How we write and check what you read

Every record on this site starts as a government document: a bill, a hearing, a file the Department of War released, a report an agency published. We do not write from memory, and we do not write from other people's articles about the documents. Each page links to the primary source, and you are meant to click it.

Some of the released files are scans of old paper. When the text layer in a scan is unreadable, we have a machine read the pages and we label the result as a machine transcription, because that is what it is. A transcription is never used as the basis for a quotation, and the official file is always one click away so you can check it against the original.

Some summaries on this site are written by Claude, an AI model. We think that is fine, and we think hiding it is not. So every passage an AI wrote carries a label saying so, and the model used is recorded with the record.

The AI is never asked what it knows. It is given the text of the source document and asked to explain what that document says. It is told, in the same instructions that govern the fact-checker, that it may not introduce a name, a date, a place, a rank, a speed, a distance, or a measurement that does not appear in the source. That single rule is the reason most drafts get rejected.

A draft is not published because it looks good. It is published because it survived each of the following, and it is quarantined if it fails any one of them.

• A second AI tries to break it

  A separate pass reads the draft against the source and looks for claims the source does not support, and for opinion, spin, or speculation. If it finds any, the draft fails.

• Every number is traced back to the source

  Code, not a model, extracts every year, every figure of three digits or more, every dollar amount and every percentage from the draft, and checks that each one appears in the source text. A number that cannot be found there is treated as invented, and the draft fails.

• Every quotation is found in the document before you see it

  Where an answer is followed by a quotation, that text was located in the source document by an exact search, and the words you read are cut from the document itself rather than copied from what the model typed. A quotation that cannot be found in the source is thrown away and no quote is shown. When we first ran this check over the 171 answers already published, 32 of the quotations offered for them could not be found in their source and were thrown away. A quote is the one thing that cannot be approximately right.

• Reading level

  We aim for a tenth-grade reading level and enforce a hard ceiling of grade twelve, computed in code. A draft written above that ceiling fails.

• No hype, no advocacy, no first person

  A word list blocks the vocabulary that makes a document sound like a rumour: shocking, bombshell, smoking gun, undeniable, cover-up. It also blocks the tics of AI writing, like delve and tapestry.

A draft that fails any check is marked needs_review and is not published. The page keeps whatever was there before, which is usually the official government summary. Failure is the common outcome for long, multi-topic bills, and we would rather show you the government's own summary of one than our confident description of something we did not read closely enough.

By default the pipeline makes a single attempt. It is not given a second try to talk its way past a check it just failed.

• We do not re-host the government's files

  We link to the official copy. A mirror of our own would add a second version of a public record for a reader to have to trust, and it would add nothing.

• We do not predict release dates

  The Department of War has not published a schedule. Where we do not know something, the page says we do not know it, rather than filling the space.

• We do not evade a website that has told us to stay out

  The official UAP sites block automated access. We respect that, take the file text from open public mirrors instead, and link you to the government's own page.

• We do not quietly correct the record

  When we get something wrong, the fix is a change to the pipeline, and the pipeline is described here.