This example shows how ARS surfaces a citation-version mismatch without auto-correcting the manuscript.
The author wants to cite Attention Is All You Need. Their corpus contains two concrete versions:
The scholar selects the proceedings record as the primary citable version, but the draft sentence quotes the arXiv v1 locator while the reference list renders proceedings metadata.
phase2_investigation/version_records.yaml:
schema_version: "1.0"
version_records:
- version_family_id: attention-transformer-family
canonical_slug: vaswani2017
primary_version_key: attention-neurips-2017
discovery_status: user_confirmed
reconciliation_note: "Proceedings record selected as default citation; arXiv v1 can be cited explicitly for preprint-specific claims."
known_versions:
- version_key: attention-arxiv-v1
citation_key: vaswani2017-arxiv-v1
kind: arxiv_preprint
title: "Attention Is All You Need"
year: 2017
venue: "arXiv"
doi: null
arxiv_id: "1706.03762v1"
url: "https://arxiv.org/abs/1706.03762v1"
publication_date: {value: "2017-06", precision: month}
metadata_provenance: arxiv_api
source_locator: "arxiv:1706.03762v1"
claim_scope_note: "Use only for text verified against the v1 preprint."
notes: null
- version_key: attention-neurips-2017
citation_key: vaswani2017-neurips
kind: proceedings
title: "Attention Is All You Need"
year: 2017
venue: "NeurIPS 2017"
doi: null
arxiv_id: null
url: "https://papers.nips.cc/paper/7181-attention-is-all-you-need"
publication_date: {value: "2017", precision: year}
metadata_provenance: user_override
source_locator: "papers.nips.cc:7181"
claim_scope_note: "Use for proceedings metadata and the final citable record."
notes: null
The preprint v1 states the Transformer architecture relies entirely on attention mechanisms (Vaswani et al., 2017).<!--ref:vaswani2017-neurips ok--><!--anchor:section:arxiv:1706.03762v1:sec3-->
The citation slug points to the proceedings record, but the prose and anchor point to arXiv v1.
VERSION_INCONSISTENT_CITATION: citation metadata, locator, or quoted claim mixes multiple records in version_family_id=attention-transformer-family. The rendered slug vaswani2017-neurips belongs to attention-neurips-2017, while the prose/anchor refers to attention-arxiv-v1. Select one version or explicitly separate the claims.
The arXiv v1 preprint states the Transformer architecture relies entirely on attention mechanisms (Vaswani et al., 2017).<!--ref:vaswani2017-arxiv-v1 ok--><!--anchor:section:arxiv:1706.03762v1:sec3-->
The proceedings version presents the Transformer architecture as relying entirely on attention mechanisms (Vaswani et al., 2017).<!--ref:vaswani2017-neurips ok--><!--anchor:section:papers.nips.cc:7181:sec3-->
The arXiv v1 preprint and the NeurIPS proceedings version both frame the Transformer as an attention-only architecture, but the manuscript cites the proceedings version as the primary record (Vaswani et al., 2017).<!--ref:vaswani2017-arxiv-v1 ok--><!--anchor:section:arxiv:1706.03762v1:sec3--><!--ref:vaswani2017-neurips ok--><!--anchor:section:papers.nips.cc:7181:sec3-->
The important rule is that each version-bound claim has its own slug and locator. ARS does not decide which version is "better"; the scholar selects the version appropriate to the claim.