How we read

Wherever this site says we read something, an AI agent read it.

Old books, and anything published in 1930 or earlier, were read in full. Those texts are free on Project Gutenberg and Wikisource.

Modern books were not read in full. For those, the agents read the author’s own free writing (articles, interviews), or free summaries and articles about the book. We cite a summary in one short credited line and never reproduce it.

We read each site’s terms before agents read the site, and we skip sites that do not allow it.

Every link shows its level, in words. Of the 72 links on the site on 11 November 2026:

1. 44

   Full text read

   The book was read in full.

2. 7

   Author’s own words

   The link rests on the author’s own articles or interviews.

3. 21

   Secondhand

   The link rests on a summary or an article about the book, and the credit names it, as in Secondhand, from Wikipedia.

A candidate link is a later book that may carry an idea. A reader gives it a label: Cites (the book names the source), Echoes (same idea, no credit found) or Conflicts (the book argues against it).

In the pilot, two readers built on different AI models, Claude Opus and Claude Fable, labelled all 74 candidate links without seeing each other’s answers. They agreed on 67 (91%). Cohen’s kappa, which corrects for agreement by chance, was 0.86 (95% range 0.75 to 0.95).

Both readers are Anthropic models. Their agreement shows the labels are consistent; it does not show they are right.

A third reader settled 6 of the 7 splits. On 1 link all three differed, and we dropped it.

A link stays only if the readers who back it say the evidence supports it. In the pilot that kept 61 of the 74 and dropped 13.

New evidence reopens a link. If the readers then disagree, or the evidence is too thin to hold it, the link moves to Still being checked. The first was William B. Irvine’s A Guide to the Good Life (2009) on Some things are up to you. On 11 November 2026, 6 links are there; the changelog lists each move.

Free summaries sometimes get a book wrong, so we measured how often. We checked 220 claims from summaries against the full text of the book.

The test needs the book’s full text, so it was run on older books, the ones we could read in full.

• All 220 claims 11 distorted 5%, 95% range 3 to 9%
• The 70 quotes inside summaries 0 wrong
• The 150 paraphrases 11 wrong 7%

Two rules follow from that. A quote inside a summary counts on its own. A paraphrase counts only if two sources we accept as evidence agree.

At 7%, one or two of the 21 secondhand links could rest on a distorted summary. That is why every secondhand link says so.

One summary site got 9 of its 41 claims wrong (22%). We use it only as a lead, never as evidence.

No person checked the quotes by hand: every quote was checked word for word against its source text by script.

Each idea page carries one quote per book.

A quote runs to at most 30 words from anything still in copyright anywhere, and at most 60 from texts that are free in both the US and the EU.

When a quote is held back, the link stays and the page shows no quote for it.

• We do not claim to have found where an idea began. First, on this site, always means the earliest source we can find.
• We do not summarise books. A book page says which ideas the book carries and where to get it.
• We show no book covers and no portraits.
• We carry no affiliate links.
• Echoes means we found no credit. It is not a claim about what an author knew or meant.

Anyone can flag a link. Write to hello@writtenbefore.com and say which link and why.

Every fix goes in the public changelog with its date.

A complaint about a quote takes the quote down within 7 days. The link stays only if other evidence supports it.