Build a small decision on a test step before adding a full fraud policy. This exercise maps a known input, produces a signal, attaches a decision strategy, and verifies the result in Investigations. You then connect that result to an application or edge action.
You need a journey step that already produces events. If it does not, complete Your first successful event first. Work in a test node and use a controlled test input.
Understand the parts
| Component | Job | Example |
|---|---|---|
| Mapped attributes | Describe the current interaction | Account ID, operation, search query, transaction amount. |
| Features | Calculate useful context, including historical aggregates | Recent attempts associated with an identifier. |
| Rules and models | Evaluate that context and produce signals or scores | A rule marks an interaction for review. |
| Decision strategy | Select the overall assessment result for the event | Pass, Challenge, Alert, Review or Reject. |
| Action integration | Make something happen based on the assessment | Your application holds a transaction, or a configured proxy action intervenes. |
The step's Workflows tab connects the executable files to the step. Dependencies determine when their inputs are available; independent work may run concurrently, while dependent work must wait. See Dependencies & execution order.
1. Choose a predictable test input
Use a mapped attribute with a value you can deliberately change. For example, the MCP catalogue-search example maps its search text to custom.general_purpose['query'].
For that test step, use this expression to recognize a deliberately chosen test value:
custom.general_purpose['query'] = 'review-demo'
This is a wiring test, not a production fraud rule. On a different step, substitute an attribute that your integration actually supplies. Verify its value in Investigations before writing a rule against it.
2. Create and attach a ruleset
Create a .rules file using the rules editor. Add a condition rule with:
| Setting | Test value |
|---|---|
| Signal | DOCS_REVIEW_TEST |
| Condition | custom.general_purpose['query'] = 'review-demo' |
| Category | DOCS_TEST |
Open the journey file from Explorer, select the test step, then open Workflows → Models. Add an operation that executes this ruleset and give it a unique name in the step.
Make the operation wait for the mapping that supplies the query. For the JSON-RPC request-body example, that means the request-body data must be processed before evaluation. Resolve the editor's diagnostics. See Rules & signals for rule authoring and data mappings for the input side.
3. Create and attach a decision strategy
Create a .strategy file with the strategy editor. Configure two ordered rules:
- When the
DOCS_TESTsignal category is present, terminate strategy evaluation with Review. - Otherwise, use a final always-true rule that terminates with Pass.
The category condition is:
has(outcome['CHAMPION'].models.signal_categories, 'DOCS_TEST')
Here, terminating the strategy stops further strategy-rule evaluation and sets the assessment result. It does not terminate the HTTP request.
In the step's Workflows → Decision Strategy section, select the file you created. A step has one decision strategy. Put specific conditions before the fallback because a terminating rule determines the result when it matches. See Decision strategy.
4. Deploy and check two cases
Follow Deployment lifecycle to review, commit, build and deploy the configuration to the test target. Confirm the build succeeded and the intended deployment is active.
Run the same interaction twice:
| Input | Expected signal | Expected strategy result |
|---|---|---|
Query is review-demo |
DOCS_REVIEW_TEST |
Review |
| Query is a different ordinary value | Test signal absent | Pass |
Open the resulting events in Investigations. Check the mapped query, the model's signals and categories, and outcome['CHAMPION'].decision_strategy.result. If the result is missing or undetermined, check that the strategy is attached to the matched step and that the deployed version contains your changes.
The result should change predictably when the input changes. If it does not, inspect the input and workflow dependencies before changing thresholds.
5. Connect the result to an action
Decide what Review means for this integration: a hold in your application, a case for an analyst, or another configured response. Follow From risk assessment to action and test that behavior separately.
For a pre-origin block, ensure the necessary assessment data is available before forwarding. Do not make that action wait for a response that cannot exist until the request has already reached the origin.
Extend the decision with real context
Replace the test condition with your actual policy. Add features using the identifiers, event filters and time windows relevant to the behavior you want to evaluate. Use the feature editor to configure them, then inspect their actual values and missing-data behavior before using them in a rule.
You can also use model scores, labels, and Marketplace enrichment. Make each consumer wait for its required inputs. Test representative legitimate activity, unwanted activity and unavailable-input cases before enabling enforcement. Use dashboards to monitor the resulting signals and decisions as you expand coverage.