@inproceedings{liu-etal-2026-auditing,
title = "Auditing and Repairing {LLM}-as-Judge Failures in a Production Text-to-{SQL} Pipeline",
author = "Liu, Haowei and
Wu, Hsin-Tai and
Fang, Yi",
editor = "Atanasova, Pepa and
Haffari, Gholamreza and
Wang, Minghan",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-industry.18/",
pages = "275--287",
ISBN = "979-8-89176-460-6",
abstract = "Production text-to-SQL pipelines often end with an LLM-as-judge that nobody has checked against humans. We audited ours and found it broken: the deployed gpt-4o-mini judge agreed with two-author gold at Cohen{'}s {\ensuremath{\kappa}} = 0.04 on a disagreement-enriched set and 0.42 on a uniform-random spot-check, with most over-flags traceable to a single mechanism we call Grade-Hallucination. Replacing it with self-hosted Qwen3.6-27B ({\ensuremath{\kappa}} = 0.72) reaches the same range as Claude Opus 4.7 ({\ensuremath{\kappa}} = 0.71); the comparison is underpowered at n = 96, but the deployment decision is driven by Qwen costing about 1/300 per call. Ensembling is not a free win: pairing the weak judge with a stronger one degrades agreement; three strong judges with unanimity routing reach {\ensuremath{\kappa}} = 0.79 at 89.7{\%} auto-coverage. Applying the audit recipe out-of-domain flags 25.5{\%} of BIRD-financial{'}s expert-authored gold SQLs as likely substantive issues under our annotation protocol. Code and pre-registration are at https://anonymous.4open.science/r/synca-audit-CBE1; the datasets are deferred to the post-publication release for anonymization reasons."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="liu-etal-2026-auditing">
<titleInfo>
<title>Auditing and Repairing LLM-as-Judge Failures in a Production Text-to-SQL Pipeline</title>
</titleInfo>
<name type="personal">
<namePart type="given">Haowei</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hsin-Tai</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yi</namePart>
<namePart type="family">Fang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track</title>
</titleInfo>
<name type="personal">
<namePart type="given">Pepa</namePart>
<namePart type="family">Atanasova</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Gholamreza</namePart>
<namePart type="family">Haffari</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Minghan</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-460-6</identifier>
</relatedItem>
<abstract>Production text-to-SQL pipelines often end with an LLM-as-judge that nobody has checked against humans. We audited ours and found it broken: the deployed gpt-4o-mini judge agreed with two-author gold at Cohen’s \ensuremathąppa = 0.04 on a disagreement-enriched set and 0.42 on a uniform-random spot-check, with most over-flags traceable to a single mechanism we call Grade-Hallucination. Replacing it with self-hosted Qwen3.6-27B (\ensuremathąppa = 0.72) reaches the same range as Claude Opus 4.7 (\ensuremathąppa = 0.71); the comparison is underpowered at n = 96, but the deployment decision is driven by Qwen costing about 1/300 per call. Ensembling is not a free win: pairing the weak judge with a stronger one degrades agreement; three strong judges with unanimity routing reach \ensuremathąppa = 0.79 at 89.7% auto-coverage. Applying the audit recipe out-of-domain flags 25.5% of BIRD-financial’s expert-authored gold SQLs as likely substantive issues under our annotation protocol. Code and pre-registration are at https://anonymous.4open.science/r/synca-audit-CBE1; the datasets are deferred to the post-publication release for anonymization reasons.</abstract>
<identifier type="citekey">liu-etal-2026-auditing</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-industry.18/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>275</start>
<end>287</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Auditing and Repairing LLM-as-Judge Failures in a Production Text-to-SQL Pipeline
%A Liu, Haowei
%A Wu, Hsin-Tai
%A Fang, Yi
%Y Atanasova, Pepa
%Y Haffari, Gholamreza
%Y Wang, Minghan
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-460-6
%F liu-etal-2026-auditing
%X Production text-to-SQL pipelines often end with an LLM-as-judge that nobody has checked against humans. We audited ours and found it broken: the deployed gpt-4o-mini judge agreed with two-author gold at Cohen’s \ensuremathąppa = 0.04 on a disagreement-enriched set and 0.42 on a uniform-random spot-check, with most over-flags traceable to a single mechanism we call Grade-Hallucination. Replacing it with self-hosted Qwen3.6-27B (\ensuremathąppa = 0.72) reaches the same range as Claude Opus 4.7 (\ensuremathąppa = 0.71); the comparison is underpowered at n = 96, but the deployment decision is driven by Qwen costing about 1/300 per call. Ensembling is not a free win: pairing the weak judge with a stronger one degrades agreement; three strong judges with unanimity routing reach \ensuremathąppa = 0.79 at 89.7% auto-coverage. Applying the audit recipe out-of-domain flags 25.5% of BIRD-financial’s expert-authored gold SQLs as likely substantive issues under our annotation protocol. Code and pre-registration are at https://anonymous.4open.science/r/synca-audit-CBE1; the datasets are deferred to the post-publication release for anonymization reasons.
%U https://aclanthology.org/2026.emnlp-industry.18/
%P 275-287
Markdown (Informal)
[Auditing and Repairing LLM-as-Judge Failures in a Production Text-to-SQL Pipeline](https://aclanthology.org/2026.emnlp-industry.18/) (Liu et al., EMNLP 2026)
ACL