Case Study: Social Media Platform

Higher precision, fewer false flags, zero friction

This global social media platform had a CAPTCHA problem, but the issue wasn't just about bots. Good users were getting flagged. Risky ones were slipping through. And their model's precision was starting to plateau. By strengthening identity signals, aligning scoring thresholds to business objectives, and validating trust before registration, the platform improved performance at every level.

The challenge

With millions of new registrations per month, this platform relied on a sophisticated lifecycle model to score risk, but the system wasn't calibrated to distinguish trust.

CAPTCHA was being over-triggered for good users. Some risky accounts weren't being challenges at all. And signal strength between identity attributes wasn't informing decision logic. They needed a way to shift from reactive risk scoring to proactive trust validation, without re-architecting their flow or raising false positive rates.

Their goals were clear:

Reduce unnecessary CAPTCHA friction for good users

Improve fraud recall at high-precision thresholds

Strengthen signal resolution across email, phone, and IP

Unlock confident approvals without increasing risk

Where static systems broke down

Good users flagged too often

Even legitimate signups were challenged bu CAPTCHA due to low scores, not low trust

Some fraud went unchallenged

Forty percent of known-risk registrations were missed by the platform's native model

Precision stalled at critical thresholds

Even at 100 percent precision, recall topped out at 5.49%, leaving credible users unapproved

Signals lacked connectivity

Phone, email, and IP attributes were scored in isolation, without the relationships that validate trust

How the system changed with Elephant

Elephant introduced trust scores calibrated to the platform's own registration outcomes, using both labeled and unlabeled events. Three scoring strategies were tested: Elephant standalone, a blended average of both scores, and a joint model with the platform's score used as an input feature.

What we implemented:

Historical CAPTCHA, fraud, and approval decisions were used to calibrate thresholds
Identity signals were evaluated for both individual strength and cross-signal consistency
Connectivity scores validated relationships such as phone to IP or email to name
Thresholds were aligned to business tolerances for precision, friction, and fraud

Proof of signal intelligence:

Email and phone matched to the same identity resulted in 3.2x lower fraud rate
Users found in Elephant's identity graph were 59% less likely to be fraudulent
Phone and address pairs matched to a known identity showed 42% lower fraud risk
Signups with three or more strong identity signals had 80% lower fraud rates

The real win? Trust at the point of entry

Once deployed, Elephant's trust score sharpened decision-making across the funnel. CAPTCHA was reduced for good users. Risky signups were caught earlier. And model lift was immediate, with blended and joint models outperforming the platform's baseline across every metric.

Recall at 100 percent precision improved from 5.49 to 10.94

Trust calibration enabled stronger detection without increasing false positives, nearly doubling recall in the safest zone

Recall at 92 percent precision improved from 73.69 to 83.09

The system captured more high-risk attempts while maintaining a tight false positive boundary

CAPTCHA accuracy improved 4.2pp at 70 percent precision

Fewer legitimate users were misclassified, reducing unnecessary challenge rates

Flagged 40% more risky registrations the old model missed

Trust signals identified high-risk users that existing, native logic scored as safe, strengthening the top of the funnel