Auditing and Repairing LLM-as-Judge Failures in a Production Text-to-SQL Pipeline
Factual evidence
What the source reports
A study reveals gpt-4o-mini used as an LLM-as-judge in text-to-SQL pipelines showed poor agreement with human annotators (kappa = 0.04).
Inspect the evidence
- Inclusion basis
- Enterprise AI
- Publisher and source type
- arXiv cs.CL — Computation and Language · RESEARCH
- Published by source
- 28 September 2026
- Collected by OneBench
- 29 Sept 2026, 03:01 UK
Stored source excerpt
arXiv:2609.30290v1 Announce Type: new Abstract: Production text-to-SQL pipelines often end with an LLM-as-judge whose agreement with human annotators has never actually been measured. When we…
Short excerpt from the collected text, not the full source. Use the source link to read it in context.
The factual summary is a OneBench synthesis, not a quotation or independent verification. Collection time is not publication time. Open the source for its full context; related reporting can share the same underlying announcement.
OneBench interpretation
Institutional assessment
So what
Relying on unvalidated LLM-as-judge setups for automated database querying introduces hidden failure modes that undermine production data reliability.
Do what
Review validation methodologies for automated evaluation pipelines with the team responsible for model risk management.