Jev as a judge of causal graphs: precomputed statistics did most of the work

Six authors from Guangdong University of Technology published a preprint called CausalGrade on arXiv on 28 October 2026. S79 It asks how to compare several candidate causal graphs, that is, diagrams of which variable directly influences which, when nobody knows the correct graph. Jev does not receive the whole task at once. It answers many small yes/no questions, and code combines the results. S79 This way of splitting the work is useful well beyond causality research. The paper is a preprint that has not yet been peer reviewed; this article is based on its full text.

How the method is put together

For every ordered pair of variables, the authors ask Jev whether the first one is a direct cause of the second. They ask through Noul, which returns a number between 0 and 1. S79 TypeSafe's documentation describes a Noul answer as the probability that the answer is yes, where 0 means no and 1 means yes. S05 The authors used version jev-1.13.0 in all experiments. S79

Two kinds of evidence go into the context of each question. The first is a statistical summary of the data that code computes in advance: descriptive statistics and the share of missing values, correlations, conditional independence tests, and indicators of direction derived from regression in both directions. The second is a matrix of edge probabilities from a pretrained causal discovery model that the authors call CDFM. S79 The graph being evaluated is not part of the question. All candidates on the same data therefore receive the same matrix of answers. S79

The graph score is then a plain sum in code. For each pair where the graph has an edge, it adds Jev's answer, and where it has none, it adds one minus that answer. The result corresponds to the expected number of correctly marked pairs. S79 It is the same principle we describe in the article on composite scoring: the model answers partial questions, while weights and the sum stay in code.

What the numbers showed

On 500 synthetic tasks, the goal was to pick the best of five candidates. CausalGrade picked the best graph in 54.4% of tasks, the comparison method κG in 48.4%, and the CDFM probabilities alone in 19.6%. S79 On the real Light dataset it reached 96.75% against 78.0% for κG, and on the Sachs dataset 66.5% against 63.5%. S79 So the lead was not equally large everywhere.

The score also tracked the actual quality of graphs more closely than the CDFM matrix for graphs from all four algorithms studied. For DirectLiNGAM the difference was only small: a Spearman correlation of −0.523 against −0.512. S79

The ablation is the most instructive part. When the authors removed the statistical summary from the input, the selection hit rate fell from 54.4% to 18.8%. Without the CDFM matrix it stayed at 53.8%. S79 According to the authors, most of the gain came from the precomputed summary, and the CDFM matrix was only supplementary evidence. S79

Where the limits are

The authors themselves point out that the candidates came from different thresholds on the same matrix, so they differed mainly in the number of edges. Part of the improvement may therefore come from the score favouring sparser graphs rather than from better discrimination of specific edges and their direction. S79 The ablation does not show which diagnostic helps most, nor whether the gain comes from the information itself or from the way it is organised. With a small data sample, the method may also remove true edges. The evaluation rests on synthetic data and two real systems, with five candidates per task. S79 The preprint reports no token counts, speed or costs. S79

What developers and online shops can take from it

A typical Slovak online shop will not be working with causal graphs. What does transfer is the division of labour the preprint shows. Code computed the numbers, and Jev received a finished, structured summary rather than raw table rows. Without that summary, selection was clearly worse. S79 This matches the documentation's advice on Noul: one condition per question, combine several questions in code, and send supporting data in structured instructions. Several Noul questions in one request run in parallel. S05

How this could look in practice is our own suggestion, not a result of the preprint. If an online shop wants to estimate whether returns of a particular product are linked to a defect, code can first compute the return rate, the number of complaints and the most common reasons. Jev then receives this summary and short yes/no questions, and code again processes the resulting probabilities. Unlike the preprint, which works directly with raw probabilities, the documentation suggests starting with thresholds of 0.7 and 0.3 when deciding individual cases and sending values in between to a person. S05 How Noul differs from Choice and Score questions is explained in the guide Choice, Score and Noul.

A larger context also has a price. Input tokens cost USD 0.042 per million and output is free. S01 Every added summary therefore enlarges the paid part of the request, and with dozens of questions per case it is worth checking pricing and limits in advance.