AI generation & verification policy

Articles are written by claude-opus-5-5 (Anthropic Claude), in Japanese and English at the same time.

• Each draft first goes through automated checks: whether every figure in the text matches its evidence (rounded figures within their rounding range), whether calculations add up, and whether the text has too many figures or formulas in it.
• Only drafts that pass are reviewed by a separate AI, in two automated stages: a Claude Sonnet pre-check, then a Claude Opus main review. Only a pass in the main review allows publication. (The two-stage review started on 19 October 2026; parts reviewed before then had a single Opus review.)
• The reviewing AI sees only the article and the data it was written from, not the writing AI's working.
• If the automated checks or the main review fail, the draft is fixed and rewritten once. If it fails again, it is not published and is recorded as rejected.
• The operator (a human) only decides whether and when an article is published, and does not touch the text.

Figures come only from data fetched automatically from official statistics and other sources, with the retrieval date recorded. The AI does not write figures from memory.

• Each claim is linked to its evidence. Evidence is numbered E-01, E-02 and so on, and a strip directly below each factual paragraph shows the number, value, source and retrieval date. The full evidence list is at the end of each article.
• The text uses easy-to-read rounded figures; exact values are in the evidence strip and list. Calculations behind policy amounts can be opened under "Show the calculation".
• Each series has a data page with its values by country and year and its source. Values from sources that permit redistribution can be downloaded as CSV; sources with redistribution limits, such as UN Comtrade, are left out of the CSV.
• Values that cannot be calculated from the data are marked as estimates or conjectures, and values the AI sets itself are marked as design values.

FFACTOAI VIEWPFORECAST

• The text is split into FACT and AI VIEW (and FORECAST), shown by both colour and label.
• The FACT parts do not assert cause and effect; they say only what the data shows.
• In the AI VIEW parts (diagnosis, policy package, view), the AI states its own judgement plainly. Figures and specific facts inside opinions are still checked against the evidence.
• Every opinion comes with an opposing view that directly challenges the cause the AI weighs most heavily.

• Each forecast is recorded when it is made: which indicator, for which country and period, whether it goes up or down, and when it will be checked.
• Forecasts are listed on the scorecard. On the check date they are compared with the actual data and marked hit, partial or miss.
• Misses are never deleted; they stay on the scorecard.
• Articles that stand at a past date and look ahead (the Back to the Future series) and forecasts about events use only data that had been published by then (annual statistics are treated as published in July of the following year).

Even in a bold "thought experiment" option, the AI never writes:

• proposals that infringe individual rights (such as coercion over having children, or penalties on people without children);
• discrimination by nationality, sex or similar grounds;
• support for or attacks on any party or candidate, or calls to vote.

The automated checks flag any wording that may fall into these lines. The main-review AI reads each flagged passage in context and records a judgement with a reason: a factual description, such as who pays for a policy, is fine; a proposal of penalties, coercion, discrimination or party support fails. An article is not saved unless every flag has a "fine" judgement with a reason.

The AI may know from its training data what happened later. Forecasts made from a past date or about past events carry a note saying so, in the article and on the scorecard.