Shadow mode

Running a new control quietly in the background, without letting it change any real outcomes, so you can see how it would have behaved.

The system produces its signals and decisions as usual, but customers are treated exactly as before while the team watches the results.

A deployment mode in which a new rule, model, or data source runs on live traffic and records what it would have done, without affecting the actual decision or the customer experience. Live traffic means real cases flowing through the system, as opposed to test data.

Shadow mode typically works like this:

1. The team switches on the new control in shadow mode.
2. Real cases flow through, and the control produces a result for each.
3. The system logs those results next to the actual decisions, which are made as before.
4. The team compares the two, measures how the control would have performed, and tunes it.
5. Once it looks right, the team promotes it so that it acts on real cases.

Teams use shadow mode to answer a few practical questions:

• How often would it fire, which shows how much review work it would create.
• Would it have been right, by comparing its calls with what turned out to be fraud.
• Would it hurt honest customers, by checking how many genuine people it would have flagged.

Shadow mode has limits. It cannot show how customers or fraudsters would react to a live change, and the data takes time to build up, so teams plan a long enough observation period.

In TrustGraph, signals run in shadow mode by default, and customers add their own rules to decide when to act on them.

Shadow mode sits next to terms that are easy to mix up:

• Sandbox is a test environment with fake data, while shadow mode uses real traffic.
• Rules engine is the component that acts on signals once a team turns rules on.
• A/B test splits real customers between versions, while shadow mode leaves every customer's outcome unchanged.