Supastro Verification

How we verify a reading

What is checked between the AI writing a reading and you seeing it.

What gets checked

A reading on Supastro is made in two parts. First your chart is calculated. Then an AI model writes the reading from that chart. Every Oracle answer is checked against the chart before you see it, and so are the written reports.

The check looks for three kinds of statement:

  • Chart facts. Which sign and house a planet is in, in the birth chart and in the divisional charts. Whether it is exalted, debilitated, in its own sign, combust or retrograde. Which house it rules. And the numbers a reading quotes from the chart, such as strength scores and planet degrees.
  • Rules from the classical texts. When a reading states a classical rule, or credits a rule to a named text, the rule is compared with the passages from the texts that the reading was given.
  • Verdicts. Before any writing starts, the chart engine works out its own verdict for an area of life on a five-step scale, from strong to severely strained. The reading's overall verdict for that area is compared with it.

Where the facts come from

The facts are calculated, not written. Supastro computes your full chart first: the birth chart, the divisional charts, your dasha periods, planetary strengths and today’s transits. Planet positions are computed from NASA JPL ephemeris data. Dasha periods are cross-checked against an independent second engine.

The AI model that writes a reading is given these computed facts. Its job is to explain them, not to work them out.

The classical rules come from a library of passages from the texts, such as the Brihat Parashara Hora Shastra and Phaladeepika. For each reading, passages are picked for the chart factors that matter to the question, and the reading is checked against those same passages.

The check, step by step

  1. The reading is written from the computed chart, together with passages from the classical texts chosen for the chart factors that matter to the question.
  2. A second model reads the finished reading and lists the claims in it: chart facts, classical rules and verdicts. It copies each one word for word. It does not decide whether a chart fact is true.
  3. Code checks each chart fact against the computed chart. If the reading says Jupiter is in the 5th house, the chart is checked for where Jupiter actually is.
  4. The second model compares each classical rule with the passages. A rule that a passage contradicts fails. So does a rule credited to a named text when the passages do not state it. A rule the passages do not address does not fail and is left in.
  5. Code compares the reading's verdict for an area with the engine's verdict. Two or more steps apart on the five-step scale fails.

An Oracle answer then gets one more check, done by code alone. The numbers and dates it quotes are compared with the chart again, and the answer must be complete.

When a claim fails

A failing sentence gets one corrective rewrite. For a wrong chart fact, the rewrite states the true value from the chart. For a wrong verdict, it states the engine’s verdict. For a rule, it uses the rule as the passage states it, or drops the credit to the text. Only the failing sentences are changed, so the rest of the reading keeps its wording.

The rewritten sentences are checked again. A sentence that still fails is removed. If removing it would leave a section empty, the computed fact from the chart, where there is one, takes its place.

Dates and timing

Dates are computed before an answer is written. The chart engine works out the dated periods from your dasha periods and the transits, and the writing model is given those windows.

An answer to a timing question must quote one of those windows exactly, with its start date and its end date. If the draft leaves it out, the engine’s window is added to the answer. A predicted date that falls outside every computed window sends the draft back for a rewrite.

Supastro keeps a record of dated predictions. A prediction goes into that record only when its window matches an engine-computed window, date for date, and it is stored with the engine’s dates.

Yoga and dosha questions

The chart engine checks every chart for a fixed list of yogas and doshas. When your question asks about a yoga or a dosha, the answer has to quote a result the engine computed:

  • Formed. The answer names it, with the status the engine gives it.
  • Not formed. If the engine checked for it and did not find it, the answer says it is not formed in this chart, checked by the engine.
  • Not checked. If the engine does not check for the one you named, the writing model is told that it is unchecked and not confirmed by name. The answer has to rest on the yoga and dosha results the engine did compute.

When a reading is refused

An Oracle question is stopped before any writing starts when:

  • a part of your chart that the answer needs could not be loaded, or
  • the question asks about an exact fact that the chart engine cannot verify.

In both cases Supastro tells you so and does not guess. No credits are used.

If an answer is written but does not finish, or fails the completeness check, it is not shown and the credits held for it are restored.

What the ✦ Verified mark means

An Oracle answer that completes this process is shown with the ✦ Verified mark. It means the chart behind the answer was complete, the answer went through the checks on this page, and it was saved only after the final check by code accepted it. No part of an answer appears on your screen before that.

The mark is about the process. It does not rate how favourable the answer is.

What verification does not mean

  • It does not check the future. Verification compares a reading with your chart and with the classical texts. It is not a promise that an event will happen, or that it will happen on a given date.
  • It depends on your birth details. The chart is calculated from the date, time and place you gave. A few minutes of difference in the birth time can move the dates.
  • It is not a promise that a reading has no errors. The check covers the claims it can identify: chart facts, rules credited to the texts, verdicts and dates. If the second model does not respond, the checks done by code still run.
  • It is not medical advice. A sentence that tells you a medical procedure or treatment is required is removed from an Oracle answer. Medical decisions belong with your doctor.

Read more