Transparent system
How an AI tarot reading is built
An AI tarot product should explain which parts are deterministic, which parts are authored, and where a model may contribute. In Neural Tarots, the server controls the card draw and the spread contract. Authored card and position meanings provide a complete baseline; optional AI synthesis may connect those pieces without changing the cards.
1. The server draws from one canonical deck
All five oracles use the same 78-card deck: 22 Major Arcana and 56 Minor Arcana. Each card has a stable ID, English and Russian names, an image, a baseline meaning, a shadow, a reflective question, and a practical vector. The client cannot request preferred card IDs.
Card of the Day is deterministic for the user’s calendar date, so refreshing does not replace it. Interactive spreads use fresh server-side draws and enforce uniqueness within the one-, three-, or ten-card result.
2. Every spread position has a fixed job
A three-card spread does not show three interchangeable definitions. Context, resource, and next step each change how a card’s theme is expressed. Relationship Dynamics uses a separate contract centered on your agency, observable interaction, and consent.
The Celtic Cross has ten canonical positions shared by the draw engine, interpretation, API, interface, and tests. Its crossed center and right-hand staff are not decorative: the visual map and detailed reading must refer to the same order.
3. Authored meaning survives an AI outage
Card-specific, position-aware text is generated from the authored semantic layer first. For multi-card spreads, an authored synthesis can identify repeated themes, tensions, resources, and a practical conclusion. This prevents a provider timeout from erasing the basic result.
Eligible readings may add personalized AI synthesis. The model receives a fixed system contract, the minimal card and question context needed, a timeout, limited retries, and a strict response schema. User text remains data; it cannot replace the system instructions or choose the cards.
4. Privacy and safety limit the claims
A typed question is validated and screened on the server. It is not placed in the URL, fixed-category analytics, or ordinary logs. Optional saving is separate from drawing, and account ownership is verified before history, profile, or subscription data is returned.
The result is framed as reflection and entertainment. It cannot establish a future event, read minds, or give medical, legal, financial, or crisis instructions. High-risk language should lead to contextual safety guidance rather than a mystical push toward action.
Continue exploring
Related guides to spread selection, card meaning, and safe reflection.
