A publication that spends its time reading other people's disclosure should be able to survive its own. This is how the work is actually done, including the parts that are done by machines.
Statutory filings, and nothing else. Annual securities reports (有価証券報告書), quarterly and earnings releases (決算短信), large-shareholding reports (大量保有報告書), and the disclosure documents companies are required by law to publish. Most of it comes from EDINET, operated by Japan's Financial Services Agency, used under the Public Data License 1.0.
We do not use aggregators, data resellers, or second-hand summaries, and we do not use the paid databases that most English-language coverage of Japan is built on. When a figure appears here, it appears because someone filed it.
Where an official statistic is involved — corporate cash, GDP, disclosure counts — we take it from the body that publishes it: the Bank of Japan, the Cabinet Office, the Ministry of Finance, the Tokyo Stock Exchange. We say which one, and as of when.
Share prices. We do not license market data, so we do not display it, and we link out to venues that do. A publication that cannot show you a chart is forced to be interesting about something else, which we regard as a feature.
Running a Japanese filing through translation software is not journalism, and a site made of that would be worth nothing. What is scarce is not the words — it is knowing which paragraph of a 200-page document matters, and why the company phrased it that way. English disclosure is not required for annual securities reports under Japanese law; the Tokyo Stock Exchange's 2025 requirement covers earnings and timely disclosure, and only for its top tier. The narrative sections — the business review, the risks, the discussion of the balance sheet — remain, for most of the market, available only in Japanese.
So we read them, decide what is worth your attention, and write that down. The translation is a by-product of the reading, not a substitute for it.
Every fragment goes through the same four steps, in order, and a figure that fails any of them does not appear:
Every fragment ends with its sources, linked, so that you can do step 3 yourself. That is the point of putting them there.
Yen figures carry a US dollar equivalent in brackets. Conversions use a single rate, stated in each piece, applied to every year — including figures from the 1990s, which nobody was converting at anything like that rate at the time. Comparing the dollar figures across years therefore tells you nothing about the exchange rate. The yen numbers are the record; the dollars are a convenience.
This publication is assembled with automated tools, including large language models, at every stage: research, drafting, checking, and the drawings. We think you should know that, so we say it in the colophon of every page rather than in a disclosure nobody reads.
What automation is not allowed to do here is invent. The numbers are constrained to the verified sheet described above; the adversarial check exists precisely because a fluent machine will otherwise write a plausible figure. What it is allowed to do is draw — the illustrations are generated, deliberately imperfect, and not representations of anything real.
No article here is attributed to an AI author, because that would be a strange thing to do. The publisher is responsible for everything on this site, whatever tool produced the first draft.
None of it makes us right. It makes us checkable, which is a lower standard and a more useful one. A process this careful still ships errors — ours will be in the places we were most confident, because that is where nobody looks twice. The drawings are deliberately imperfect for the same reason the sources are listed: a publication that presents itself as flawless is asking you to trust it, and we would rather you verified it.
Within 72 hours, at the top of the piece, with the history left public. If you find an error — and eventually you will — write to [email protected]. Telling us we are wrong is a contribution, not a complaint.