This week: teaching the AI books to stop making things up

Xingge draws Zi Wei Dou Shu (Purple Star Astrology) charts and explains them in plain English. The chart and readings are free. The paid part is a set of AI-written books: your life, your decade, your year, or two charts together.

This week the books got most of the attention, because I caught them inventing things.

What went wrong. The writer model gets a structured chart and a set of checked source passages for each chapter. Even so, a few chapters added a Ba Zi (Four Pillars) detail that wasn't in the material, named an illness the source never mentioned, or called next year "this year".

The fix. Each chapter now goes through a checker before it's accepted. The checker compares it against its own source material and flags anything the material doesn't support: invented pillars, illnesses that aren't in the source, fatalistic wording. The chapter is then rewritten with that list attached. The year book always speaks about the current year.

Also shipped:

• Ba Zi as a second lens next to the chart. It shows themes that complement the chart. It doesn't score one system against the other.

• Full sample books in English, Chinese and Japanese, so you can read one before paying.

Distribution: listed on a few free directories this week. Next I want to see which ones send anyone at all.

If you've shipped LLM-written long-form content, how do you catch the claims it makes up?