Actio pauliana: rules as code and LLMs

Article 42 of the Dutch Bankruptcy Act (Faillissementswet) lets the trustee challenge a voluntary transaction of the debtor if it prejudiced creditors and the debtor, and for a transaction for consideration also the counterparty, knew or should have known of the prejudice; Articles 43 and 45 introduce rebuttable presumptions of knowledge for transactions within a year before bankruptcy. The structure of the rule formalizes well, but three concepts - voluntariness, prejudice to creditors and knowledge - are developed only in case law. The authors build a system in which the rules are written as code and the evaluative concepts are decided by LLMs relying on precedents. The retrieval corpus: out of more than 270,000 civil decisions from rechtspraak.nl, 846 were selected that mention Articles 42, 43 or 45.

The core is a model of Articles 42, 43 and 45 in the Catala language, written by hand together with a practicing trustee and a legal scholar, with explanations for each line of the statute. The evaluative concepts in it are logical inputs filled by neural components: one subroutine extracts facts and dates from the case description, three others decide voluntariness, prejudice and knowledge, relying on retrieved decisions. For retrieval, the decisions are split into sections (course of proceedings, facts, dispute, reasoning, operative part) by a deterministic segmenter whose vocabulary was compiled with an LLM, and labeled by which of the three concepts the court addressed; retrieval runs only over decisions on the relevant concept and is supplemented by decisions of other instances in the same case. The system was tested with three LLMs - Gemini 3.5 Flash, GPT-5.5 and Mistral Medium 3.5 - on ten cases outside the corpus, mostly from the JOR case law journal, with 0, 5 and 10 retrieved decisions.

The segmenter on 100 manually labeled decisions gives precision 0.98 and recall 0.97; concept labeling by the Mistral model gives recall 0.94 at precision 0.80. On the ten cases, across all criteria and the final conclusion, precision is 0.63-0.68, recall 0.82-0.90, and the correct outcome is reached in 6-7 cases out of 10. Voluntariness and prejudice are decided well (F1 0.75-0.89), knowledge poorly (F1 0.10-0.25). The number of retrieved decisions barely affects the figures but changes the reasoning: in one case GPT-5.5 cited the same decision the court itself had cited. In another case Gemini and GPT-5.5 reached opposite conclusions on prejudice from the same precedent, and it was GPT-5.5’s conclusion that matched the court.

The authors name the limitations themselves. The sample of ten cases is small and skewed toward difficult cases selected by the case law journal, while uncontested cases are resolved by the trustee’s out-of-court declaration and do not reach open data. The reference labels were made by one author. In one of the cases the authors explain the error on knowledge by the court having treated it as established because it was insufficiently contested, which is not in the input facts; in addition, the model’s textual explanation may diverge from the conclusion of the symbolic core. For court analytics the paper shows a division of labor: the formalizable part of the rule is computed as code and can be verified, while the LLM is responsible only for the evaluative concepts and cites specific decisions, but the reliability of that part is still limited, and next the authors propose testing the system together with trustees.