The verdict nobody dares write: the old system was right

The €100 Lakehouse — 10/12

Every rebuild assumes the legacy is wrong. This one's written verdict: the 12 missing payments were real — "the platform lost them; the legacy didn't invent them."

That sentence had to be earned:

→ each of May's 12 payments appears in 13–30 independent legacy exports; 11 of 12 carry an issued invoice number; all have MB Way/Multibanco references
→ the 12 athletes, searched in the new platform by six identifiers — email, phone, name, DOB, NIF. Every apparent hit resolved to a different person
→ the legacy's known over-counting mechanisms, measured: zero in May

Then the referee: the fiscal invoice registry, external to both systems. Pull the 11 invoice numbers; confirm date, amount, client. Ruling recorded: for May, trust the legacy total. The new system is right in method and wrong in population.

When two systems disagree, the tiebreaker is a third record external to both.

July showed the stakes: −16.54%, five times worse than any month. The daily series: €185 on Jul 11, zero rows Jul 12–18, €65 on Jul 19. Not weak sales — ingestion had died. Seeding legacy numbers would have silenced that alarm forever.

The report's last line: €1,502 (17.6%) "attributed on evidence this investigation could not independently verify." Every unknown dated, quantified, bounded.

Have you ever had to write down, in a rebuild, that the legacy was right? What did it cost you to admit it?

Databricks · dbt · Airflow · Terraform · AWS

P.S. New tech post every Wednesday.

#100EuroLakehouse #DataQuality #EngineeringCulture