What Do LLM Judges Actually Verify? Matched Controls and Record Gaps for Document Currency

Unnati Shah, Samruddhi Kakade, Aum PatelarXiv 2026 · arXiv

This research paper investigates a critical vulnerability in RAG (Retrieval-Augmented Generation) systems: the tendency of LLM judges to blindly trust metadata over textual evidence. When verifying if a document citation is still "in force," these judges often defer to incomplete amendment records, failing to catch errors even when the documents themselves clearly show a replacement has occurred.

TL;DR

LLM judges frequently "shut off" their reasoning capabilities and defer to provided amendment records, even when those records are incomplete and contradict the document text. While explicit warnings about record incompleteness can restore some reasoning, it often comes at the cost of increased false alarms, highlighting a fundamental tension between metadata trust and textual analysis.

What Problem Does This Paper Address?

In domains like law, regulation, and technical standards, documents are constantly amended or replaced. Systems that answer questions based on these collections often use an "LLM judge" to verify that the citations in an answer are still current. To help, builders often provide these judges with an amendment record—a lookup table showing which documents replaced which.

However, these records are rarely perfect; they lag behind new updates or miss links during extraction. The authors ask: What happens when the record is silent about a replacement, but the document text itself provides evidence of it? Does the judge act as a smart reasoner, or does it become a glorified (and fallible) lookup tool?

Methods

The researchers developed a test harness that generates matched sets of answers to probe a judge's reasoning. For every replacement link (e.g., RFC A is replaced by RFC B), they created four types of items:

  • Fault: Cites the old doc after it was replaced (should be flagged).
  • Control: Cites the successor doc (should pass).
  • Date Control: Cites the old doc before it was replaced (should pass).
  • Decoy: Cites a doc that was updated but not replaced (should pass).

The core of the experiment involves "damaging" the amendment record by removing half the replacement links, creating gap faults. They then tested four frontier models (Gemini 3.1 Pro, GPT-5.5, Claude Opus 4.5, and Gemma 4 26B) across different context settings, varying how the record was presented (e.g., as an authoritative "lookup," an "unverified note," or with a "warning").

Datasets

  • IETF RFC Index: The primary dataset, consisting of 48 replacement links from the Internet Engineering Task Force. This provides a high-quality "gold standard" for technical standards.
  • Indian Rooftop-Solar Regulations: A descriptive case study involving 111 regulatory instruments where the amendment record was noisily extracted by an AI pipeline.

Key Insights and Ideas

  • The Deference Trap: When a record is present, judges often stop looking at the text. If the record is silent (effectively saying a doc is "in force"), judges pass the answer, ignoring the successor document sitting right in front of them.
  • Authority Matters: Simply changing the wording of a gap (e.g., from "in force" to "no replacement recorded") does little to help. However, lowering the record's perceived authority (labeling it an "unverified note") or providing a specific warning about incompleteness nudges models back toward textual reasoning.
  • Surface Similarity Shortcuts: The study found that judges often use title similarity as a proxy for replacement. While this works for RFCs (which often share titles), it leads to false alarms on "updates" that aren't actually "replacements."
  • Reasoning vs. Lookup: A judge given a perfect record looks highly capable, but this is an illusion of performance. The true measure of a judge's value is what it adds beyond a simple database lookup.

Key Results

The results showed a dramatic drop in error detection when a flawed record was introduced. As shown in the "gap faults" analysis, models that were nearly perfect using text alone became significantly worse when given a silent record.

Catches on the 24 gap faults per judge
  • Gemini 3.1 Pro: Caught 23/24 gap faults with text alone, but only 1/24 when the record was present.
  • GPT-5.5: Dropped from 23/24 to 9/24.
  • Claude Opus 4.5: Dropped from 17/24 to 9/24.

Interestingly, while warnings restored catch rates, they also triggered "validity false alarms" on decoys. For GPT-5.5, false alarms on updates rose from 0/48 (with a correct record) to 8/48 (with a warning).

Limitations and Future Work

  • Small Sample Size: The study relied on 48 links; while statistically significant for these models, larger-scale testing is needed.
  • Corpus Specificity: Technical standards like RFCs have specific patterns (like abstract overlap) that judges might exploit. Legal contracts, where amendments may look nothing like the original, might yield different results.
  • Mechanism: The paper identifies that deference happens but cannot fully explain why (e.g., is it a result of RLHF, prompt salience, or training data?).
  • Future Work: The authors suggest that future evaluations must include "gap faults" to ensure LLM judges are actually verifying content rather than just echoing metadata.

Summary updated

Abstract

LLM judges often check, against an amendment record, whether the documents an answer cites are still in force. Such records are rarely complete. From a table of replacements, our test generates matched answers citing a replaced document, its successor, the replaced document before its replacement, and an older document that a newer one updated without replacing; it removes links from the record and scores each finding by its stated reason. On 48 replacements from the IETF RFC index, judges given a correct record caught nearly every replaced citation, as a lookup would. On the 24 faults whose link we removed, every judge caught more from the texts alone than with the record: Gemini 3.1 Pro 23/24 against 1/24, GPT-5.5 23/24 against 9/24, Claude Opus 4.5 17/24 against 9/24 and Gemma 4 26B 13/24 against 0/24. The drop recurred on a second draw of removed links. Most misses carried no finding; in an exploratory run with a conflict code, only Gemini used it often. Rewording the gap as "no replacement recorded" did not help; presenting the record as an unverified note, or warning that it was incomplete, restored most lost catches for Gemini, GPT-5.5 and Claude (significant after correction for the first two) but at most 5 of 13 for Gemma; where catches returned, so did false alarms on updates. With both texts, judges mostly told replacements from updates; every update flagged had a near-identical title. A descriptive case study on Indian rooftop-solar rules is appended; we release the harness.
BibTeX
@misc{shah2026what,
  author = {Unnati Shah and Samruddhi Kakade and Aum Patel},
  title = {{What Do LLM Judges Actually Verify? Matched Controls and Record Gaps for Document Currency}},
  year = {2026},
  eprint = {2611.01311},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.01311},
}
Figure