The judgment is public domain. The headnote is not.

Take a Supreme Court judgment from a commercial law report, train an embedding model on the page as printed, and you have quietly baked someone else's copyright into your weights. The judgment itself is public domain. The headnote printed above it is not. Most "legal AI" corpora never notice the difference, and it is the difference that decides whether you can ship the resulting model at all.

This is the provenance story behind embed-legal-en, our open retrieval adapter for Indian Supreme Court judgments. The weights ship Apache-2.0 with no share-alike obligation. That licence was not a decision we made at the end; it is a property we had to earn in the corpus, one cut at a time.

One distinction decides everything

Under the Copyright Act 1957, §52(1)(q) places the text of a court judgment outside copyright. It is statutory public domain, which is the cleanest provenance a model can have: there is no licensor who might later be found not to have held the right, in a way even a Creative Commons grant cannot fully rule out.

What the section frees is the judgment. What it does not free is the reporter-written headnote, the syllabus a law report prints above the decision to summarise it. That headnote is original editorial work, and it is separately copyrightable, a line the Supreme Court drew in Eastern Book Company v. D.B. Modak (2008): the raw judgment text stays public domain, while the reporter's original editorial contribution can attract its own copyright. Train on it and your weights are no longer purely public-domain: they now carry a fragment of a publisher's copyrighted text, and the clean Apache-2.0 story is gone. So "download some judgments" is not a corpus. The corpus is the judgment text with the editorial layer surgically removed.

Source discipline comes before cleaning

The first and biggest lever is where the text comes from. We pull judgments only from official court portals, never from a commercial reporter such as SCC or Manupatra. The reporter's product is precisely the editorial layer: the headnotes, the catchwords, the arrangement. The court's own copy on an official portal never had that layer to begin with, so the cleanest input is the one you never have to scrub.

It is tempting to treat a government host as a green light, and it is the wrong instinct. A .gov.in domain does not launder third-party copyright: a company's annual report filed with a ministry is still the company's copyright, merely hosted by the state. Provenance is a sourcing decision before it is a cleaning decision, and the sourcing rule here is narrow on purpose: the court's own order text, from the court's own portal, resting on §52(1)(q). Nothing wider.

How the headnote is actually cut

Even from an official portal, a judgment PDF can carry a reporter-style headnote and page furniture, so the corpus goes through a four-step cleaning pass. The step that matters most is the third.

1. Extract with a layout-faithful parser, not a naive one. A justified two-column law-report page fools a simple extractor into splitting words (en gaged becomes two tokens), and garbled tokens poison every training pair downstream.
2. Drop page furniture: the running margin letters and bare page numbers that carry no legal text.
3. Excise the headnote. This is the cut. We remove the span between the coram line, the [... JJ.] marker naming the bench, and the start of the reported body, anchored either on the CIVIL/CRIMINAL APPELLATE JURISDICTION marker or on the "The Judgment of the Court was delivered by" line. That span is exactly the reporter's syllabus. What survives is the cause-title, the coram, and the full judgment body.
4. Audit for anything still carrying markup, and repeat until the corpus auditor returns zero findings.

The honest part is the fallback. When no reliable anchor is found, the document is never cut blind. It is pulled out of the automated path and flagged for manual review. A confident cut at the wrong offset would slice into the judgment itself and corrupt the very text we are trying to preserve, and a bad cut is worse than no cut. A flagged document only enters the training corpus once a person has confirmed by hand that no headnote remains, and anything still carrying editorial matter is held out entirely. We would rather hand-clear a handful of documents than trust a regex to guess where a syllabus ends.

Why the whole exercise was the licence

Because nothing copyrighted ever entered the corpus, the weights can be Apache-2.0 with no share-alike floor. That last phrase is the payoff. Compare the common alternative: a model trained on Wikipedia text honestly inherits CC-BY-SA, whose share-alike term reaches downstream into anything built on the model, along with an attribution obligation. That is the correct, clean licence for a Wikipedia-derived corpus, because the licence simply follows the source, and our own flagship multilingual model carries exactly that, for exactly that reason. What is different about embed-legal-en is the corpus, so the licence follows differently: because every input is statutory public domain, the weights ship Apache-2.0 with no share-alike floor. Use them commercially, redistribute them, with attribution and nothing more. The framework code is Apache-2.0 as well, kept separate from the weights.

The point of the cut was never tidiness. It was that a model you cannot ship is not a model, and provenance is the part of the pipeline that decides which one you have. The retrieval numbers are on the model card (judgment-to-judgment Recall@1 rises from 0.309 on the base to 0.545 with the adapter), but the number that took the most care to earn is the one on the licence line.