AD CAMPAIGN DATA INTEGRITY 

Rebuilt the data model from the live schema instead of the document: which master table feeds each client category, which keys carry the association, and how a soft delete marks a row as gone while leaving it joinable.

Read every record three ways, querying the database directly in Python and comparing it against the same record from the API and the admin UI, rather than trusting any single surface.

Reconstructed the campaign state machine from live data and found the documented status mapping was inverted for two of the four states, which meant every draft count derived from it was describing the wrong set of rows.

Traced an asset field that never populated back to its source and showed the application was reading a table that had never held that data, while the real values sat in a different master table all along.

Separated genuine spec violations from legacy rows that predate the current rules, so the fix list reflected the policy rather than the history, and flagged the ambiguous ones instead of quietly picking a side.

Established what could not be verified at all: one feature's backing table was empty in the test environment, so the conflict and locking behaviour had no path to proof. Reported it as untestable rather than passing it.

Kept direct database writes out of permission and editability testing, because writing straight to the table bypasses the exact validation layer under test and turns a real check into a green tick that proves nothing.

Wrote every finding with the query and the request that reproduce it, so each one could be confirmed by someone else without rerunning my session.