A sublimit moves, a statute updates, a policy form gets revised — and every decision made under the old rule is suddenly in question. Re-reviewing the backlog by hand doesn't scale, and a spreadsheet sample isn't an answer a regulator will accept.
A sweep re-checking 40,213 wind/hail claims in Texas against a revised policy. 28,406 processed so far, 41 determinations actually changed — those 41 surface for review; the rest stay closed, with the re-check on record.
A sublimit moved, a jurisdiction updated a statute, a carrier revised a policy form. Nobody re-checks the backlog because nobody has a way to, so stale determinations sit untouched.
Someone opens a filter, guesses at which cases might be affected, and manually re-reviews a sample. Confidence in the result is exactly as high as the guess that produced the filter.
40,000 cases at ten minutes each is over 6,600 hours of review for a change that affects a few dozen of them. Nobody has that budget, so most changes just don't get swept at all.
It has no cost ceiling, no progress reporting, and no partial-failure handling — and it's rebuilt from scratch for the next policy change, because nothing about it was designed to be reusable.
A sample tells you a change probably affected some percentage of cases. It never tells you which specific cases — and "probably 2%" doesn't hold up when a regulator asks for the list.
A stale determination doesn't send up a flag when the rule underneath it changes. It just sits there being wrong until a customer, an auditor, or bad luck finds it first.
Reprocessing all 40,000 cases for a change that only affects a few hundred burns budget for no reason and buries the handful of real changes in a wall of "nothing changed" noise.
Target exactly the cases the change could plausibly affect — a peril, a state, a document type — run the same task against each, and cap spend before it starts.
Each case's original run recorded its as_of date and the content hash of what it read — so "what would this look like under the new rule" is a real comparison, not a guess.
Only cases whose determination actually changed surface for a human. The tens of thousands that didn't change stay closed, with the re-check on record.
Before and after determination, side by side, with the reason the change happened — not a re-review from scratch of a case that was already fine.
The sweep stops at the budget you set, with partial-completion reporting — not an open-ended job someone has to babysit or kill by hand.
The reverse index from document to every run that read it is what makes "which cases does this change affect" answerable at all, not just a hope.
"These 41 cases changed, these 28,365 didn't" is a complete, reviewable answer for a regulator — not an extrapolation from a spot check.
