For the last fortnight I have been building a corpus of Indian primary law on my own machine. Bare Acts, State rules, departmental G.O.s, judgments. Nothing about it is clever. The point is that when I cite a section, I have read that section from a source I can name.
Most of the raw material came from one place: open-india-law, published by vaquill-AI. Judgments from the Supreme Court and all twenty-five High Courts, some twenty-two thousand enactments, tribunal decisions, all scraped from official government sites, scrapers under Apache-2.0 and data under CC BY 4.0. Someone has pointed a crawler at every High Court website in the country and lived with the consequences. Anyone who has tried to get a page out of one on a Tuesday afternoon will know what that means. The scrapers are published, so the chain from the court’s server to the entry in front of you can be checked. That matters more than the volume does.
I took three of the twenty-six judgment files — Supreme Court, Telangana, Andhra Pradesh — a little over thirteen lakh judgments. Every other High Court is persuasive in a Telangana matter, and I would rather hold three files I have been through than twenty-six I have not.
What came with the judgments
The Supreme Court entries carry citations. The High Court entries carry none. Not a neutral citation, not SCC, not AIR — twelve and a half lakh judgments where I have the full text, the parties, the case number and the date, and no reference to put in a plaint. Useful for research. Useless for a citation.
The dates are the quieter problem. The Telangana file runs to a maximum decision date of 27.12.2079, the Andhra Pradesh file to 30.06.2030. Those are not curiosities; they mean any date filter is unreliable until the column has been checked. And the year field in the Andhra Pradesh file does not hold the year of decision at all — it holds the date the record was added. The field name gives no hint of that.
None of this is a fault of the project. Court websites publish what they publish, and a scraper cannot conjure a citation that was never on the page. But speed and apparent completeness are not verification.
Why I set the statutes aside
The legislation files I downloaded and then did not use. That is the part worth writing about, because it is not visible in the answer a tool returns.
I went looking for section 105 of the Transfer of Property Act — the definition of a lease, which I have had open for weeks for a different article. Two entries came back, both labelled section 105. The first is section 105. The second, with the same label, is the text of section 106. Ask for 106 instead and you get a fragment of the arrangement-of-sections list: “106. Duration of certain leases… 107. Leases how made. 108. Rights and liabilities of lessor and lessee.” Section 106 is in the file. It cannot be reached by its own number.
The mechanism shows in the identifiers. A long section is split into parts, and when a part runs past the end of its section into the next one it keeps the earlier number. By a crude test — comparing the number on the entry against the first section number appearing in its text — about one entry in thirty across the central legislation file disagrees with its own label. The test cuts both ways and I would not defend that figure to a decimal. The section 105 example is the part that matters: nothing on the entry tells you which kind you have, and a search for “lease defined” returns both with equal confidence.
So I rebuilt the statute side from India Code and the Telangana Law Department portal — about seventeen hundred instruments against the twenty-two thousand in the dataset. A fraction of it, each one from the department that administers the Act.
The primary source is not clean either
India Code is the Government of India’s own statute portal, and it is where bare Act text ought to come from. It is also incomplete in ways nothing on the page announces. Of the Acts I hold with numbered sections, 142 show gaps in the numbering — 552 section numbers missing between the first and the last. Most of those I have not examined, and many will be genuine omissions in the enactment. Of the ones I did examine, 49 sections across 23 Acts were in the file all along, filed under a neighbouring section’s number because a heading failed to parse.
The Hyderabad Metropolitan Development Authority Act, 2008 is the clearest case. India Code’s copy omits ss.5, 6 and 8, and its section 9 opens with a repeal provision that belongs elsewhere. I established that by reading it against a printed copy. There was no other way to establish it.
Then a gap that no amount of parsing fixes. Schedule I-A of the Indian Stamp Act, 1899 — the schedule that fixes stamp duty in Telangana and erstwhile Andhra Pradesh — is not on India Code. The portal carries State amendments for Assam, Himachal Pradesh, Orissa and Uttar Pradesh, and stops there. That is a limit on its remit rather than a defect in it; Schedule I-A is a State amendment to a central Act and the portal does not undertake to carry every one. The practical consequence is the same. For a Telangana conveyancing question, the national statute portal is not where the answer is.
The licence article, and an open question
The lease side of Schedule I-A checks out. Article 31 charges a lease of under one year at 0.40% of the whole amount payable, rising through 0.50% and 1% for one to five years, 6% for ten to twenty, 15% for twenty to thirty, and 3% of market value beyond thirty. Those figures appear in the department’s own Ready Reckoner, in the copy of Schedule I-A I hold, and in third-party calculators — three sources that did not copy each other.
What gets mis-stated is the authority. The rate sites cite G.O.Ms.No.59 dated 20.07.2021 for lease duty. That G.O. covers Articles 6(B), 16, 20(c), 42, 47-A, 49 and 49-A. It says nothing about a lease.
