1. Introduction

A citation performs two jobs at once. It credits prior work, and it lets a reader retrace the path from claim to evidence. When the second job fails silently, the reference list stops being a chain of custody and becomes decoration. That failure has become inexpensive: a fluent drafting system will produce a reference that has the shape of a real citation — plausible authors, a real journal, a well-formed identifier — without a corresponding record anywhere.

Editorial practice has largely responded by asking for more references rather than better-resolved ones. We argue the opposite: fewer, resolved, and metadata-matched references are strictly more informative than long lists that no one has checked.

2. Data and method

We drew 4,120 preprints deposited between January 2025 and March 2026 across six subject areas, extracted reference lists with a rule-based parser, and resolved each candidate identifier against the public DOI registry. For every resolution we compared four fields against the citation as printed: container title, article title, first author surname, and publication year. A citation was recorded as mismatched when any field diverged beyond normalisation.

Three outcomes were possible for each reference: no identifier present, identifier present but unresolvable, or identifier resolved. Resolved references were then classified as matched or mismatched.

3. Results

Absent identifiers were common but visible: an editor scanning a reference list can see that a DOI is missing. Unresolvable identifiers were rarer and were caught reliably by a single registry lookup. Metadata mismatch was the persistent case — the identifier resolves, so an automated check that stops at resolution reports success, while the resolved record describes a different paper than the one cited.

Mismatch clustered in citations supporting the most consequential claims in a manuscript, which is consistent with citations being added late, to shore up an argument, rather than accumulated during reading.

4. A minimal verification protocol

First, require an identifier for every reference that has one, and require an explicit statement for those that legitimately do not. Second, resolve every identifier. Third — and this is the step usually skipped — compare the resolved record against the printed citation field by field, and return divergences to the author as corrections rather than silently repairing them. Fourth, check retraction and withdrawal status, and treat a retracted source supporting a central claim as blocking.

The protocol is cheap enough to run on every submission and strict enough to make a fabricated reference list an unproductive strategy.

5. Limitations

Our corpus is preprints, not accepted articles, so it describes what authors submit rather than what journals publish. Rule-based parsing under-extracts from unstructured reference formats. Field comparison cannot distinguish a genuine transcription slip from an invented citation, and we make no attempt to infer intent.