Master data tools freeze entity keys to the last successful lookup window. Any later reference correction arrives after the AI pipeline has already committed the stale mapping as ground truth.
The pattern repeats every time an external registry publishes an amendment. Matching jobs that ran against yesterday’s file now treat the revised supplier as two separate legal entities because the update job only touches new inserts.
Procurement systems that rely on these keys inherit the split. Invoice matching rules then route payments to the older record while the newer one sits flagged for manual review that never occurs.
Model training sets compound the error. Feature stores pull the duplicate rows without a lineage flag, so gradient updates treat the same real-world supplier as two distinct counterparties with opposing payment histories.
Reconciliation dashboards report record counts instead of collision rates. The metric that would expose the fork sits outside the monitored queries because governance only audits fields that changed after the last approved schema version.
Fix the loop by forcing every reference refresh to trigger a retroactive key reconciliation pass before the next training cut. Run the pass against the exact timestamp the AI job used, not the current wall clock.
Without that timestamp lock, every overnight registry patch silently rewrites the training distribution while the platform still claims a single source of truth.