Fraud signals · Admin

Open signals

Investigating

Dismissed

Detectors enabled

of 7

mediuminvestigatingduplicate_identity

Mailglow

Three accounts sharing a device fingerprint reviewed within one hour

Detected Aug 19, 2026

Detectors

Detector	What it looks for	Threshold
Review burst	Unusual review volume from one referrer or inside a narrow time window.	10+ in 48h from one source
Duplicate identity	Multiple accounts sharing a device fingerprint, email pattern or IP range.	2+ accounts, same fingerprint
Vendor-authored	Reviews submitted from IP ranges belonging to the vendor.	Any match
Incentive language	Review text matching known incentivised-campaign phrasing.	Model confidence > 0.8
Category spam	A product applying to jobs with no supporting fit votes.	2+ rejected applications
Referral campaign anomaly	Invite links producing improbably uniform ratings.	Rating variance < 0.2
Rating inflation drift	A product's average moving faster than its review volume explains.	Not yet enabled

Automatic responses

Hold flagged evidence out of the scoreApplies while a signal is open, and reverses on dismissal.Suspend BetterBid on a high-severity signalBad-actor status blocks bidding until the signal is resolved.Notify the product ownerThey can see that a signal exists, but not the detection detail.Auto-remove on model confidence above 0.95Off by default — a human should make removal decisions.

Why signals do not act alone

Every detector here has a false-positive mode. A genuinely successful launch looks like a review burst; a company with an office VPN looks like a shared fingerprint; a niche product with three honest applications looks like category spam. Holding evidence pending review costs a product a few days of score movement. Removing it wrongly costs them a rank and costs us the claim that the ranking is earned.