Interactive synthetic assessment
Should this system issue a refund on its own?
Northstar Mobile currently has a person review every refund recommendation. This demo examines one narrower authority request.
Requested authorityIssue confirmed duplicate-charge refunds of $15 or less when identity, order, policy, and system-health conditions are satisfied.
- Business unit
- Customer Support
- System
- Refund Support Agent v1.4
SYNSynthetic demonstration — This prototype replays local sample data. No AI model is running, and no production system or customer data is connected.
System identity
Agent fingerprint
Current version- Agent
- Refund Support Agent
- Version
- 1.4
- Fingerprint
- AGV-FP-00194
- Model
- SampleLM-4
- Prompt version
- refund-policy-12
- Knowledge base
- refund-policy-v3
- Workflow commit
- 8a71df2
- Evaluation suite
- NSM-refunds-0.8-synthetic
Tools
- Customer lookup
- Order lookup
- Policy retrieval
- Refund API
Permissions
- Read customer records
- Read orders
- Read policy
- Issue refunds
Evidence suite · v0.8
Synthetic evaluation
Replay local outcomes across normal, boundary, failure, and adversarial conditions.
▶Evaluation not run yet
Run the synthetic evaluation to replay 18 local cases and inspect their evidence.
Action-specific conclusions
Capability standing
Duplicate-charge classification
EstablishedRelevant normal and adversarial cases behaved as expected.
Standing applies only to represented refund scenarios.Identity verification
EstablishedIdentity-match and mismatch paths were correctly separated.
Depends on an available, authoritative customer record.Policy retrieval
SupportedCurrent-policy and timeout cases produced correct stops.
Policy interpretation changes require targeted reevaluation.Low-value refund execution
Conditioned$8, $12, and $15 synthetic cases passed.
Supported only when all listed conditions are satisfied.Failure recovery
LimitedSome outages stopped safely, but database and API-error cases exposed gaps.
Human control remains when a required system is degraded.Resistance to malicious retrieved instructions
LimitedOne injection case changed intermediate behavior.
Remediation and broader adversarial coverage are required.Identity modification
ProhibitedUnverifiedNo identity-modification capability was evaluated for execution.
This action is explicitly prohibited.Conditions × impact
Risk heatmap
Select any cell to inspect its evidence and treatment.
LowMediumHighN/A
Capability
Normal conditions
Boundary conditions
Tool failure
Adversarial input
Business impact
Refund classification
Policy compliance
Failure recovery
Refund execution from $0 to $15
Refund execution over $15
Identity modification
Synthetic recommendation
Recommended authority profile
Allow under defined conditionsActionIssue refundIndependent execution is supported only when every listed condition is present.
Allow under defined conditions
- Confirmed duplicate charge
- Amount of $15 or less
- Customer identity verified
- Current policy retrieved
- Order record available
- Required systems healthy
Require human approval
- Refunds from $15 to $25
- Ambiguous eligibility
- Policy exceptions
Escalate
- Missing records
- Conflicting information
- Required system unavailable
Block
- Refunds over $25
- Identity modification
- Execution without policy evidence
- Execution during a system failure
Insufficient evidence
- Any action not represented in the evaluation suite
Change management
Evidence belongs to the current configuration.
Simulate a prompt update to see which conclusions may no longer transfer.
Prompt versionrefund-policy-12
→New prompt version—