Citations as an evidence contract: locate, support, limit

A model to inspect

Evidence-bearing answer

01State claim
02Attach locator
03Check reader access
04Show limit

No evidence is a first-class outcome.

Original conceptual diagram · not a live trace or measured result.

Define what a citation promises

Before choosing a UI, state the promise. A citation should let the intended reader locate a source, identify which portion supports the nearby claim, and distinguish source fact from the assistant’s synthesis. It should not imply that the source endorses every inference in the sentence. Separate a direct statement, a calculation, a comparison, and a recommendation; they need different evidence or a clear label.

The W3C Web Annotation model separates a body from a target and supports selectors that identify part of a resource. That is a useful mental model: a link to an entire policy is an identity, but a section, page, quote range, or stable heading is the locator. Use only locators your renderer and source lifecycle can maintain.

Build a claim-to-evidence record

At answer time, create a small record for each material claim: claim ID, source ID and version, locator, retrieved passage identifier, access result, and support type. Support type may be direct, calculation input, background, or insufficient. The record need not be shown in full to a user, but it allows an operator to inspect why a citation appeared. Do not fabricate a page number when extraction lost pagination.

Worked example, hypothetical: “Returns are accepted within 30 days for unopened items” needs the active returns policy and the section that contains both the period and condition. A shipping FAQ may be related background, not support. If the policy says 30 days but omits unopened items, answer the supported part and label the condition as unresolved.

Design the empty-evidence response

No supporting evidence is a valid result. Tell the user that the available material does not establish the answer, identify the missing source type where that is safe, and offer a narrower question, a fresh lookup, or a human owner. Never fill the gap with a plausible rule merely to retain conversational momentum. This is particularly important for eligibility, safety, legal, medical, and contractual questions.

Partial evidence calls for a split response. State the supported fact with its citation; then state the limitation: “This policy gives the standard window, but I could not verify an exception for your product.” That is more trustworthy than a single confident paragraph with one link.

Make citation access part of rendering

Check that the caller can open the cited source at the time the answer is delivered. A retrieval trace may identify a document that has since been revoked or moved. If direct opening is not appropriate, provide a permitted internal title and route rather than an opaque URL. Do not expose a secret filename in an error state.

PROV-DM distinguishes entities, activities, and agents in provenance descriptions. Use that distinction modestly: store the source artifact, the retrieval or transformation activity, and the service version that produced the answer. It helps reviewers investigate a claim without pretending the provenance model proves factual correctness.

Review the support, not the link count

Sample answers and ask four questions: can a permitted reader reach the source; does the locator land near the asserted fact; does the passage support the wording; and are the limitations visible? Review citations after source updates, extraction changes, and access-policy changes. A higher citation count can make an answer harder to audit if each marker is vague.

Counterexample: three citations at the end of a paragraph about pricing, eligibility, and timing do not show which source supports which fact. Prefer claim-adjacent references or a structured evidence panel. Keep aggregation and judgment visibly separate from quotations.

Set a modest release gate

Release gates can be simple: every high-impact answer has a support record or an explicit no-evidence outcome; every displayed citation passes an access check; sampled locators are reviewed after ingestion changes; and source versions are retained long enough to investigate an answer. Do not claim universal factuality from those checks.

Sources can be wrong, stale, ambiguous, or inconsistent. An evidence contract makes that condition visible and actionable. It cannot turn a weak corpus into a reliable decision system. Record the reviewer, access result, and reason whenever a citation is suppressed or replaced. Recheck changed sources before reusing old answer evidence.

Take it into the review

Claim evidence card

Claim	Support type	Locator	Access check
return window	direct	Returns / standard window	open now
opened condition	insufficient	no active clause	do not assert
refund estimate	calculation input	order value + policy	operator review
recommendation	synthesis	linked facts	label as advice

A starting artifact to adapt to your service—not a ready-made policy, compliance certificate or test result.

Primary reading

Sources and limits

These links support the architecture, policy, or product behavior discussed above. Vendor documentation describes vendor features; it is not independent proof of performance. Current details should be rechecked before a production decision.

What the sources establish

Web Annotation Data Model

The model describes annotations with bodies, targets, and selector mechanisms for parts of resources.

Limits: It does not prescribe a chatbot citation UI or determine whether a source supports a claim.

Checked 2026-09-19 · W3C · source publication date not established.

Open original source ↗

PROV-DM: The PROV Data Model

PROV-DM defines concepts for describing entities, activities, and agents in provenance.

Limits: Provenance records explain lineage; they do not independently verify truth or source authority.