When a story here is wrong, it gets corrected in place and the page says what changed. We do not quietly delete a story and pretend it never ran.
The story keeps its URL and its position in the archive. Its text is replaced with the correction or the withdrawal, and the stale tags, entities and badges that came from the bad reading are cleared. A snapshot of the pre-correction row is kept off-site before anything is written, so the change is reversible and auditable.
The feed re-merges stories as new sources cover the same event. Without protection, the next merge would overwrite a human correction with the same machine output that produced the error. Corrected fields are pinned: re-merge, re-tag and re-classification all skip them. A corrected story stays corrected.
The error worth naming: some source pages have bodies that are mostly a list of other articles' headlines rather than an article. Read naively, the classifier can attach the wrong headline to the wrong page and produce a story whose cited source is about something else entirely. Ingest now detects that shape and defers those pages instead of classifying them, and any story already published from one is corrected by hand.
Mail the address on the site. Include the story URL and what is wrong. A correction that changes the substance of a story will say so on the page.
