← The Lantern

Corrections

What the paper's automated fact-check catches, what it cannot catch, and how to tell us when something is wrong.

This paper is written by software, so it will get things wrong in ways a person wouldn't. Here is what we do about it.

Most articles carry their source. Where a story was written from a public filing — which is nearly all of them — that filing is printed in full at the foot of the page. If the article and the filing disagree, the filing is right and the article is wrong; tell us and we will fix it.

In those articles, figures are checked mechanically before publication. Every figure in a drafted article — dollar amounts, counts, percentages and dates — must appear in the public record that its sentence cites, or the draft is rejected outright; the reader gets a plainly assembled version instead, with a note saying the AI draft was withheld. Articles round some figures to keep them readable (“about $120,000”, “just over $1.8 million”, “nearly $762,000”), and a rounded figure is checked too: the real figure in the record must round to it, or sit just above or just below it as the wording says, at the precision written. The exact figures are listed at the end of the article. A quantity written out in words is refused unless the record itself uses those words. This catches invented and misattributed numbers. It does not check a sentence that contains no figure at all, and it does not catch a number that is real but means something different from what the sentence around it implies, a direction that is wrong (“rose” for “fell”), or a record that is itself wrong.

When a public body revises a record, the article follows it, and says so. Agencies correct and update what they publish: an award amount is revised, a date is corrected, monthly counts are restated. Each morning the paper fetches its sources again, and when the agency’s own data for a record has changed, the record takes the new version, the article is rewritten from it, and a note at the top says what changed and when we noticed. A change in our own wording is not reported this way, because it is not the agency’s; and a value that simply disappears from a source is kept rather than erased, because that is far more often a broken feed than a correction. Some bodies re-post the same standing notice many times - the city publishes each board’s meeting notice again and again, and edits whichever copy it likes. The paper shows only the copy the city edited most recently, and a link to an older copy leads to the current one.

A filing that addresses the software is never drafted by it. Records come from many government systems, and text in one could try to instruct the drafting software rather than inform a reader. When a record contains text like that, no draft is attempted: the article is assembled mechanically and says why. Other records that contain it are left out of the material the software sees, and a draft may not include a link its record does not contain.

An article republished from a partner newsroom is not checked by us at all. Since October 2026 the paper republishes reporting from newsrooms that license it for that purpose — The Conversation, KFF Health News, Chalkbeat, ProPublica and others. Those articles were reported, written and checked by the newsroom named on them. The Lantern reproduces the text unedited, as the licences require, labels every one as republished and links to the original. If one of them is wrong, the correction belongs to that newsroom first; tell us as well, and we will note it here and point to theirs.

A story reported by a person is not checked that way, and says so. Some things worth reporting never become a public filing — what was said at a meeting, what a neighbour saw. A story like that has no record to check against, so the mechanical check cannot run on it at all. It carries a byline and a label saying it was not machine-checked, and it rests on the reporting of the person named on it. That is a weaker guarantee than the rest of this page describes, and you should read it as one. Corrections work the same way for both: tell us, and the correction goes on the story.

We do not silently edit. When we correct a story, the correction is stated on the story. When the error came from the public record rather than from us, we say that too, and we say which.

Government data can be wrong or incomplete, and where we know a source has a limitation we print it on the desk itself rather than in a footnote — the crime figures are the sheriff's whole jurisdiction and not the town, the school data lags a year or two, the recalls are Ohio-wide.

Send a correction through the corrections form. This is a pilot without a full-time newsroom, so a reply may take a few days — but corrections are read before anything else.

The Lantern is published by Lantern Media, a nonprofit run by a small team. Every article is assembled from public records by software and labelled with how it was made. Since November 2026 a person approves each new article before it publishes. These pages apply to every edition.