Project: simulate a receipt pipeline and prove its limits

Specify the recipe

Define merchants, receipts, line items, payments and refunds with stable keys and temporal rules. Generate 47 receipt events across three scanner sources, including discounts, duplicates, late settlements and cropped totals. Mark every event with a scenario and generator version. Relational checks should reject orphan refunds and impossible chronology before the clean fixture is published.

Exercise the pipeline

Feed the clean and intentionally broken versions through a parser or event pipeline. Assert idempotent handling of repeated receipt IDs, correction after a late event and reconciliation of subtotal, discount, tax and total. The broken fixture should quarantine its labeled exceptions rather than silently succeeding. Compare outcomes across a second set of unseen text templates to expose dependence on the generator’s style.

Evaluate utility and privacy

Choose an untouched real holdout if access is authorized; otherwise state that deployment utility remains unverified. Compare a baseline with a synthetic-assisted workflow by source and scenario, reporting support counts and errors. Test for exact content matches and near-copy candidates against any restricted training input. A disclosure review is required before a release containing learned outputs.

Package the evidence

Deliver code, seed, configuration, input manifest, dataset checksum, schema report, scenario matrix, pipeline assertions, holdout definition, utility table and disclosure findings. Force one failed gate and show that publication is blocked. Include a withdrawal procedure for a previously released version. A reviewer should reproduce one ordinary receipt and one late-refund correction from the same versioned recipe.

Implementation

def reconcile_receipt(receipt):
    expected = receipt["subtotal_cents"] - receipt["discount_cents"] + receipt["tax_cents"]
    return receipt["total_cents"] == expected and receipt["discount_cents"] <= receipt["subtotal_cents"]

Performance and operating cost

Reconciling one receipt is O(1); processing N events and indexing IDs is O(N) expected time and space. Real utility and disclosure checks add model evaluation and neighbor searches beyond this fixture-level assertion.

Common Mistakes

Do not present simulated results as measured production behavior.
Do not release rows from a failed privacy or integrity gate.
Do not omit scenario and generator versions from the packet.