Jev-as-a-Judge for Brazilian Portuguese Information Retrieval: Relevance Assessment and Reranking

Abstract: This study investigates Jev, a System One model designed for fast, structured decision-making, in information retrieval. We use Jev-as-a-judge to generate relevance judgments for four Brazilian Portuguese IR collections covering different domains and compare them with the official relevance judgments of each collection. We also evaluate Jev as a reranker. For relevance assessment, exact agreement ranged from 52.9% to 64.9%, while quadratic weighted Cohen's kappa ranged from 0.52 to 0.82. Most disagreements involved adjacent relevance levels, with at least 94.5% of annotations matching exactly or differing by only one relevance level. Despite differences at the individual query--document level, Jev-generated qrels produced similar retrieval system rankings, with Kendall's tau ranging from 0.90 to 0.93 for NormasTCU and Ulysses-RFCorpus across the two evaluation metrics. For reranking, Jev reordered the top 1,000 BM25 candidates for each of the 50 Quati test queries, increasing nDCG@10 from 0.3329 to 0.6331 (+90.2%) and MRR from 0.6852 to 0.9033 (+31.8%). These findings support Jev as a relevance assessor and show improved ranking effectiveness over BM25 in the evaluated reranking setting.

• The paper evaluates Jev, a non-generative model, in relevance assessment and reranking tasks for Portuguese IR, using Brazilian Portuguese datasets.
• Exact agreement: Jev qrels matched 64.9% to 52.9% of individual relevance decisions but preserved systemic-level similarity with high Kendall’s τ (0.90–0.93).
• BM25 + Jev reranking substantially improved nDCG@10 (90.2% improvement) and MRR (relative 0.69%) in Quati, proving Jev operational effectiveness.

Study objectives and positioning

The paper evaluates Jev, a System One model designed for structured, non-generative decision-making, in two Brazilian Portuguese information-retrieval settings: automatic relevance assessment and candidate reranking. Its central premise is that both tasks can be expressed as bounded decisions rather than open-ended text generation. For relevance assessment, Jev selects a predefined relevance level for a query–document pair. For reranking, it estimates the probability that a document directly answers a query and uses that probability as a pointwise ranking score.

This framing addresses two practical problems in IR. First, manually constructing qrels is expensive, particularly for specialized collections with long documents and domain-specific relevance criteria. Second, generative LLM rerankers can require repeated pairwise, setwise, or listwise inference, creating substantial computational and operational overhead. The paper therefore asks whether a structured decision model can produce qrels that preserve retrieval conclusions and whether it can improve a conventional BM25 ranking.

The evaluation is conducted exclusively on Brazilian Portuguese resources, which is methodologically important because much of the LLM-based IR literature is centered on English. The four collections span web passages, jurisprudence, normative acts, and legislative proposals. They also differ in document length, relevance scale, and source of reference judgments. Quati and JurisTCU use four-level labels, with LLM-generated references and expert review in the latter; NormasTCU and Ulysses-RFCorpus use three-level labels derived from domain-expert annotation. This diversity permits an assessment of Jev under heterogeneous annotation conditions rather than on a single uniform benchmark.

The paper’s motivation is consistent with prior work showing that LLM-generated judgments can produce high system-level agreement despite imperfect query–document-level agreement (Upadhyay et al., 2024). It also distinguishes its reranking formulation from RankGPT-style approaches, which commonly use pairwise or listwise comparisons (Qin et al., 2023). Jev instead scores each candidate independently, groups multiple typed questions in one request, and sorts candidates using the returned relevance probabilities.

Experimental methodology

For qrel generation, Jev version jev-1.13 was presented with each query–document pair and instructed to select a relevance label using the original collection terminology. The resulting judgments were compared against the reference qrels using exact accuracy, confusion matrices, and quadratic weighted Cohen’s $\kappa$. The weighted statistic is appropriate for ordinal labels because a disagreement between adjacent levels is less consequential than a disagreement between the extremes of the scale.

The analysis covered 1,933 Quati pairs, 2,250 JurisTCU pairs, 812 NormasTCU pairs, and 8,205 Ulysses-RFCorpus pairs. To measure evaluation-level consequences rather than only annotation-level similarity, the authors recomputed retrieval effectiveness for 15 NormasTCU systems and 16 Ulysses-RFCorpus systems. They compared system rankings under reference and Jev-generated qrels using nDCG@10 and MRR, with Kendall’s $\tau$ as the agreement measure.

The reranking experiment used Jev version jev-1.13.0 on the 50-query Quati test set. BM25 retrieved the top 1,000 passages from the one-million-passage corpus. Jev independently scored every candidate, with ten document-specific questions grouped into each API request, and the candidates were sorted by the estimated probability of directly answering the query. The experiment used the reference qrels, not Jev-generated qrels, thereby separating the evaluation of Jev as an assessor from its evaluation as a reranker. nDCG@10 used linear gains, while MRR treated every document with a positive relevance label as relevant.

