context
Ring history as an account feature, and why it did nothing.
