These are cases where the source is wrong or ambiguous, not parser bugs. That distinction is
the reason this list is short and the reason it exists at all. When our reading of a record is wrong in a way
that generalizes, we fix the extraction rule and re-run it across every affected case — that is not a
correction, it is a repair, and it leaves no entry here.
A correction is the other thing: a single record where the document itself cannot be read the way it is
written, and a person had to decide. We do not teach the parser a rule from one of these,
because the same pattern is read correctly in thousands of other cases and a rule drawn from an exception
would break them.
Each correction is applied as an overlay on top of the parse, so it survives a re-parse, and the originally
parsed value is kept beside it. The disagreement between the two is the auditable part.
How that works, and what it means for the accuracy figure →