Source & Content Policy

We read publicly available news articles and official releases, extract the facts they report, compare reports of the same event, resolve the place, and add our own assessments — severity, what imagery could show, market relevance, related events. One record is one real-world event, not one article. How the records are produced is described in the Methodology; the sources are listed under Data Sources.

We do not sell articles. We do not train models on them. We do not give article text to users.

• Public pages only. We do not log in, pay, or get past paywalls, geo-blocks or bot challenges. A page that answers with one is recorded as not retrieved, and a site that keeps refusing is left alone.
• robots.txt. When we fetch articles from news publishers, we follow the site's robots.txt rules for all crawlers (User-agent: *), re-read daily. If a robots.txt cannot be read, we keep fetching and try again later.
• Reservations that apply to everyone. We follow machine-readable reservations addressed to all automated use: /.well-known/tdmrep.json, the tdm-reservation header and meta tag, and noai. An article covered by one is not collected.
• Notices. When a publisher or rights holder tells us to stop, we stop (see below).
• Government open data. Official releases that a government publishes as open data under a licence permitting reuse are collected under that licence, with a named fetcher and a handful of requests a day. Today that is one source: the daily activity reports of Taiwan's Ministry of National Defense, registered on Taiwan's open-data platform under the Open Government Data License; the site's published terms place no restriction on automated collection.

In the first day after these rules took effect (11–12 October 2026) they stopped about 2.6% of the articles we would otherwise have fetched.

We do not apply rules written for other companies' crawlers to ourselves. Many publishers disallow named crawlers such as GPTBot or CCBot while allowing everyone else. Measured over seven days in October 2026, about 41% of the articles we use came from sites with such rules.

Those rules are addressed to systems that copy articles to train models, or to answer questions with the article's text. We do neither. We extract facts, link to the article, and do not reproduce it. A robots.txt rule is, by the standard's own design, a rule for the crawler it names, with a default group for all others — and we follow that default group.

Our article fetcher identifies as a browser. Sites that reject anything announcing itself as a bot would otherwise be lost, including outlets that are the only reporting for some regions. The consequence is that a site cannot single us out in robots.txt. A site that wants us to stop can tell us, and we stop for that site.

• Article text is working material. It is kept for analysis and deleted 14 days after the event it belongs to stops receiving new reports. Text of articles we decided not to use is deleted after 7 days. Search excerpts used to corroborate a location are deleted on the same rule.
• What stays: the facts we extracted, the link, the outlet, the publication time, the article title, a hash of the text, and — as internal evidence for a stated figure or claim — the short sentences it was taken from.

• Summaries and assessments are ours. Each summary is compared with the source article for copied wording. One that follows the article too closely is rewritten, and the rewrite is used only when every number and name in it survives; otherwise the summary is written again from the extracted values. The comparison is a measure, not a guarantee: a small share of summaries still sit close to their source.
• Titles are short statements of fact. For an event known from a single report, the title is often the same as, or close to, that report's headline.
• Sources are shown as outlet, time, article title and a link to the original. We do not show article text or search excerpts. A short phrase quoted from a report (at most 160 characters) may appear beside the fact it supports.
• Delta Agent reads search results to answer a question and cites the articles by title and link; the excerpts it read are not shown.

We set out our reading plainly. It is our own interpretation of how the law applies to what we do; it is not legal advice to anyone else.

• Facts are not protected by copyright. That a strike hit a named refinery on a given day is a fact. Japan's Copyright Act states that miscellaneous reports and news of current events that merely convey facts are not literary works (Article 10(2)).
• Analysing text to extract information is a permitted use. The Act allows works to be used for information analysis — extracting and comparing information from a large number of works — where the purpose is not to enjoy the expression (Article 30-4), and allows minor incidental use when providing the results of such analysis (Article 47-5). Off-Nadir Delta operates from Japan and relies on these provisions. We consider that our use does not unreasonably prejudice rights holders: a record does not substitute for the article — it links to it and does not reproduce it.
• Short quotations beside a fact are made within the bounds the Act sets for quotation (Article 32): limited in length, marked as the report's words, with the source linked.
• In the European Union, the text-and-data-mining exception applies unless the rights holder has expressly reserved the use by machine-readable means (Directive (EU) 2019/790, Article 4). We read and follow the reservation signals listed above. We do not parse website terms and conditions automatically; where a publisher's terms reserve these rights, tell us and we will apply the reservation.
• robots.txt is a request to crawlers, grouped by crawler name, with a default group for all others; the standard itself says the rules are not a form of access authorization (RFC 9309). We follow the default group for news publishers, as described above.

A commercial licence covers Off-Nadir Delta's own output — event records, titles, summaries, assessments, and exports of them. It does not include any right to reproduce or redistribute the source articles. For that, follow the link and the publisher's own terms. See the Terms of Service.

Write to support@offnadir-lab.com and tell us the site or the article, and what you want. We can:

• stop collecting from a site;
• delete the article text we hold from it;
• remove an article from the sources of our records.

We act on the specific material identified and keep a record of what was asked and what we did. If an article has been corrected or retracted, tell us and we will update or remove it.