Use rules where they fit
Deduplication, reconciliation and preservation of human notes need reproducible rules. I kept these checks in code rather than generative AI.
Case 04 · Operations
Collecting more data was different from making it useful. I reconciled records and distinguished corrected information from newly obtained information.
Deduplication, reconciliation and preservation of human notes need reproducible rules. I kept these checks in code rather than generative AI.
Collecting a list during sending left records missing. I added collection after completion and reconciled the results by invitation.
Keep corrections, new contacts and recovered records distinct. Deliver the results in separate tabs for the people responsible.
Separate incidents—overlapping runs and mismatched delivery results—led to checks on both writes and completion.
Scheduled and manual runs wrote duplicate rows. Cleanup preserved rows containing human-entered information.
Add an execution lock, protect review status and notes, and roll back if the before-and-after check of reviewed entries differs.
A successful request did not always mean successful delivery. Re-query delivery counts before reporting completion and reconcile records again.
These are separate improvement outcomes. They should not be summed or treated as a partition of the full dataset.
Source rows and individual creators are different units.
Contact reconciliation and verification records. Source rows, unique creators, changes, and newly obtained records are counted separately.
Verified recovery records. Missing-record recovery is a separate improvement from contact reconciliation.
Operating and decision records describe concurrent-run duplicates, execution locking, preservation checks and delivery re-querying. These examples show operational safeguards; private selection rules and implementation details are omitted.