भेटिएन भने, भनिँदैन।

Two citations, one gate

LikhatAI writes a token, never a citation. Code walks the act’s own section markers. A marker renders the citation with the act’s words beside it. No marker deletes the token from the sentence.

token

cite · muluki-civil-code-2074 · 422

act

मुलुकी देवानी संहिता, २०७४

marker

४२२. दोहोरो पारी सम्पत्ति हस्तान्तरण गर्न नहुने

result

marker found · text clean · rendered with its footnote

token

cite · muluki-civil-code-2074 · 952

act

मुलुकी देवानी संहिता, २०७४

result

no marker for ९५२ — the code ends at दफा ७२१ · token deleted from the sentence, reason written under “Lawyer review required”

Six stages. One writes.

• Reading your briefmodel, then code parses it
• Finding the lawmodel
• Writing the documentmodel — the one call that writes the draft
• Polishing the registermodel, then code compares every name, number, date and citation byte-for-byte and discards the polish on one mismatch
• Running every checkcode
• Citation reportmodel, advisory — the draft is already written

Every model call runs with no tools and a single turn. Every check but one is ordinary code.

What every run checks

1. Input floor — refuses before any model runs
2. Document-type refusal
3. Citation gate — every cited section must exist in the corpus, or the citation is deleted
4. Residual citation sweep
5. Date gate
6. नजीर gate
7. Act-name gate
8. Personal-data gate
9. Footnote gate
10. Section verification
11. Required-provision check
12. Relief completeness
13. Pleading order
14. Internal consistency
15. Vague-grounding check
16. Caption check
17. Register byte-net
18. Party-name net
19. Storage-path backstop
20. Form conformance — the one check that uses a model, and only to repair

What it reads

15,000

statute files

40,965

grounding cards

360

case files

2,379

templates

68

statutory forms served

53

document types