This separation is a strength of the design. It avoids circularly evaluating a Jev reranker using judgments generated by the same model and makes the two claims empirically distinguishable: qrel substitution should preserve system comparisons, whereas reranking should improve the ordering of a fixed candidate set.

Agreement at the query–document level

Exact agreement between Jev and the reference qrels ranged from 52.9% to 64.9%. Ulysses-RFCorpus produced the highest exact agreement at 64.9%, followed by JurisTCU at 64.3%, NormasTCU at 54.8%, and Quati at 52.9%. These values do not indicate near-identical annotation, and the low Quati accuracy is particularly important: almost half of the individual judgments differed from the reference labels.

The weighted agreement statistics provide a more favorable interpretation because most discrepancies were ordinally local. Quadratic weighted $\kappa$ ranged from 0.52 for NormasTCU to 0.82 for JurisTCU. At least 94.5% of all judgments were either exact matches or differed by no more than one relevance level. Differences of two levels occurred in only 2.5%–5.5% of cases, and differences of three levels were absent in Quati and occurred in only 0.1% of JurisTCU cases.

Collection	Exact agreement	One-level difference	Two-level difference	Weighted $\kappa$
Quati	52.9%	44.5%	2.5%	0.68
JurisTCU	64.3%	31.1%	4.6%	0.82
NormasTCU	54.8%	39.7%	5.5%	0.52
Ulysses-RFCorpus	64.9%	29.7%	5.4%	0.65

The disagreement patterns are not uniform across collections. In Quati, for example, Jev frequently assigned label 1 to documents assigned label 0 by the reference qrels, while highly relevant documents were often assigned lower labels. In JurisTCU, the model showed stronger agreement at both relevance extremes, whereas NormasTCU exhibited the weakest weighted agreement. These differences may reflect domain-specific ambiguity, variation in reference-label construction, document length, or the semantics of the relevance instructions.

The paper appropriately compares these values with reported human annotation agreement, where $\kappa$ can also vary substantially. However, this comparison is only contextual. Jev is not being evaluated against an independent consensus ground truth: the reference qrels themselves differ in provenance, with some generated by LLMs and others produced by experts under different protocols. Consequently, overlap with human agreement ranges does not establish that Jev is equivalent to human assessors. It establishes only that its disagreement structure is compatible with the variability observed in relevance annotation.

Preservation of retrieval-system rankings

The paper’s stronger result concerns system-level evaluation. Despite modest exact agreement at the individual pair level, Jev-generated qrels produced highly similar rankings of retrieval systems on NormasTCU and Ulysses-RFCorpus. Kendall’s $\tau$ ranged from 0.90 to 0.93 across both metrics and both collections.

For NormasTCU, $\tau$ was 0.92 for nDCG@10 and 0.90 for MRR. For Ulysses-RFCorpus, it was 0.93 for nDCG@10 and 0.92 for MRR. Under the paper’s adopted criterion, $\tau \geq 0.9$ indicates functional equivalence between evaluation rankings. Thus, the principal contradiction in the results is also the most consequential one: Jev agrees with the reference qrels on only approximately 53%–65% of individual labels, yet the resulting system rankings remain strongly aligned.

Figure 1: NormasTCU system rankings under reference and Jev-generated qrels for nDCG@10 and MRR.

The NormasTCU heatmaps show that the largest ranking changes are limited rather than global. For nDCG@10, one system moves from eighth to eleventh position, while three semantic systems each move upward by one position. MRR exhibits more local reordering, including changes among highly ranked systems, but the overall ordering remains stable. This demonstrates that query–document errors do not necessarily propagate into system-level conclusions when they are distributed across systems and relevance levels.

Figure 2: Ulysses-RFCorpus system rankings under reference and Jev-generated qrels for nDCG@10 and MRR.

Ulysses-RFCorpus shows similar stability, with fewer changes under nDCG@10 and somewhat more movement under MRR. The result is practically relevant for benchmark construction: if the purpose of qrels is to discriminate among retrieval systems rather than reproduce every human label, Jev may be adequate for preserving comparative conclusions in these two collections.

This finding should not be generalized to all evaluation metrics or all system populations. The analysis uses only nDCG@10 and MRR, and it evaluates a fixed set of systems. A different collection, a smaller number of topics, a more tightly clustered set of systems, or a metric sensitive to recall at larger cutoffs could produce different rank perturbations. Nevertheless, within the stated experimental scope, the high Kendall correlations support the paper’s claim that Jev-generated qrels can function as evaluation substitutes at the system-ranking level.

