Madeleinex/icj-precedent-cite-bench

ICJ Precedent-Citation Benchmark

A benchmark for one task over the International Court of Justice: given the record of a case as it stood before the Court issued its decision, predict which earlier ICJ cases (and which provisions of the Court's own instruments) the decision will cite.

The five cases were all decided after the knowledge cutoffs of current frontier models, so a model cannot have read the decisions during pretraining, and each case is removed from the graph so it cannot be retrieved either. It is a companion to the ICJ-Graph corpus; the graph supplies the pool of earlier cases and instrument provisions a system searches.

Scope and caveats

The test set is five cases because that is what the domain allows. The ICJ has decided only about 200 cases since 1947, and only a handful postdate current models' knowledge cutoffs, which is what a leak-free held-out set requires. Two properties bound what the scores can show:

• Five cases means high variance. Per-case F1 ranges from 0.04 to 0.49 across the reference runs, so read the numbers as directional rather than as a precise ranking between close systems.
• The gold is machine-extracted and not yet human-verified. Each case's cited-precedent set comes from an LLM reference pass over the majority opinion, resolved to case ids by fuzzy name match at a 0.90 threshold. Exact-id scoring also counts a legally valid substitute authority as wrong, so F1 understates a good answer.

Task

Input: one question per case, built only from documents available before the decision (the question put to the Court, the facts, and each party's arguments), with the Court's own findings and citations removed and leak-checked. Output: the earlier ICJ cases the decision cites, optionally the instrument provisions it cites, and the jurisdiction disposition.

Data fields (data/test.jsonl, split test, 5 rows)

field	type	description
case_id	str	ICJ General List number, zero-padded (e.g. 179)
title	str	case name
type	str	judgment / advisory_opinion / order
decided	str	decision date (all after 2025-01)
question	str	the leak-free input record
gold_cited_cases	list[str]	gold set: earlier ICJ cases the decision cites (case ids)
gold_cited_case_names	dict	id to case name, for reference
gold_cited_case_freq	dict	how often the opinion cites each case (weights wRecall)
gold_cited_provisions	list[str]	cited provisions of the four Court instruments (e.g. statute:article_36)
jurisdiction	str	gold disposition: yes / no (prima facie for the order)
decline	str	for advisory opinions, whether the Court declines; n/a for contentious cases
holding	str	one-line gold holding, for reference (not scored here)
gold_source	str	provenance note for the gold

Cited cases that are themselves held-out (one advisory opinion cites another) are excluded from the gold, because their target is absent from the retrievable graph.

Metrics

• Case citation and provision citation precision / recall / F1, reported both micro (pool true/false positives and negatives over all answers) and macro (per-case, then mean over the five cases).
• Weighted recall (cases): recall with each gold case weighted by how often the opinion cited it, so missing a load-bearing precedent costs more than missing a one-off. Averaged over answers.

Jurisdiction sits outside the citation score. Each case carries its gold jurisdiction disposition, and score.py reports accuracy on it as an auxiliary signal.

Evaluate

Write one JSON object per answer (you may submit several answers per case; they are pooled):

{"case_id": "179", "cited_cases": ["075", "120", "130"], "cited_provisions": ["statute:article_36"], "jurisdiction": "yes"}

python score.py --gold data/test.jsonl --pred your_predictions.jsonl

score.py is dependency-free and computes every metric in the tables below from a predictions file.

Reference results

Two non-model baselines (cite the graph's k most-cited cases, or its k most-cited provisions) and one generation model in two sizes under three cumulative settings: Parametric (closed-book, no tools), Cypher only (read-only Cypher over the graph), and + semantic (adds meaning-based passage search, Qwen3 embeddings, top-k 20). All scores are micro; each model cell pools three samples per case, and each most-cited row is one fixed prediction.

Case citation

Setting	Model	Prec	Rec	wRec	F1
Most-cited (k=6)	-	0.23	0.09	0.03	0.13
Most-cited (k=18)	-	0.20	0.23	0.15	0.21
Parametric	gpt-5-mini	0.47	0.13	0.19	0.20
Cypher only	gpt-5-mini	0.48	0.19	0.36	0.28
+ semantic	gpt-5-mini	0.62	0.24	0.45	0.34
Parametric	gpt-5	0.57	0.32	0.42	0.41
Cypher only	gpt-5	0.52	0.27	0.40	0.36
+ semantic	gpt-5	0.60	0.29	0.43	0.39

Provision citation

Setting	Model	Prec	Rec	wRec	F1
Most-cited (k=6)	-	0.17	0.12	0.13	0.14
Most-cited (k=18)	-	0.12	0.27	0.25	0.17
Parametric	gpt-5-mini	0.83	0.20	0.28	0.32
Cypher only	gpt-5-mini	0.87	0.22	0.34	0.35
+ semantic	gpt-5-mini	0.93	0.20	0.30	0.33
Parametric	gpt-5	0.77	0.33	0.45	0.46
Cypher only	gpt-5	0.75	0.27	0.36	0.40
+ semantic	gpt-5	0.82	0.30	0.48	0.44

Read against the five-case caveat: precision stays higher than recall in every row, so recall caps F1. The graph raises the smaller model's scores more than the larger model's. Most cited cases a model misses were passages it had already retrieved, so selecting among retrieved candidates looks like a larger source of error than retrieval.

License

Mixed per source, matching the ICJ-Graph corpus (license: other, see LICENSE). What this project produced (the questions, the gold annotations, the scorer) is licensed CC-BY-4.0; the ICJ material quoted in the questions is a work of the Court and is passed through under the corpus card's per-source terms.

Citation

@misc{icjprecedentcitebench,
  title  = {ICJ-Graph: An Open Citation Graph and a Layered Document Corpus of the International Court of Justice},
  note   = {Anonymized for review},
  year   = {2026}
}