RAG · cross-check · conflict gate

rag-multidoc-crosscheck

Fictional policy Q&A retrieves across every source, exposes contradictory claims, and refuses to manufacture a winner.

Public · fictional corpus
The problem

The top chunk can hide a stale or regional contradiction.

In policy and compliance, silently selecting one source is often worse than refusing. A safe system must show both claims and route disagreement to a person.

Input

One question across two fictional policy versions

Q: How many unused PTO days can I carry over into the next year?
Sources: pto_policy_2023.md + pto_policy_2024.md
The money shot

Both claims survive retrieval; the conflict becomes the answer

Verdict accuracy1.012 questions
Citation coverage1.0Non-refused questions
Conflict recall1.03 labelled conflicts
pto_policy_2023#1

5:day

Employees may carry over up to 5 unused PTO days into the following calendar year.

pto_policy_2024#1

8:day

Employees may carry over up to 8 unused PTO days into the following calendar year.

VerdictCONFLICT
Decisionneeds_review
ReasonComparable values differ; no source is silently preferred.
How it's verified

A labelled harness tests every verdict class.

The 12-question offline evaluation covers CONSENSUS, CONFLICT, SINGLE_SOURCE, and INSUFFICIENT. It separately scores verdict accuracy, all-source citation coverage, and recall on the three conflicts because a missed contradiction is the costly failure.

Honest limitations
  • All five documents and claims are fictional.
  • Offline claim extraction uses simple number-and-unit regex; production needs unit conversion and richer normalization.
  • The small labelled set does not establish production accuracy on unseen policies.
  • Conflict detection does not decide which source is authoritative; a human resolves the discrepancy.