TAIYI LAB (IN FORMATION)
Correct, faithful, and open to testing.
Taiyi Lab is an internal research group at Terramontao, still in formation. Geomontao treats the chart as a mirror; the lab's work is to make that mirror true, and to make that claim testable.
WHY A LAB
Talking about charts is not the same as getting them right.
General AI models can talk about anything, but they work out a chart on the spot: the same birth details, asked another way, may produce a different chart. Quotations are hard to check word for word, and schools are not told apart.
Geomontao closes half of that gap with deterministic engines and retrieval from the classics. The other half is method, expert knowledge and reproducible versions, which have to be built up through our own data and evaluation. That is the lab's work.
The Great Change is the state before qi appears. — Liezi, ‘Heaven's Gifts’
THREE LAYERS
Accuracy has three layers, each reviewed on its own.
We do not fold them into one vague accuracy score. The first two are what a person can feel today. The third is a long-term research question, and Geomontao's value does not depend on its outcome.
Correct · charts and calendars
Deterministic engines cast the chart; reference charts are signed off blind by two advisers and checked item by item.
Faithful · readings and sources
Every sentence of a reading traces back to the chart's facts and the classics, with schools labelled separately.
Open to testing · what readings claim
Blind matching studies and pre-registered prospective studies, with thresholds and statistics registered before anything is unblinded.
HOW WE TEST
Set the threshold first. Look at the result second.
If a sample is too small, we report that and claim nothing. Positive and negative results are both published.
Sealed evaluation
A third party holds the test set and thresholds are locked before results are seen. Rewrites and splits of one case never cross between training and test sets.
Blind adviser review
Advisers score separately, then review together. Only tasks that reach agreement between scorers are counted.
Blind matching
Charts are matched against de-identified life-event records, to see whether results rise above chance.
Prospective registration
With consenting users, each judgement is recorded when written, held by a third party and opened only when the time comes.
This is a research plan. The lab is still being formed: enrolment has not begun and no test results have been published. When research involves user data: separate consent, the minimum necessary, de-identified, and withdrawable at any time. Public figures' cases are used only to evaluate charts and readings, never to test what readings claim.
TWO TRACKS
Research does not hold back the product, and the product does not wait for research.
Today's Geomontao runs on the production track: deterministic chart engines, retrieval from the classics, registered large-model services and word-by-word quotation checks. The Taiyi model is on the research track, and reaches the product only in a controlled way, after sealed evaluation and registration.
Everything the research track produces returns to production first: reference charts correct the engines, and rules and test sets improve retrieval. Even if the model itself does not succeed, the data, the evaluation and the engine improvements stay in the product.
SCOPE
Seven systems, near to far. Bazi comes first.
Birth charts and question-based systems need different evaluation, so they are planned separately. Passing one stage does not approve the next.
Stage one
Bazi, end to end: sources, method, tools, model and evaluation. Ziwei is limited to reference charts and engine checks.
Stage two
Ziwei and Qizheng, also birth charts, planned according to stage one's results.
Stage three
Qimen Dunjia, Da Liuren and Meihua Yishu: question-based systems, with evaluation designed separately.
Research list
Taiyi Shenshu: literature and calendar research only.
No system is used for readings about nations, public events, political figures or disasters; readings are only for the person themself, with their consent. This is written into each system's method and refusal rules.
THREE GATES
Each gate can be checked. Each gate can stop the work.
What comes out are documents, data and reproducible results. If a gate is not passed, the work stops there.
Set the rules
Freeze the Bazi method and measure agreement between scorers. Task types that fall short do not enter training.
Run the comparison
Report what retrieval, tools and weights each contribute. Two rounds without stable gains end weight training.
Fix the version
Sealed evaluation, blind adviser review, reproduction in a clean environment and a model card. Without a pass, it stays an experimental version outside the product.
A claim of accuracy should be open to testing.
For research collaboration and scholarly exchange, start with the contact page.