How each edition is made
Every edition is produced the same way, on a schedule, with no manual step. This page describes that process in full, including the parts that constrain what the digest can tell you.
1. Headlines
We read a fixed set of public RSS feeds: gold-price and central-bank news via Google News, Mining.com, USAGOLD, the Wall Street Journal markets feed, and Federal Reserve monetary press releases. From each we take the headline, the link, the publisher and the timestamp — never the article body. Stories carried by more than one feed are collapsed into one, and stories already covered by an earlier edition are dropped so the same news is not reported twice.
2. Price
The price block is captured once, at publication, from the XAUTicker spot API, and then frozen onto that edition. An archived edition therefore shows the price it was written against, not today's price. When the feed reports that it is running in delayed mode, the digest is told so and will not describe the price as real-time.
3. The write-up
The digest is written by a large language model (Anthropic's Claude), working only from the two inputs above. It is instructed that it may not introduce a number, institution, quote or event that does not appear in the material it was given; that it has seen headlines and not articles, and must write at the level a headline supports; that it may not forecast; and that specific claims must be attributed by name to the publisher who made them. Each section records which headlines it drew on, and those are the links printed beneath it.
The output is checked before publication — structure, and that every source reference points at a headline that actually exists in the edition. Anything that fails that check is not published.
4. Publication
The edition is written to disk and served at a permanent URL. Nothing is regenerated afterwards.
Limits worth knowing
- Headlines only. Nobody — human or model — read the underlying articles. A headline can be misleading about its own story, and that error would carry through to us.
- Feed bias. The digest can only reflect what our fixed feed list published. A story no listed source covered does not exist as far as an edition is concerned.
- No human editor. There is no journalist checking each edition before it goes out. The guardrails above are structural, not editorial.
- Machine-written prose. Every digest is labelled as such. When the writer is unavailable, the edition publishes as a plain headline list and says so rather than substituting something that looks edited.