What reconciliation forbids

A differential never assigns blame. That needs a designated authority, and designation is organizational.
Agreement between two systems is evidence only to the degree that they are independent. Reconciling a copy against its source verifies transport and nothing more.
Two datasets with no common time cut cannot be meaningfully compared at all.
Fuzzy matching cannot be made exact. It is a classifier with a recall ceiling set by its cheapest component.

Escapes and prices

Escape	Price
Climb the localisation ladder instead of diffing everything	several round trips and a comparison protocol
Canonicalise to remove representation noise	a defined class of real differences you can no longer see
Sample	detection of rare defects (and rare correlates with expensive)
Diff on code change	2× compute, and complete recall only over your own change
Designate a system of record	a political decision; it resolves every later disagreement
Retire the dual write	a re-architecture; it removes the need for the audit entirely

Case, resolvedConstructed case · Ravi at Freshlane

A ₹4.2 lakh gap, and an argument every morning

The warehouse's monthly revenue differed from the payment provider's settlement by ₹4.2 lakh, and the two teams argued every night about whose number was right. Where it started: The reconciliation ladder →

The ₹4.2 lakh gap. Ravi climbs the ladder instead of loading everything:

Row counts per day: equal on all days except two.
Per-partition sums of amount by store region: the difference sits in one region (West).
Hashes of rows sorted by order_id inside that partition, compared as a tree: 37 rows differ.
A cell-level diff of those 37 rows: all are refunds that the warehouse recorded as positive amounts, and they add up to ₹4.2 lakh.

The whole search exchanged a few kilobytes, and it ran in minutes. He also fixed the combiner: his first row-hash used XOR, and a duplicated refund pair would have been invisible. He replaced it with addition modulo 2⁶⁴. Before hashing he normalised amounts to exact decimals with two digits, and wrote down the rule, so the team knows that errors below one paisa are intentionally not visible.

The nightly argument. The payment provider is declared the system of record for settled amounts; the warehouse is audited against it as a star. The legacy billing database is the real cause of the argument: orders are written to it and to the order service independently (a dual write). Ravi proposes a single writer, with legacy billing fed from the order service's change log. After that, the nightly audit is cut to a weekly check.