Key takeaway
Retain the DOI relationship and the actual resolved version and files. A persistent identifier is not, by itself, an immutable-byte guarantee.
A citation can stay stable while its described resource changes
Two evaluation reports cite the same dataset DOI but disagree about their record period. Before diagnosing a model difference, determine what each report actually resolved and downloaded. A common citation string is useful discovery evidence, but it may leave the version question open.
DataCite’s versioning guidance permits the same DOI with updated metadata for minor content changes, recommends a new DOI for major changes and leaves the major/minor boundary to stewards. It also describes a canonical DOI representing several versions. Interpret the identifier under the publisher’s policy rather than treating every DOI as a frozen delivery.
Complete a resolution ledger before accepting a replay claim
Hypothetical example: a publisher uses group label G for a canonical DOI and labels V1 and V2 for two version-specific DOIs. These labels stand for invented records; no unregistered DOI or real publisher inventory is asserted. The evaluator has an older result for V1 and proposes a new run from the group landing page.
The group metadata lists both versions. V2 describes an added operating year. A later minor correction also updates the resource associated with V1. The following decisions preserve the uncertainty instead of selecting the newest-looking landing page and calling it the old input.
| Observed illustrative evidence | What it identifies | Replay decision |
|---|---|---|
| G hasVersion V1 and V2 | A group of versions | Choose and record the intended member |
| V2 isNewVersionOf V1 | A sequence of resource versions | Do not relabel a V1 result as a V2 result |
| V1 metadata notes a minor correction | Same DOI; described content updated | Compare actual files and correction history |
| Original result lacks retained files/version evidence | Citation alone remains | Hold exact-input replay claim pending evidence |
Read the relation type rather than a generic related link
DataCite distinguishes sequence relations such as IsNewVersionOf and IsPreviousVersionOf from HasVersion and IsVersionOf for grouping. Its related-identifier guidance also distinguishes corrections and alternate formats. A related resource is not automatically the next version of the same input.
For the fictional record, V2 pointing to V1 as its predecessor explains the sequence. G pointing to both explains the group. If another identifier describes JSON instead of CSV, inspect whether it is a format variant rather than an additional observation period. Record the actual relation type with the target.
The publisher’s version policy remains relevant. A steward may call a correction minor and keep the DOI, while that correction still changes a decision-critical label in the evaluator’s task. The evaluator needs a specific change assessment rather than assuming minor means immaterial to every downstream result.
Bind the run to the retrieved artifact
Retain the requested DOI, resolution time, resulting landing page, version statement, relation metadata and files actually obtained. For each file, preserve its applicable byte identity and the approved package manifest. A landing page URL and a checksum answer different questions and are useful together.
In the example, the original report cannot establish an exact-input replay until its missing artifact evidence is recovered. The new run can still be a valid evaluation of a specifically recorded V2 package. It must be labeled as that evaluation, rather than a reproduction of the unverified V1 input.
Where an old version is unavailable, state that limit. Do not invent a historical hash, infer a file’s old contents from today’s metadata or silently substitute a newer release. A documented substitution may be useful, but it changes the comparison and should be reviewed as such.
Keep citation, integrity and permission distinct
The resolution ledger gives the reviewer a concrete answer: which resource relationship was followed, which version was selected and which artifact the result actually used. It does not establish that the observations are accurate, the evaluation is representative or the intended reuse is licensed.
Use due diligence for the publisher’s version policy and inventory for the retained artifact references. A VOID introduction does not register a DOI, guarantee access to historical versions or grant use of a cited dataset. The owner and recipient still decide the actual sample and license scope separately.