ACCOUNT TRUST

Early access

Know which accounts are one person. And see every link that says so.

Every account gets a Real, Suspicious or Fake tier from device, network, bot and linked-account evidence, with each part shown. You set the thresholds.

Join early access Talk to an engineer

Accounts / acct_91c2Example data

acct_91c2Fake

Model estimate0.77

0.500.75

Account score, compared with your two thresholds. Not calibrated yet.

acct_91c2acct_7f3eacct_2d41acct_58b0acct_e613acct_0c9a

Hover or focus a link to see why it counts, or why it does not.

Device first seen 40 minutes ago with account acct_7f3e
pendingfact
Second account, acct_91c2, on the same device
pendinginference
Shared device, counted
No automation signals
pendingnone found

Thresholds you set
A block beats an allow
Every link has a reason
Scores stay on your server

QUESTIONS THIS ANSWERS

Is this one person running ten accounts?
Linked accounts, counted and not counted, with the reason for each.
- acct_7f3eShared device, counted
- acct_2d41Shared network, counted, weakly
- acct_58b0Shared network, counted, weakly
- acct_0c9aShared device, not counted: identical hardware
Why did this account turn Suspicious yesterday?
The evidence timeline shows what arrived, and when.
1. Signed up from a VPN address
  pendingfact
  Sep 12
2. Shares a network with two other accounts this week
  pendinginference
  Sep 24Shared network, counted, weakly
Suspicious
Which links did you ignore, and why?
Links we did not count are shown and labelled: identical hardware, a device shared by more than nine accounts, or a weak session.
- Shared device, not counted: identical hardware
- Shared device, not counted: device shared by more than 9 accounts
- Shared device, not counted: weak session
Can my team overrule the model?
Yes. Put the account on an allow, block or review list. The model's tier stays visible beside your decision.
Your decision
allowTrusted customers list
Model's tier
Suspicious

How it works

Set the line yourself.

Drag the two thresholds and watch 24 example accounts re-tier. The defaults are 0.50 and 0.75, and you can change both in your settings.

Real13 allowed

0.03
0.06
0.09
0.12
0.15
0.19
0.23
0.27
0.31
0.36
0.41
0.46
0.49

Suspicious6 challenged

0.52
0.56
0.61
0.66
0.70
0.74

Fake5 blocked

0.77
0.82
0.87
0.92
0.96

Example accounts. Counts show how many would be challenged or blocked at your settings.

Links drawn, and links refused.

Counted links are solid. Links we refuse to count are dashed and labelled: identical hardware, a device held by more than nine accounts, or a session we could not tie to the device strongly. A shared office network counts, but weakly, and at most three network links count toward any account.

acct_91c2acct_7f3eacct_2d41acct_58b0acct_e613acct_0c9a

Hover or focus a link to see why it counts, or why it does not.

Old evidence fades. Automation does not.

Older sessions count for less: a session's weight falls from 1 toward one half, losing half the distance every 30 days (0.75 after 30 days, about 0.63 after 60), and after 90 days the session no longer counts, so an account is not judged forever on one odd evening. Hard automation evidence is the exception: time never brings an account that showed it back to Real.

Accounts / acct_3d80 / EvidenceExample data

Sep 24Second account on the same deviceweight 0.99
Aug 26Signed up from a VPN addressweight 0.75
Jul 27Time zone far from the address's locationweight 0.63
Jun 12navigator.webdriver was truenever reads Real

Evidence fades toward half weight and stops counting after 90 days. Hard automation evidence stays.

Your decision, beside the model's.

Allow, block and review lists apply to accounts. A block always wins. The model's tier stays visible, so you can see where you and the model disagree.

Accounts / ListsExample data

acct_3d80Confirmed fraud (block)DecisionblockModelFake
acct_7f3eTrusted customers (allow)DecisionallowModelReal
acct_58b0Trusted customers (allow)DecisionallowModelSuspicious

An account on a block list reads Fake, with the list named as the decision, whatever else holds it. Your server acts on that decision.

API

Fetch the tier from your server.

Your server asks with its secret key. Account scores and tiers stay on your server: the browser's answer never carries them.

What your server gets back

Tier: Real, Suspicious, Fake or Unscored, from the account's score and the thresholds in force.
Decision: The effective decision. Your allow or block list decides when it holds the account; a block always wins.
Multi-accounting probability: Model estimate that one person runs this account and the accounts linked to it.
Linked accounts: How many linked accounts counted toward the score.
Hard automation evidence: True once the account showed hard automation evidence. It then never reads Real.
Tier thresholds: The two thresholds in force. The defaults are 0.50 and 0.75, and you can change both.
Calibration status: The model's calibration status. Every score is a model estimate.
First seen: When the account was first seen.
Last seen: When the account was last seen.

Spec

Tiers: Real, Suspicious, Fake, Unscored
Default thresholds: 0.50 and 0.75
Tier rule: Real below the first threshold; Suspicious from the first up to and including the second; Fake above the second
Device link window: 90 days
Network link window: 7 days
Links counted: At most 3 device and 3 network
Crowded devices: A device held by more than 9 accounts adds no device links
Evidence decay: Weight falls from 1 toward 0.5, losing half the distance every 30 days; sessions older than 90 days do not count; hard automation evidence never decays to Real
Scores: Model estimates, shown with their calibration status
Where scores go: Your server, fetched with your secret key

Limits

FAQ

The key your server sends with each session, usually your own user ID. Without a key, the session attaches to the device alone.

The score is compared with your two thresholds. Below the first is Real, from the first up to and including the second is Suspicious, above the second is Fake.

No. A shared network counts weakly, and at most three network links count toward any account.

A device ID that identical hardware could share is never counted as a strong link.

No. The browser's answer never carries account scores or tiers: your server gets them when it asks with its secret key. Send scores to the browser adds the session's scores to the browser's answer, never the account's.

Not during early access.

We have not published an accuracy figure. Talk to an engineer about testing it on your own traffic.

Apply for early access. We onboard a few teams at a time and work with each one directly.

Build account trust with us.

Join early access Talk to an engineer

Know which accounts are one person. And see every link that says so.

QUESTIONS THIS ANSWERS

Is this one person running ten accounts?

Why did this account turn Suspicious yesterday?

Which links did you ignore, and why?

Can my team overrule the model?

How it works

Set the line yourself.

Links drawn, and links refused.

Old evidence fades. Automation does not.

Your decision, beside the model's.

Fetch the tier from your server.

Spec

Limits

FAQ

What counts as an account?

How is the tier decided?

Can one busy office network flag everyone on it?

What about identical laptops?

Does the browser see the score?

Is it billed separately?

How accurate is it?

How do we join?

Build account trust with us.