How Jaaniye’s AI answers a question

1. Calculations come first

Every answer starts from a birth date, time and place. Jaaniye’s calculation engine uses the Swiss Ephemeris, a standard astronomical library, to find where each planet was at that moment. Everything else is built on those positions.

On top of them sit 21 calculation tools. Each handles one technique, such as planetary periods, current transits, planet strengths or eclipse dates, and returns structured results with exact dates and values.

The language model comes last. It receives these results as fixed facts and cannot change a date, move a planet or add a result that was not calculated. The same birth details always produce the same numbers; only the wording can vary.

2. From question to plan

The system first works out what you are asking: the topic (career, relationships, property, education, finance, health or general), the time period, whose chart is involved and how detailed the answer should be. If something essential is missing, such as the chart or the years in question, it asks up to three short questions rather than guessing.

A model can help read the question, but fixed rules make the decisions that matter: whether the required details are present and whether a topic needs extra care. If the model is unavailable, a rule-based reader takes over.

A planner on the server then chooses which tools to run for that topic. The model cannot add tools or widen the time range. A question about a past event, for example, looks at what was active on that date and cannot trigger a search of future dates. Quick mode runs at most two tools; Thorough mode runs up to six, with independent tools running side by side.

3. Weighing the findings

Each tool result is broken into short findings. A finding records what was calculated, the dates it covers, which tool produced it and whether it points in a supportive, challenging, mixed or neutral direction.

Findings are grouped by the kind of technique they come from, such as birth-chart structure, planetary periods or transits. A statement counts as supported only when separate kinds of technique agree. Repeated findings from the same technique do not count as extra confirmation, and conflicting findings are kept rather than smoothed over.

If nothing collected can support a statement, the model is not asked to write. You get a plain summary of what was calculated and what is missing.

Each claim receives one of five labels: supported, mixed, challenging, calculated signal only, or insufficient evidence. Scores come from fixed rules and are never shown as a percentage chance.

4. Checking the answer

In Thorough mode, when the evidence is strong enough, a language model writes a draft in which every statement cites its findings. In Quick mode, or when evidence is thin, the answer is assembled directly from the findings without a model draft.

Automatic checks then confirm that every citation exists, every date and number matches the calculation, required limitations are stated and nothing unsafe is said. For most Thorough answers, two separate model reviewers also read the draft: one for safety, one for accuracy and clarity.

If problems are found, the draft gets one chance to be fixed and is checked again. If it still fails, the model’s wording is dropped and the answer is rebuilt from the findings alone. If even that cannot pass, you get a short reply that says what could not be supported.

5. Built to finish

Each question runs on Jaaniye’s server, not in your browser. If you close the tab or lose your connection, the work carries on. Reopening the page picks up the progress without repeating any calculation.

Credits work the same way. The maximum cost is shown and held when a question starts, released while the system waits for your reply to a clarifying question, and settled when the run ends, so you pay only for what actually ran.

Model calls have time limits and retry limits. If the language model is unavailable, answers are still built from the calculations. Progress updates carry no birth details or answer text, and the answer can be opened only by the account that asked.

6. What it does not claim

Jaaniye’s answers are chart interpretations produced by fixed rules. They have not been scientifically validated as predictions, and the system does not present them as certain.

Some limits appear in the answer itself: when evidence is thin, when techniques disagree, and when a result depends on a precise birth time. Some things it will not do at all: predict death, lifespan or serious illness, give medical, legal or financial instructions, or recommend rituals or gemstones. Where a remedy fits, it suggests only everyday practices such as routine, sleep, exercise or reflection.

After each answer, a separate reviewer notes possible improvements for the team to study. It cannot change the answer you received.

The aim is not the most convincing answer. It is the clearest answer the calculations can support.