The source was recorded in two fields, and one batch job split them apart
Situation: our knowledge base is built on this principle: every concept node is maintained in one place, and everywhere else links to it rather than copying the content. Each node also records where its verification came from, in a field that points at the report used to check its sources. On 16 July 2026 a batch job added new-format verification fields to 56 old-format nodes, including a new source pointer; the old-format source field was left in place. The new field was filled, across the board, with the report the batch as a whole relied on, and was never changed to each node's own evidence file. From that day, where a node's evidence lives was recorded twice on every node.
Decision point: on 6 August 2026, while opening the report behind each pointer for a new shelf, we found a node on bounded rationality whose new field pointed at a report with zero hits on the concept. The report named in the old field held the real evidence, and both reports sat in the same folder.
There were at least four workable options. One, change the new field to the old field's value on the spot and carry on with the shelf. Two, shelve the node at its recorded tier and leave the pointer for later. Three, take the node off the shelf and report the finding, with the location of both reports, to the line of work that maintains the nodes, and leave the fix to them. Four, halt the whole shelf and first sweep the entire knowledge base for the same problem.
What we did and what happened: we took the third option; under our division of work, the shelf side reads nodes and does not change them. Three days later, on 9 August, the check on their side found the same defect in more than one place: on 9 nodes the new field carried the batch's report, and for 6 of them that report had zero hits on the topic. All 6 were repointed to the evidence file named in the old field, each one checked. On 5 further nodes the new field was already right, so the old field, which held more detail, was moved up into it and then removed. 11 duplicate fields were cleared in one pass.
The negative result: as of 25 October, 7 nodes from the same batch still carry both fields. Their new field originally pointed at a file name that does not exist. On 24 July it was changed to match the old field, but the duplication itself was never removed.
Take-away: if a format migration records the same fact in an old field and a new one, fill the new field node by node from the old one and check it, then delete the old field that same day. Filling one value across a whole batch writes an unchecked pointer onto every node.