Ask any assistant for your birth chart without giving it one and it will produce a chart. Confident, specific, wrong — planets in signs they were nowhere near, degrees to two decimal places, a rising sign it had no way to compute. It is not lying. It is doing the only thing it can do when asked for a fact it does not have, which is to produce the most plausible-looking answer.

Giving it a computed file removes the need to invent. It does not remove the habit. So here is how to check.

The instruction that makes checking possible

At the top of your file, before any chart data, sit seven instructions in order of precedence. The first one is the reason this article can exist:

“Never invent chart facts. Every position, sign, house, dignity or date comes from a row in the data block. If no row says it, say it is not in the file.”

That gives you something to hold it to. An assistant that has read the file has been told, in writing, to admit a gap rather than fill one. When it fills one anyway, you can catch it — because the file is the authority, and the file is right there.

Four checks

Ask for the row. After any specific claim, say: which line of the data block says that? A real answer quotes a coded row. An invented one paraphrases the claim back at you in different words, or explains the general principle instead of citing the specific fact. The dodge is the tell.

Ask the same thing twice, twenty minutes apart. Computed facts are stable. If your Moon was in one sign at the start of the session and a neighbouring one later, nothing was ever read — the answer was generated both times, and generation is not reproducible.

Ask for something the file does not contain. This is the most useful check and almost nobody runs it. Ask it for the date of your own death, or what your neighbour’s chart says, or which stock to buy. The correct response is a refusal that names the reason. If it produces an answer, it is not respecting the file's boundaries, and nothing else it told you is trustworthy either.

Watch for confident jargon. A long answer thick with terms and thin on dates is usually the model padding. Your file has actual numbers in it — six strength measurements per planet, house scores, period dates. Ask for the number. Padding cannot produce one.

Why a free chart makes this worse

The common shortcut is to generate a chart on a free site, paste the positions in, and ask away. Two things go wrong, and they compound.

Free calculators disagree with each other. The ayanamsa alone differs between tools by enough to move a planet into the next house, and most of them do not state which one they used — why that happens is worth understanding once. So your starting data may already be wrong, and the assistant has no way to know.

And a positions list is not the material a reading needs. No dignities, no combustion orbs, no divisional charts, no Ashtakavarga, no period dates. Asked about any of it, the model does not stop. It produces something plausible, and you cannot tell the difference from the outside.

What a file cannot fix

Be clear about the limit. A computed file makes answers checkable. It does not make them correct, and it does not make the interpretation right. The model can still quote a genuine row and then draw a conclusion from it that no classical source would support.

Interpretation is where you stay involved. Ask why whenever an answer surprises you, ask what would have to be true for it to be wrong, and treat a reading as a hypothesis with evidence attached rather than a verdict. Medical, legal and financial decisions need appropriate evidence and qualified help, not a chart.

Which assistant you use changes how often you need these checks. A model holding the whole file at once contradicts itself less than one reading it in fragments, which is why the per-app notes say plainly which ones do what.

If you want to see the material an assistant is supposed to be working from, the files are here and a complete one is readable before you pay.