Now the licence. Schedule I-A has a charging article of its own for one — Article 33, “LICENCE of immovable or movable property… granted by owner or authority for rent or fee or by whatever name it is called”. Its first limb reads as a graduated scale, ₹2 per ₹100 for the first ₹1,000 and ₹10 per ₹500 above that, charged “for the whole amount payable, or deliverable under such licence”. Both legs come to the same thing. It is a flat 2%, on the same base Article 31(a)(i) uses.
| If the instrument is | Article | Duty | On ₹5,50,000 |
|---|---|---|---|
| a lease | 31(a)(i) | 0.40% of the whole amount payable | ₹2,200 |
| a licence | 33(a)(i) | the graduated scale — 2% of the whole amount payable | ₹11,000 |
Before that gap tempts anyone: this is not a choice between two articles for one document. Duty follows what the instrument does, not what it is called, and Article 31 itself reaches “any agreement to let or sublet”. A leave-and-licence that hands over exclusive possession for a term at a rent is a lease, and it goes to Article 31 whatever the heading says. Article 33 reaches the genuine licence — non-exclusive, revocable, no interest passing. The ₹8,800 turns on which of those has actually been drafted.
Which leaves the question I cannot answer. I have not found Article 33 applied to a leave-and-licence in Telangana. The Ready Reckoner has no licence card. The rate sites put such an instrument under Article 31 at 0.4% or 0.5%, where they name an article at all. Article 33 sits in the Schedule, in terms that fit a licence of immovable property, and the apparatus around it proceeds as though it were not there.
There is an argument against me on the text, and it deserves stating. Article 33 says “granted by owner or authority”, which can be read as aimed at licences from a public authority — a market stall, a municipal space — rather than at a private arrangement. I read “or authority” as covering a grantor who is not the owner. The narrower reading is available, and I have not seen it tested.
One decided case comes close, and I want to be careful about what belongs to the Court and what belongs to me. In New Bus Stand Shop Owners Association v. Corporation of Kozhikode, (2009) 10 SCC 455, the Supreme Court held that the instrument before it was a licence rather than a lease, and fell under the residuary entry of the Kerala Stamp Act rather than the lease entry. That is the holding. The inference I draw — that the outcome would run differently in a State whose schedule carries a licence article of its own — is mine. The Court was not asked about Andhra Pradesh and said nothing about it.
I should be plain about my source. My copy of Schedule I-A is a private compilation, not a gazette. A 2010 edition of the same compilation prints Article 33’s scale in identical words, so the text has been stable; the notification history is harder to pin down, and I could not trace the 2013 G.O. that my copy records as having reduced two of its clauses. The lease figures I could corroborate three ways. The licence figures I could not corroborate at all.
If the sum is large enough to matter, the route is not a blog post and not a guess at the counter. Section 31 allows the instrument to be brought for adjudication — to the Collector, and in Telangana to the District Registrar exercising that power — and section 32 gives a certificate that settles the duty. That is worth the fee, because an instrument not duly stamped is inadmissible in evidence under section 35 until duty and penalty are paid. The exposure is not the ₹8,800. It is the document you cannot put before the court on the day you need it.
What this does not replace
A fair question from anyone who has been at this longer than I have: why do any of it, when the commentaries are on the shelf, the court library is open, and a search takes ten seconds?
For most questions the shelf wins, and I would not argue otherwise. Mulla on the Transfer of Property Act carries what no file of mine does — an editor who has read the cases, weighed the lines that pull against each other, and told you where the weight sits. A parquet file has no view. It cannot tell you that a section has been read down, that the High Courts have divided on it, or that the point turns on a paragraph the headnote leaves out. For the questions that decide a matter the commentary is the better instrument, and the corpus does not come near it.
What I built is narrower. It answers one question — what does this section say, in these words, and where did the words come from — at eleven at night, in the middle of a draft, without leaving the document. That is a small thing to want. It is also the thing I was getting wrong most often, because the temptation is to work from memory when the book is in the other room. The second reason is less about research than about what is already happening: I use these tools, and I expect more of us will. A model asked about s.105 will answer either way. The only choice is whether it answers from a text I have checked.
And the Schedule I-A episode makes the traditional case better than it makes mine. What answered that question in the end was a compilation sitting in my own files since 2019, read against the department’s own counter. The corpus told me a question existed. It could not settle it. The settling came from paper and from the people at the counter, which is where it has always come from.
What I would say to anyone doing this
Open legal data in India has arrived, and vaquill-AI has done a great deal to make that true. Use it — at the right distance. It is a research surface, not a citation source. Between a government website and the entry in front of you sit a scraper, a text extractor and whatever splits long sections into parts, and any of the three can put the right words under the wrong number without leaving a mark.
If you are pointing a language model at Indian legal material — and I am, daily — the question is not whether the corpus is good. It is whether you can open the entry and read the source it came from.
Figures checked against my own corpus on 01.09.2026, from the open-india-law v2026.08 snapshot downloaded 25.08.2026; the project may have corrected any of it since. Nothing here is advice on what duty to pay on a particular instrument.