Synthetic SmartCards and grant cohorts: testing without harming beneficiaries

Postal-financial products fail in production when teams test with toy CSV files. Grant weeks fail when sequencing, PoL gates, and MDU catch-up are rehearsed only on slides. The NamPost Innovation Sandbox treats synthetic data as a first-class product asset—shaped to Namibian income bands and tariff tables without copying real beneficiaries.

Why static mocks break down

Spreadsheet samples miss:

• Grant sequencing (days 1–8 pension, child/disability, BIG cohorts).
• SmartCard balance behaviour and 911-prefix conventions.
• Agency commission tables for NamWater, municipalities, and prepaid utilities.
• Edge cases: missed sequence day, PoL failure, insufficient float, idempotent USSD retries.

When test data is too clean, Observation Deck metrics look green while production breaks on day one.

What NamPost synthetic seeds include

High-utility sandbox namespaces provide:

• SmartCard accounts with realistic transaction graphs—not random numbers.
• Social grant cohorts with sequence_day, PoL status, and MDU flags.
• Agency billers with PSD-10 fee disclosure fixtures (DQ-03).
• NAMQR TLV samples for generate/scan/offline batch replay.
• Personas (Amuche, Ndishi, Elizabeth, David, Sarah, Johannes, Petrus, Maria, Thandi)—selectable in /dashboard/sandboxes/:id/personas.

Use cases with immediate value

G2P and PoL: Simulate pension credit after POST /api/v1/pol/verify; block disburse when PoL fails (INT-04).

Merchant NAMQR: Dynamic QR, step-up MFA, settlement reconciliation for Johannes.

Stokvel AIS overlay: Consent-scoped reads for Elizabeth—never grant eligibility from AIS (X-02).

MoF dashboards: Aggregate leakage views for David—no export of live GRN files without regulator role.

What “good” looks like

• Schema fidelity to OpenAPI contracts.
• Relational integrity across accounts, grants, and payments.
• Scenario libraries: normal, edge, adversarial.
• Versioned metadata: generation date, known limits.

Key lessons

1. Start from decisions (fee quote, grant release, QR settlement)—not from dumping fake rows.
2. Build repeatable INT suites; regulators compare runs over time.
3. Test failure paths early: PoL reject, consent revoke, rate limit, channel not supported.
4. Pair datasets with observability—fee drift (RK-01) must surface on the Observation Deck.

Synthetic data is not a shortcut around governance. It is how NamPost ICT lets you learn fast without touching live grant money.