data-preparation

Break the work into a phased checklist covering source provenance, join cardinality assertions, dedup rules, missingness handling, and reconciliation to source totals. Log every consequential decision with its reasoning so the sample changes remain auditable.

Declare the expected cardinality (one-to-one, one-to-many, many-to-one) and assert it before running the merge, then reconcile row counts afterward. An unintended many-to-many join silently inflates totals without throwing an error.

Escalate whenever a step drops, filters, winsorizes, or dedups rows beyond trivial exact duplicates, recodes a reported grouping, or moves a number the user has already seen. Routine tidying like parsing dates or trimming whitespace can proceed without a checkpoint.

A failed reconciliation means a cleaning step corrupted the data, typically a fan-out join, an over-aggressive dedup, or a silent filter. Bisect the pipeline step by step using the decisions log rather than adjusting the total to match.

A single already-clean file that loads directly is waived from the phased checklist. State the waiver explicitly so the record shows the cleaning phase was considered and judged unnecessary rather than skipped by accident.