All articles Fraud & ATO

Mirror Mode Fraud: Test Rules Before You Block

Engineer pausing in front of a split glass panel in a dark studio, one half glowing blue and the other amber

Mirror mode fraud rules let you run a new policy against live traffic and record what it would have decided, while every request still comes back as allow. In Prynt Control, each new use case — signup, free trial, bonus, login, checkout, SMS, paywall — starts with a 14-day mirror period: signals are scored, a decision with reason codes is written down, and nothing is stepped up or blocked until you switch enforcement on.

The point is cost. A block rule that reads well in a spreadsheet can land on paying customers, shared office networks, or a browser update that shifts a device signal overnight. If you enforce on day one, your users become the test set and your support queue becomes the bug report.

This is the operational version: what to log, which numbers to read, and how to move from mirror to enforcement without guessing.

What mirror mode changes, and what it does not

In mirror mode Prynt returns a decision with a reason code and a suggested step — captcha, verify_email or card_or_passkey — but the response your server acts on stays allow. You build an evidence trail on real traffic without touching the user experience. How decisions, signals and products fit together is laid out on our platform page.

Mirror is not staging

Staging traffic is your team on a clean network. Mirror traffic is real devices, real VPNs, shared IPs and, if you have an abuse problem, real OTP-pumping scripts hitting your endpoint at 3 a.m. That difference is the whole value.

Your code still owns the outcome

Prynt does not block by itself: Control recommends and your server decides. The final call stays with the team that owns the customer relationship, and a mirror run produces a file you can review, not a switch that flips.

Pick one use case and define pass/fail first

Start with the narrowest surface that already costs you money — usually free_trial or sms. One use case produces a readable dataset; five at once produce a dashboard nobody opens.

Data-center aisle with blue light strips and a floating translucent panel above a keyboard

Write the three numbers you will compare

Before day one, commit to how many decisions would have blocked, how many would have stepped up, and how many of those belong to accounts that paid or converted recently. You already hold the third number in your own database; do not estimate it. Everything else is commentary on those three.

Log enough to argue about later

Each mirrored decision should keep the request ID, the decision, the risk score, the reason list, the suggested step and a visitor identifier. Pseudo-code, roughly:

event = {
  requestId, visitorId, useCase,
  decision, riskScore, reasons, suggestedStep,
  accountId, plan, outcome   // filled in later
}

The outcome field is the one most teams skip. Write the signup, trial or payment result back onto the decision record and the mirror run becomes a precision table instead of a log. Prynt Automate can push the same events to Slack with {{trigger.decision}}, {{trigger.riskScore}}, {{trigger.reasons | json}} and {{trigger.requestId}}, so the risk team reads decisions where it already works.

Read the results in three passes

Analyst reviewing an abstract pattern wall in a dark room lit by blue screen glow

First pass: the block bucket

Sample the decisions that would have blocked and watch a few sessions behind them. Prynt Replay masks every form field in the browser and again on the server, so a reviewer sees behaviour without seeing customer data. Look for returning users, paid accounts, and browsers that look odd but behave normally.

Second pass: the step-up bucket

For each risk band, ask which step costs least: captcha is cheap and annoying, verify_email is slower but verifiable, card_or_passkey is expensive and strong. The right answer differs between signup and login.

Third pass: the false-positive cohort

Pull the blocked candidates that are trusted devices, known customers or corporate networks. If a VPN or datacenter signal alone drives the decision, that is a policy bug, not a fraud finding.

Enforce in bands, not with one switch

Turn on step-up for the middle band first and leave the tail in mirror one more week. Keep the block band small, name each reason code in your support macros, and keep a way back: returning a use case to mirror should be one change, not a release. Re-run mirror whenever you add a rule, move a threshold or ship a checkout redesign.

What mirror mode cannot tell you

Mirror measures who you would have touched, not what you prevented. Abuse volume becomes visible only once enforcement starts, so pair the 14 days with a narrow, reviewable first week and read the numbers again. Reason codes are defined in our glossary, and the implementation detail lives in the docs.

Three glowing translucent glass bands above a black surface with a fingertip reaching for the middle one

FAQ

Does mirror mode block anything?

No. During the mirror period Control answers allow and attaches the decision, reason code and suggested step for your records. Blocking starts only when your code acts on the recommendation.

How long should the mirror period run?

Fourteen days is the default and it covers a full weekly cycle. If a promo, launch or seasonal spike is coming, mirror across that event rather than a quiet week — the abuse you care about usually arrives with the campaign.

Which use case should I mirror first?

The one already creating manual clean-up: repeated trial abuse, SMS pumping, or signups your team reviews by hand. Fix that surface, then extend the same policy pattern to login and checkout.

Ready to see what your rules would do before they do it? Start a use case in mirror mode at pryntid.com and take it live with the enforcement steps in our docs.

Try it free

Prynt is device intelligence with a free plan: 5,000 identifications a month, visitor IDs with confidence and bot detection — no credit card. Upgrade to Pro for VPN, proxy, tampering and behavioral signals.

Keep reading