Why every AI answer about the law should show its sources

Ask a general-purpose AI a question about Nepali law and you will usually get a fluent, confident answer. The trouble is that fluent and correct are different things. A model can quote a section number that does not exist, attribute a rule to the wrong Act, or blend two provisions into one. For a lawyer the cost is not an awkward moment. It is a filing that cites something that is not there.

Why instructions are not enough

The usual fix is to tell the model, in its instructions, to cite its sources and never invent anything. That helps, but it is a request, not a guarantee. A model that is asked not to invent a citation can still produce one, because nothing in the system stops it.

What Bahas does differently

Bahas does not let the model answer from memory. It works in a fixed order, and the last step is a check written in code, not a polite request.

1. 1

   Search the library

   Your question is used to search the legal library for passages that bear on it.

2. 2

   Grade what came back

   Each passage found is scored for relevance. If the results are weak, Bahas rewrites the question and searches again, up to twice.

3. 3

   Write from those passages

   The answer is written only from the passages that passed, with a number beside each claim.

4. 4

   Verify the citations

   Every number in the answer is compared with the passages that were actually retrieved. A number that does not match a retrieved passage is removed before you see the answer.

Because the final check is code, the guarantee does not depend on the model behaving well. A citation that survives to your screen points at a passage that was really retrieved for your question.

What this does not promise

It is worth being plain about the limit. The check proves that a citation points to a real passage. It does not prove the passage supports the conclusion drawn from it, and it cannot tell you whether a provision applies to your client's facts. That judgement is yours.

So the working habit we suggest is simple: read the answer, open the numbered passage, and read the original wording before you rely on it. Bahas makes that fast, because the source is one click from the claim.