Benchmarking Kepler Across 1.5 Million Real E-Commerce Records: What Clean Code Misses

In operations intelligence, synthetic data is clean and comforting—but real-world enterprise data is chaotic, inconsistent, and full of edge cases. To prove Kepler's deterministic audit engines in the wild, we audited over 1,500,000 real-world commercial records across 7 open-source enterprise datasets spanning UK/EU retail, Brazilian logistics, Amazon/Flipkart catalogs, payment chargebacks, SaaS subscriptions, and Customs Bills of Entry.

The Real-World Data Breakdown (1.5M+ Records)

• UCI Online Retail II (1,067,371 rows / 53,628 orders): Audited for sub-COGS loss, discount stacking collisions, serial customer wardrobing, and EU IOSS VAT ceiling breaches.
• Olist Brazilian E-Commerce (100k orders / 112k items): Reconciled courier delivery transit delays, 3D volumetric packaging creep, split shipment routing bugs, and multi-installment gateway fee inflation.
• Enterprise Marketplaces (Amazon 10k & Flipkart 20k): Scanned 20,000 catalog variants for mandatory 6-digit HSN codes / 0g weight anomalies, and 10,000 Amazon ASINs for algorithmic Buy Box suppressions caused by off-platform price desyncs.
• Global Enterprise Operations (Chargebacks, EXIM Customs, SaaS): Detected 333 expiring bank chargeback deadlines, 115 unclaimed Free Trade Agreement tariff notifications, and 250 terminated employees retaining active paid Google Workspace seats.

What Synthetic Unit Tests Miss (And Real Data Teaches Us)

1. Negative Quantities as Cancellations & Returns

Synthetic fixtures represent returns with dedicated booleans (isReturn: true). Real ERPs (SAP, Tally, NetSuite) encode returns as negative numbers in the Quantity column with 'C' prefixed invoice numbers (e.g. C489449). Without native signed quantity normalization, parsers crash or miscalculate gross sales.

2. Promotional Discounts as Line-Item Pseudo-SKUs

Modern D2C stores store discounts at the cart level, but legacy POS and wholesale exports insert discounts as discrete line items with special StockCodes (D, Discount, Manual). Normalizers must scan both line-item SKUs and checkout-level coupon arrays.

3. Multi-Installment MDR Fee Creep

Merchants assume a flat 1.9% or 2.0% gateway fee, but real checkout installment plans (2x to 12x credit card installments) inflate effective merchant discount rates (MDR) to 3.5%–5.2%, creating substantial settlement variances.

Zero-Crash Scale & Performance

Running the full 16-engine suite across 1.5 million rows executes in under 12 seconds using pure TypeScript deterministic algorithms, with zero V8 heap overflows and 100% type safety.