Encyclopedia › How the Charts Are Computed › Why LLMs Get Chinese Astrology Charts Wrong

Why LLMs Get Chinese Astrology Charts Wrong

A language model is not a calendar engine

A failure everyone can reproduce

Give a chatbot your birth date and time and ask it to "cast a BaZi chart" or "cast a Zi Wei Dou Shu chart", and the result is frequently wrong: the day pillar off by one, the month pillar mis-assigned around a solar-term day, the major stars seated in the wrong palaces, the four transformations attributed to the wrong stars. During the AI fortune-telling wave of early 2025, users cross-checked these charts at scale — the same birthday could produce different charts in different conversations.

This is not a defect of any one model. It is the nature of the task: chart casting is calendar arithmetic plus a chain of table lookups, not language generation. Asking a language model to cast a chart in its head is asking it to mentally recompute two centuries of almanac.

Where exactly it goes wrong

First, the calendar data is not stored precisely. The year pillar changes at Lichun (立春) and the month pillar at each solar term — at the exact minute of the term, not at midnight of that day. Term moments differ every year (Lichun 2025 fell at 22:10), and no model reliably carries a minute-precision table of two hundred years of them, so births near a term boundary often lose a whole pillar.

Second, the boundary rules are ambiguous unless fixed by convention. Whether the late Zi hour (23:00–23:59) advances the day pillar is a school choice that must be made explicitly and applied consistently; leap months, hour boundaries and true solar time are all places where "slightly off" means "a different pillar". Models answer these edges confidently and inconsistently.

Third, star placement is a dependent chain. Zi Wei Dou Shu seats the Zi Wei star from the five-element bureau, then derives every other star from it — one slip corrupts the whole board. A language model generating token by token is not executing that chain; it is recalling text that looks like the output of the chain.

Fourth, the model never says "I cannot compute this". A wrong chart arrives with exactly the same confidence as a right one, which makes it more dangerous than an honest inability.

The right division of labour: engines compute, AI reads

The emerging consensus is a layered architecture: a deterministic engine does the computing (verifiable, reproducible — the same birthday always yields the same chart), and the language model only interprets the chart facts the engine produced. However freely the reading layer writes, it can no longer miscast the chart.

That is how this site is built: iztro computes the Zi Wei chart and lunisolar the BaZi chart, with three correction layers on top — a minute-precision solar-term table covering 1900–2100, an explicit late-Zi-hour convention (the day pillar advances at 23:00), and true solar time backed by a 259-city longitude table. The same conventions are implemented a second time, independently, in Python (serving the MCP endpoint), and the two implementations verify each other through 12 golden test vectors covering Lichun boundaries, terms falling inside a birth hour, both Zi hours, and leap months — re-checked weekly in CI.

How to verify any AI-cast chart

Four checkpoints expose most errors. Born within a few days of Lichun: check whether the year pillar respects the exact term moment. Born between 23:00 and 23:59: check that the day pillar follows one school consistently. Born in a leap month: check the converted solar date. Finally, give the same birthday to two independently implemented calculators — trust the chart only when they agree.

If you want AI involved, involve it correctly

Two sound paths. Cast the chart with a deterministic tool first and hand the chart text to the AI for interpretation — the export panel on this site’s results page exists for exactly that. Or connect the engine to the AI directly: this site offers an MCP server, so Claude, Cursor and other MCP clients can call the same chart engines the website uses. The AI then reasons over a computed chart, not an imagined one.