The fastest way to build an AI astrology feature is obvious: send chart data to a language model and ask for a profile. The result may read well, but several engineering questions immediately become harder.
AETHERCORE operates GenesisCore, a product that calculates Western astrology and Jyotish as separate routes and presents the results on screen and in PDFs. On September 9, 2026, we added a detailed Five Paths view to both paid plans. The new surface includes a radar chart, matched conditions, interpretive notes, reference people, and a one-page PDF.
The Five Paths judgment itself makes no new LLM call. It uses fixed rules and a prepared explanation dictionary. Existing Western and Jyotish analysis still uses AI; the boundary is local to this feature.
birth input
-> Western / Jyotish chart computation
-> fixed-rule matching
-> path-group and detailed-rule IDs
-> explanation and reference-data lookup
-> result tabs and one-page PDF
This is a public behavior model, not a claim that every internal component has been open-sourced. Its purpose is to make the responsibility of each output layer visible.
The radar chart is easy to mistake for a personality or aptitude score. It is neither. The values compare the number of matched path-family types with a 420-person development reference. They are not ability ratings, career recommendations, or success probabilities. A zero means that no current fixed rule matched for that path; it does not mean the person lacks the corresponding ability.
The page keeps three data layers distinguishable:
| Layer | Visible material | Responsibility |
|---|---|---|
| Judgment | placements, rule IDs, match conditions | what was calculated and matched |
| Interpretation | path names, meaning, rationale, practical contexts | how the symbolic structure is explained |
| Record | occupations, activities, comparison notes, sources | what was documented about reference people |
That separation matters more than the chart design. It prevents a calculated match from quietly turning into a claim about ability or causation.
The product dictionary currently covers 73 path groups and 386 detailed rules.
A path group is a reading unit that can contain several concrete routes. A detailed rule is the unit used to show which system, chart, position, or relation matched. Several rules may lead to the same social function, and one person may display several functions at once.
The interface preserves distinctions such as D1 versus D10 and the rule position versus a reference position. Display fields vary with the rule. The goal is a traceable answer to “why did this path appear?” rather than a decorative score alone.
The current dictionary size describes product coverage. It does not mean that all 386 detailed rules have been independently validated one by one.
With the same input, rule version, and dictionary version, the Five Paths result can follow the same route again. Sampling variation from an LLM does not enter the judgment layer.
A label can be followed back to a path group, a detailed rule, and the calculated placement. When something is wrong, we can ask whether the defect belongs to computation, matching, wording, or reference data.
An unclear description or weak source can be corrected without rewriting the matcher. A rule change can be reviewed against the smaller set of affected explanations and screens.
The product can show three different statements without merging them:
Those statements have different evidential weight. The UI and the data model should respect that difference.
The reference-person layer contains 509 edited person-by-path records in Japanese and English. It is not a dataset of 509 unique people, and it is not 509 independent experiments. The same person may contribute material to more than one path.
This may sound like a small wording detail, but it is a data-contract issue. Once marketing copy calls a join-table row a unique subject, every later interpretation becomes unreliable.
After release, the team checked 40 Japanese and English pages, eight paid-result variants, desktop and mobile rendering, public images, and consistency with the PDF sample. The PDF process included 1,522 layout inspections plus visual checks of representative outputs.
These checks support implementation, rendering, layout, and document consistency. They do not establish independent scientific replication, causation, personal ability, or future-prediction accuracy. A full production purchase run including paid AI generation and an actual customer email delivery has not yet been completed end to end.
An AI product does not need to use generative AI at every layer. Use flexible generation where interpretation benefits from it. Keep deterministic rules where users need stable labels, traceable conditions, and reviewable changes.
For a domain in which calculation, tradition, interpretation, and personal meaning easily blur together, that boundary is part of the product.