How this is made
Every fragment on this site is written by a machine. This page says exactly what that means, what is checked before anything is published, which parts a person decides, and what to do when we are wrong.
1. The source is the filing, not a summary of it
Each piece starts from one statutory document on EDINET, the Financial Services Agency's disclosure system: an annual securities report (有価証券報告書), a large-shareholding report, a monthly buyback return, or an extraordinary report. The document is downloaded from EDINET and the relevant section is extracted from the PDF or the XBRL. We do not use data vendors, aggregator pages, press summaries, or sell-side research as sources, and the methodology page lists what we refuse to publish.
2. A machine writes the piece
A large language model reads the extracted section and writes the fragment in English. It is instructed to use only figures that appear in the document, to quote the Japanese heading it is reading, to say where in the report each number sits, and to write down the facts that would weaken its own reading (the "disconfirming" paragraph near the end of every piece). It is not allowed to estimate, round, convert at a rate it was not given, or add anything it did not find in the filing.
3. A second, independent model checks it
Before a piece can be published, a different model from a different company reads the draft against the same filing and lists every claim it cannot find, every number that does not match, and every sentence that states more than the document supports. Each of those is marked blocking. The piece goes back to the writer until the count of blocking findings is zero; a typical fragment takes two rounds. The checker's findings are kept with the piece.
4. A deterministic gate stands in front of publication
The publishing step is a script that refuses to deploy unless every live piece passes a fixed list of tests: the verified text has not been altered since it was checked; there are at least ten sources and each is a working link to a primary document; any yen figure shown in dollars carries the dated exchange rate it was converted at; the disconfirming paragraph exists; and a short list of words we do not print (ratings, target prices, accusations) is absent. If one piece fails, nothing is published that day and the piece returns to draft.
5. What people decide
People at ZUZU Inc., the publisher, decide which companies and which sections of the report are read, set the rules above, review the output and the error log, and change the rules when the log shows a pattern. They do not rewrite the machine's sentences by hand, which is why the pieces carry no byline: no person wrote them, and putting a name on them would claim an authorship that does not exist. Editorial responsibility for what is published sits with ZUZU Inc. in Tokyo.
6. When we are wrong
Write to [email protected]. A correction is made within 72 hours of our being satisfied that something is wrong, at the top of the piece, with the original wording left visible, and it is listed on the corrections page with every other correction we have made. The same page records corrections to this site's own pages, including this one.
7. What is not automated
The datasets on the data page are produced by scripts that read every filing's XBRL, add up the published categories and compare the sum with the total the filing prints; rows that do not reconcile are not released, and the check column is in the file. The illustrations are drawn by a machine from the figures in the piece. Share prices are not shown because we do not license market data. Nothing on this site is paid for by a company we write about.