Evidence, claims and time

One visit, two reports, a later receipt

In our fictional Harbour Court dossier, an inspection takes place on 12 August. The first report is produced on 14 August. A corrected report follows on 20 August, and a receiving system records it on 21 August. Both reports refer to the same visit. A query returning 21 August as “the inspection date” is not merely imprecise: it answers a different question.

Identify the inspected subject, inspection activity, findings, report versions and receipt separately where the questions require them. A report can carry several assertions, and an assertion may depend on evidence beyond that report. An address on the cover does not establish which of Harbour Court’s two dwellings was inspected.

Start with the assertion whose evidence matters

“The report exists”, “the report states that the wall has a defect” and “the wall has that defect” are three distinguishable claims. A file can establish what text it contains without establishing the accuracy of the observation. A model should not silently upgrade evidence of an assertion into evidence that the assertion is true.

Suppose report version 1 identifies Flat 2 on its cover, but its room descriptions indicate Flat 1. Version 2 corrects the cover. Preserve the correction and the relationship between versions; do not silently edit the old subject reference and leave consumers believing it was always unambiguous. The correction concerns a description, not necessarily the inspected dwelling or the visit’s identity.

Now suppose only the conclusion changes after another professional reviews the photographs. That revision need not imply a second inspection either. The model needs the identity and provenance of the revised assertion, not an invented event to explain why the text differs.

Describe production, responsibility and derivation separately

PROV-O distinguishes entities, activities and responsible agents. A report version can be an entity, its production an activity, and its author an agent. Derivation links an entity to evidence used in producing it; attribution identifies responsibility. Neither relation is a truth certificate. PROV-O’s starting-point terms provide the vocabulary distinctions.

A plan belongs with the activity association when that qualified pattern is used: activity → qualified association → agent, role and plan. This preserves which participant followed which procedure. Do not scatter plans onto generic attribution records or impose a single agent or plan merely because the first example has one. PROV-O’s Association pattern explains this qualification.

The adopted provenance concern also distinguishes four origins: causal extraction, editorial modelling history, mapping-decision provenance and discovery-system origin. A tool that extracted a statement is not the source author. The place where a document was discovered is not necessarily its publisher. The person who revised an ontology definition is not the person who inspected the property.

These distinctions are useful within one dossier. A report can be the source of a candidate domain definition, and separately evidence for a finding about a dwelling. The two uses describe different claims. A single “source” label cannot reliably explain both.

Choose the smallest honest unit of provenance

Resource-level provenance is efficient when all relevant statements share origin and confidence and the resource is created and validated as an immutable unit. ODR-0052 permits that pattern under those conditions. It is not permission to attach one source to an evolving resource whose facts come from several places.

Imagine an extracted report resource containing the visit date and conclusion from version 1. A modeller later adds a corrected subject from version 2. Retaining a resource-wide claim that everything came from version 1 is now misleading. Different statement evidence requires selective RDF 1.2 reification under the adopted carrier profile.

Statement reification makes the particular assertion addressable for provenance. Qualified generation instead describes details of the event that generated an entity. They can coexist, but they answer different questions. ODR-0052 makes qualified generation conditional on an activating local amendment; selecting statement-level provenance does not silently activate that additional pattern or deferred shapes.

Respect the reused vocabulary’s property kinds. A source-file path is a literal string, not automatically a PROV Location resource. The adopted correction keeps the local source-file datatype property independent of the object-valued prov:atLocation. A useful-looking hierarchy is not justification for an incompatible subproperty declaration.

A transition has a subject; a record has its own creation time

ODR-0053 selects immutable transition records for applicable state history. A transition record is a prov:Entity identifying its subject, previous state, new state and effective instant. Its optional prov:generatedAtTime describes the record’s creation time; prov:wasGeneratedBy identifies the recording activity. An activity has its own start and end times rather than borrowing an entity’s generation timestamp.

The state values belong to the same locally closed state space and are distinct. Explicit allowed-successor relations represent adjacency. A point-in-time snapshot can be derived where it has a use, but the transition chain is primary. A “current status” field alone cannot reconstruct what an earlier query would have found.

The selected OWL-Time profile supplies bounded temporal resources and relationships. Instants require timestamp values; proper intervals require a beginning and an end later than the beginning. These constraints check declared time data. They do not infer an absent observation, invent a domain lifecycle or settle whether an entity survives a subdivision.