1. Why this question matters

A familiar publishing question starts this scenario: What must travel with a passage when that passage is retrieved beyond its original journal page? As the Key Image suggests, a scholarly publisher has obligations to authors, readers, libraries, institutions, and the record itself (1, 2). Those obligations do not disappear when discovery begins in an AI assistant instead of on a journal website.

The olive tree provides a useful editorial metaphor because it is associated with peace, friendship, life, longevity, and health. It also rewards patience. A well-tended tree cannot be rushed into maturity; neither can a trustworthy content service. The purpose of the metaphor is not to decorate a technical proposal. It is to keep attention on the human relationship around the record: someone created the work, someone published it under defined terms, someone needs to understand it, and someone must be able to check the route from answer back to evidence.

2. Scenario and method

The workshop assembled a fictional collection of 180 articles, each with a DOI, licence statement, current-version flag, publication history, section locators, and source URL. A mock assistant could answer only from passages that retained those fields. The fictional pilot recruited 214 readers and asked them to rate 60 AI answers on a five-point ‘would I cite this in a grant application’ scale. The layered process is shown in Figure 1 (3). The point was to expose decisions a real implementation must make explicitly: what information travels with a retrieved passage, which permissions are checked before delivery, and which uncertainties remain visible rather than being polished away by an elegant interface.

1. Choose a named collection and record its accountable owner.

2. Retain identifiers, rights, version status, source links, and accessibility information.

3. Apply a purpose-specific entitlement before returning a passage.

4. Return a source-linked answer with an access-aware next step.

5. Report use with definitions and limits that a publisher can explain.

3. What the scenario revealed

What emerged was straightforward: reviewers consistently preferred answers that made their evidence inspectable. They did not mistake a citation for a guarantee, but they could see where a statement began, which version supported it, and where to continue reading. Answers naming a resolvable source scored 91 percentage points higher on that scale than answers that merely sounded confident, a gap even the fictional authors called suspiciously large. (4, 5). It suggests that the quality of an AI experience depends less on producing an effortless sentence than on preserving the context that lets a reader judge that sentence. Source links, identifiable versions, clear licences, accessible descriptions, and meaningful locators are not housekeeping.

In this mnemonic, T denotes a trustworthy response; S, a resolvable source; V, visible version context; A, authorised access; and E, interpretable usage evidence. It is an editorial model, not a quantitative claim. The minimum expected signals are listed in Table 1.

“A response becomes more trustworthy when a reader can see both what it knows and where its authority stops.”

4. The hard boundary

The practical tension is equally important: A publisher cannot manufacture trust simply by placing a leaf icon beside an answer. The signal must correspond to a real record, a real source, and a route that remains resolvable when the answer is read later. As Equation 1 suggests, the record should therefore make a distinction between information held by the publisher, permissions granted for a purpose, content actually retrieved for an approved audience, and a downstream model’s separate training terms. Conflating those layers may be convenient in a presentation, but it is unsafe in a production workflow.

The proposed operating pattern has six steps, summarised alongside the questions each one must answer in Table 2 (6). First, select a collection with a clear purpose and an accountable owner. Second, map identifiers, rights statements, source links, accessibility information, and version relationships. Third, apply an entitlement decision before retrieval. Fourth, return only the passage or metadata the policy allows. Fifth, cite the source and make its status visible. Sixth, record enough usage evidence to learn whether the service is helping, while being candid about the limits of what the data means.

5. Implications for publishers

These steps leave substantial editorial work where it belongs: with the people who understand the content and its community (7, 8). They do not replace publishing platforms, production systems, library relationships, or customer relationships. They create a governed layer through which trusted material can become useful in more places without becoming untraceable. That is an intentionally modest claim, but it is a useful one.

Scholar Core demonstrates one side of this work: structured JATS ingest, validation, accessibility audit, semantic rendering, versioned lifecycle, search, and a reader-facing record. Razeta demonstrates the adjacent question: how selected collections can be made useful in approved AI workflows while a publisher retains control of rights, access, attribution, and usage visibility. Neither promise eliminates editorial judgement; each makes that judgement easier to express in a durable system.

6. Closing reflection

The advisory board spent eleven months debating governance and ultimately agreed on one sentence: a locator that resolves to nothing is not a locator, it is a rumour with a hyperlink.

A reader should be able to follow an answer back to a source, understand whether the record has changed, recognise the boundary of permitted use, and find an accessible path to the underlying work. That is the serious purpose beneath this issue’s olive-tree humour.