A Citation Is Not Support.

A footnote appears. A link turns blue. A retrieved passage sits under an answer in a tidy box. The reader relaxes. The sentence feels grounded.

That relaxation is often premature. A citation shows that something was pointed at. Support requires that the cited material actually warrants the specific claim being made. Conflating the two is one of the fastest ways to make a weak claim look sourced.

The distinction

I separate four layers that systems and reviewers often collapse:

1. Citation presence: a source is named, linked, or displayed.
2. Retrieval relevance: the passage is on topic for the question.
3. Entailment or support: the passage, with acceptable inference rules, warrants the claim.
4. Decision usefulness: the supported claim is the one the decision actually needs.

Layer one is cheap. Layer three is where audits live. Layer four is where buyer diligence lives.

Academic writing has long distinguished citation norms from logical support. A paper can be cited without entailing every sentence it appears beside. RAG systems recreate the visual habit of scholarship without guaranteeing the logical relation. Evidence Bundler, in my own apparatus, deliberately stops at nominating candidate passages. Claim Audit Lab applies a separate support judgment. The split is architectural, not pedantic: nomination is not verdict.

What established practice already covers

Information retrieval has always had a gap between “found a relevant document” and “the document proves the answer.” Fact-checking pipelines, entailment research, and legal citation practice all stress that proximity is not warrant.

In regulated review, subject-matter experts learn to ask whether the record answers the exact question on the form, not merely whether a record exists. Buyer diligence repeats the same move when a vendor appendix is thick but off-target.

Evaluation practice is catching up. Teams scoring “groundedness” or “faithfulness” are trying to operationalize layer three. The hard part is that layer two alone can look impressive in a demo.

What remains poorly named or operationalized

Three failure modes stay common.

Citation presence bias: a link or reference number is treated as finished work. Relevance laundering: a passage mentions the topic but does not establish the proposition. Inference smuggling: a reasonable-sounding bridge is added in the answer text while the citation only covers the bridge’s first step.

Near-miss passages are especially dangerous. They reward the reviewer who skims for topical words, not the reviewer who asks what would need to be true for the claim to hold.

A working model

Treat each atomic claim as the unit of audit. For every claim sentence, ask four questions mapped to the four layers. If the process stops after layer one or two, the claim may be decorated, not supported.

For human review, I prefer explicit labels over binary “grounded / ungrounded” when stakes are high: supported, partially supported, reasonable inference, unsupported, contradicted, not checkable. Binary scoring hides the exact failure mode.

For automated systems, keep candidate retrieval separate from support verdict generation. When one component does both, measurement and debugging become harder.

A concrete example

Synthetic minimal pair.

Claim: “The system prevents duplicate CAPA assignment.”

Cited passage: “CAPAs are tracked in the QMS and assigned to owners for investigation.”

The passage is relevant. It does not entail prevention of duplicate assignment. At best it supports a weaker claim: that CAPAs are tracked and assigned. The stronger claim smuggles a guarantee the source does not offer.

Synthetic RAG case.

Claim: “The 2024 guidance requires annual model retraining for all ML systems used in GxP decisions.”

Retrieved passage: A guidance index page listing document numbers with one sentence noting that computerized systems should remain in a validated state.

The system produced a citation box. The decisive regulatory language is not in the box. The answer feels sourced because the UI pattern says it is.

Citation-vs-support checklist

Step	Review question	Pass / fail signal
1. Claim text	What is the exact proposition?	Claim is atomic and falsifiable
2. Cited artifact	What document or passage is offered?	Artifact identified with location
3. Topic match	Does it mention the subject?	Relevance without entailment is not enough
4. Warrant	Does it establish the proposition or only context?	Distinguish support from background
5. Inference gap	What bridge did the author add?	Label inference explicitly
6. Decision fit	Is this the claim the decision needs?	Decision-useful proposition

Decision rule: do not mark a claim supported in a register, scorecard, or buyer response unless step 4 passes for the exact wording used. If step 5 carries load, downgrade to partial support or reasonable inference.

Current conclusion

The philosophical distinction between citation and support is not new. The operational contribution is to keep the layers separate in tooling, review language, and buyer-facing materials. I have high confidence in the distinction as a review practice. I do not claim that any single tool fully solves entailment in open domains.

This note is concept analysis, not a benchmark result. I have not published comparative false-support rates from Claim Audit Lab runs here.

Open questions and limits

Some domains tolerate stronger inference rules than others. Legal, scientific, and quality-review contexts differ on what “reasonable inference” means. A public codebook has to declare those rules.

Long documents introduce position effects: decisive sentences buried late are easier to miss. That is a measurement design problem for evaluators, not an excuse to skip human review on consequential claims.

Related work and version notes

Related concepts include candidate evidence versus support verdict, documentation-presence bias, near-miss evidence, and unsupported specificity. Apparatus references include Evidence Bundler and Claim Audit Lab. The research map proposes minimal-pair micro-evaluations on related versus entailing passages.

What this changes

RAG and agent builders should design UI and logs that show nomination separately from support. Claims reviewers gain a checklist that resists the comfort of a link. Buyers can ask a sharper question: “Show me the passage that establishes this sentence, not one that mentions the topic.”

The smallest next action is to take one answered claim from a recent system output and run the six checklist rows. If step 4 fails, fix the claim or fix the evidence before reuse.

Commercial bridge

Buyer-readiness work and Evidence Intelligence research sprints use this separation routinely when decomposing vendor claims and building evidence registers. The checklist travels without the engagement. Company-specific claim registers and remediation sequencing remain paid work.