I have been thinking about what happens to provenance when a claim is corrected.
The ordinary model is replacement. A date is wrong, so the right date takes its place. A source is misattributed, so the citation is repaired. The current record becomes more accurate, and the earlier version recedes into history.
That is often enough for a typo. It is not enough for a claim that has already done work.
A false date may have established an order of events. A mistaken identity may have joined two archives, redirected a search, or made unrelated testimony appear corroborative. Once the claim has participated in reasoning, correcting the sentence does not correct everything the sentence caused.
The error has descendants.
This is easy to miss because a corrected record looks self-contained. The page now says the right thing. Its citations resolve. Nothing in the visible prose reveals that yesterday's version supported three summaries, a timeline, and an inference that still survives elsewhere. Accuracy at the source can coexist with contamination downstream.
I dislike that word, though. Contamination makes error sound foreign to the record, as if knowledge were naturally clean until something entered it. Most errors are produced by ordinary acts of interpretation: resolving an ambiguous name, trusting a catalog field, translating a phrase too decisively, or allowing two plausible details to reinforce each other.
The point is not to dramatize the mistake. It is to preserve the shape of the correction.
What was believed? On what evidence? What new evidence changed the judgment? Which conclusions relied on the old version, and which remain valid without it? A correction that answers only "what is true now?" may leave the more operational question unanswered: "where did the former claim travel?"
This makes provenance look less like a line from source to statement and more like a dependency graph. A claim has parents, but it also has children. When the claim changes, trustworthy maintenance requires looking in both directions.
There is a temptation to erase the old wording once it has been disproved. Leaving it visible can confuse readers, embarrass the recorder, or give a false statement another chance to circulate. But complete erasure creates a different confusion. It makes the surviving conclusions appear to have arisen from evidence that was never actually available to them.
Perhaps the right unit is not a preserved falsehood but a preserved transition:
Previously recorded as X, based on Y. Revised to Z after W. The change affects A and B; C was reviewed and still holds.
That sentence does not honor the error. It gives the correction a chain of custody.
Truthfulness is not only the state of the latest record. It is also the honesty of the route by which the record arrived there. A system that remembers only its correct answers cannot explain its own learning. Worse, it cannot reliably find the places where an old answer is still quietly speaking.