Extracting Governing Law and Jurisdiction Clauses From a Contract

“Governed by the laws of New York” and “the courts of New York County shall have exclusive jurisdiction” are two sentences that share a place name and answer entirely different questions. They are routinely printed in the same numbered section, and an extraction that returns one section returns whichever came first.

Three clauses, one section

Near the end of most commercial agreements sits a section headed something like “Governing Law; Jurisdiction”, or folded into a miscellaneous section with waiver, severability and notices. It usually contains up to three distinct propositions:

• Choice of law. Which body of substantive law supplies the rules for interpreting the agreement. The output field is a jurisdiction: a state, a country, occasionally a sub-national unit named unusually.
• Forum or venue. Where a dispute is heard. The output field is a court or a place, plus whether the submission is exclusive or non-exclusive — a distinction carried by a single word that a summarising extraction drops.
• Dispute resolution. Whether disputes go to a court at all, or to arbitration, and if so under which institution’s rules, with what seat, before how many arbitrators, in what language. Several sub-fields, none of which is a jurisdiction.

These can and do disagree in ways that are perfectly deliberate: an agreement governed by the law of one place, arbitrated under the rules of an institution headquartered elsewhere, seated in a third, with court proceedings for injunctive relief permitted anywhere. A schema with one governingLaw string flattens four facts into one, and the flattening is not recoverable from the stored value.

The structural consequence is that the unit of extraction is the proposition, not the section. Sectioning a contract by heading and asking a model to label each section is the approach that produces this error, because the heading is honest — the section really is about governing law and jurisdiction — and the model summarises it into whichever the first sentence establishes.

The wording that tells them apart

The vocabulary is conventional enough to be a reliable classifier, and the discriminating tokens are few. Choice of law is signalled by governed by, construed in accordance with, the laws of, and the object is a body of law rather than a building. Forum is signalled by jurisdiction, venue, submit to, the courts of, forum non conveniens, and the object is a court or a geographic seat of one. Arbitration is signalled by arbitration, arbitrator, final and binding, and by naming an institution and a rule set.

Two false friends account for most misclassification. The phrase the laws of appears inside forum clauses that are not choice of law — “courts located in the State of Delaware” carries a jurisdiction name with no choice-of-law content at all. And arbitration clauses contain a governing-law-shaped sentence about the arbitration agreement itself, which is a different thing from the law governing the contract; the two are distinct propositions and in international agreements they are sometimes deliberately different.

For arbitration the sub-fields worth having are the institution, the rule set with its version where stated, the seat, the number of arbitrators and the language. Institutions such as the American Arbitration Association and the International Chamber of Commerce revise their rules periodically and a clause referring to rules “in effect at the time” means a different document than one referring to a dated edition. Extract the reference as printed rather than normalising it to an institution name.

Without regard to conflict of laws

Most choice-of-law clauses carry a trailing qualifier: without regard to its conflict of laws principles, or excluding any rules that would require the application of the law of another jurisdiction. It reads as boilerplate, is one clause in perhaps four sentences, and its presence or absence is a separate boolean field.

For extraction purposes the reason it matters is that its absence is the interesting state and it is the state a summarising extraction cannot report. A model asked “what law governs” answers with a place name whether or not the qualifier is there. A model asked for a structured object with a conflictOfLawsExcluded boolean has to look, and a null tells you the sentence did not settle it. That is the general shape of every useful clause extraction: sub-fields you require the model to fill force a reading that a free-text summary does not.

The same applies to the exclusion of the UN Convention on Contracts for the International Sale of Goods, which in cross-border sale agreements is commonly disclaimed in the same sentence. It is a boolean with three states — expressly excluded, expressly applied, not mentioned — and the third is genuinely different from the first. Do not model it as a nullable boolean where null and false mean the same thing in the consuming code.

The carve-out that changes the answer

Almost every forum and arbitration clause has an exception, and the exception is where the drafting effort went. The usual shapes:

• Either party may seek injunctive or equitable relief in any court of competent jurisdiction, notwithstanding the arbitration agreement. An arbitration clause with this carve-out does not send everything to arbitration, and an extraction reporting a flat “arbitration” is wrong for the fastest-moving category of dispute.
• Claims relating to intellectual property or confidentiality are excluded from arbitration, which pushes them back to the forum clause — so the two clauses have to be read together to answer anything.
• A jurisdiction submission qualified as non-exclusive, which means the clause names a court without ruling out others. The word non-exclusive may be the only difference between two otherwise identical clauses.
• Small claims or collection actions preserved for the courts, often in a subordinate sentence at the end.

Model the carve-outs as a list attached to the clause, each with its trigger and its destination, rather than as prose in a notes field. The exception-restores-the-general-rule pattern is exactly the one that insurance exclusions demand, and the structural answer is the same: the exception is a child of the rule it modifies, never a sibling.

A place name is not an extraction

A word about why the outputs need to be conservative. Certain jurisdictions appear far more often than any distribution of counterparty addresses would predict, because some statutes expressly permit parties to choose that jurisdiction’s law for sufficiently large transactions without a local connection — New York’s General Obligations Law section 5-1401, published by the New York State Senate, is the best-known example. So a governing law that has nothing to do with either party’s address is ordinary and a validator that flags a mismatch as suspicious will flag most of your corpus.

What is worth flagging is structural: a contract with a forum clause and no choice-of-law clause, a choice of law with no forum and no arbitration, an arbitration clause naming no rules, or two sections in one document that both purport to choose law. Each of those is a fact about the document that you can state without interpreting anything. Report them as findings with the section numbers attached, alongside the termination clause and the other operative provisions, and leave the question of what any of it means to the people whose job that is. The extraction’s job is to be complete and to be honest about where it is not.