Pointwise reranking effectiveness

The reranking experiment reports a substantial improvement over the authors’ BM25 run on Quati. Jev reranked the top 1,000 BM25 candidates for each of the 50 test queries using independent relevance-probability estimates.

System	nDCG@10	MRR
BM25, published configuration	0.3991	—
BM25 + mT5-XL, published configuration	0.6593	—
BM25, current configuration	0.3329	0.6852
BM25 + Jev, current configuration	0.6331	0.9033

Within the controlled comparison between the current BM25 and BM25 + Jev runs, nDCG@10 increases from 0.3329 to 0.6331, an absolute gain of 0.3001 and a relative gain of 90.2%. MRR increases from 0.6852 to 0.9033, a relative improvement of 31.8%. The MRR result implies that the first relevant passage is moved substantially earlier in the ranked list, while the nDCG result indicates broader improvement in the top-ten ordering.

The comparison with mT5-XL requires caution. The published BM25 + mT5-XL score is 0.6593, numerically higher than Jev’s 0.6331, but the corresponding published BM25 score is 0.3991 rather than the 0.3329 obtained in the current experiment. Since the candidate sets, BM25 configuration, qrels implementation, and other experimental details are not fully aligned, the paper correctly declines to interpret these values as a controlled Jev-versus-mT5 comparison. The valid claim is narrower: under the current Quati configuration, Jev substantially improves BM25 ranking effectiveness.

The pointwise formulation is also operationally significant. Unlike pairwise or listwise rerankers, Jev does not need to compare candidate documents explicitly or generate a permutation. This simplifies batching and permits independent scoring, but it also means that the model’s probability estimates must be sufficiently calibrated and comparable across documents for the same query. The paper uses the returned probability directly and does not fit a calibration model or combine it with BM25 scores. The observed gains therefore support the usefulness of the raw decision probability in this setting, but they do not establish that probability calibration is unnecessary in general.

Limitations and open questions

The relevance-assessment experiment is restricted to four Brazilian Portuguese collections, and the reference qrels are heterogeneous. Quati and JurisTCU include LLM-derived labels, whereas NormasTCU and Ulysses-RFCorpus rely on expert judgments with different numbers of annotators. This prevents a clean attribution of disagreement to Jev rather than to the reference-label process. The study also does not evaluate inter-annotator agreement under a common protocol, so comparisons between Jev’s $\kappa$ and human agreement remain indicative rather than causal.

The system-ranking analysis covers only NormasTCU and Ulysses-RFCorpus, despite generating Jev qrels for four collections. It uses 15 and 16 systems, respectively, and focuses on nDCG@10 and MRR. The reported $\tau$ values demonstrate ranking preservation for these system sets, but no confidence intervals, topic-level significance tests, or sensitivity analyses are reported. The threshold of 0.9 is a useful interpretive convention, not a guarantee that all downstream conclusions are interchangeable.

The reranking study is narrower still. It evaluates one corpus, 50 queries, one first-stage retriever, a candidate depth of 1,000, and pointwise scoring. Documents outside that candidate set cannot be recovered, so the result measures reranking quality conditional on BM25 recall. The paper does not report latency, throughput, monetary cost, variance across repeated model calls, or statistical significance. It also does not compare Jev under matched conditions with mT5-XL, monoT5, RankT5, or pairwise and listwise LLM rerankers. The claim that Jev is operationally cheaper therefore remains unsupported by this experiment, despite the broader cost motivation for System One models.

The most direct open question is whether the same pointwise probability formulation transfers from short web passages to full-length legal documents, where relevance may depend on distributed evidence, document structure, and specialized terminology. Additional questions concern candidate-set depth, score calibration, query-dependent truncation, listwise versus pointwise prompting, and the relationship between Jev confidence estimates and human escalation policies.

Conclusion

The paper presents evidence that Jev can support two distinct IR functions in Brazilian Portuguese. At the annotation level, its exact agreement with reference qrels is moderate, ranging from 52.9% to 64.9%, but disagreements are predominantly adjacent-level deviations and weighted $\tau$0 reaches 0.82. More importantly, Jev-generated qrels preserve retrieval-system rankings with Kendall’s $\tau$1 between 0.90 and 0.93 on the evaluated human-annotated collections.

As a reranker, Jev improves the authors’ BM25 run on Quati from 0.3329 to 0.6331 in nDCG@10 and from 0.6852 to 0.9033 in MRR. These results establish strong within-experiment effectiveness gains, but not superiority over dedicated rerankers, cost efficiency, statistical significance, or generalization beyond the tested corpus and candidate-generation setup. The paper’s principal contribution is therefore empirical and bounded: structured, probability-based decisions from Jev can preserve system-level evaluation conclusions and substantially reorder BM25 candidates in the evaluated Brazilian Portuguese setting (2610.23346).