Research task 04 / provenance and reporting

Carry Uncertainty Into the Record

A reusable classification preserves more than its preferred label. It carries the source trail, competing interpretations and the conditions under which the decision may change.

Publish a claim, not a detached label

A spreadsheet cell reading “Gothic Revival” tells a future reader very little. It does not identify the building component, show whether the label came from a model or a historian, or explain which source supports it. Treat the classification as a claim with an author, evidence, scope and review status. That structure makes the record useful even when the preferred interpretation changes.

Keep source evidence separate from derived data. The original photograph, a cropped image, an extracted feature and a predicted label are different objects. Link them with explicit relationships and stable identifiers. A future researcher should be able to trace the label back through the transformations without confusing a generated description with the source itself.

Conceptual archival folders, a facade model and stacked record sheets arranged as an evidence bundle — Conceptual evidence bundle. It does not represent an official heritage register.

Define a minimum record

Assign a record identifier and identify the building and component being classified. Include the category vocabulary and its version, since a label can change meaning when definitions are revised. Store source references with access information and rights notes. Preserve dates as qualified statements about events rather than forcing every date into an unqualified construction-year field.

For a model-assisted claim, record the model name or identifier, version, relevant configuration and input references. If text prompts were used, preserve the prompts and the output that informed the decision, subject to privacy and rights constraints. Include the evaluation context needed to interpret scores. A number without its method, category set and scope is not a durable explanation.

Make review status unambiguous

Use a small, documented status vocabulary. “Proposed” means a suggestion exists. “Under review” means a named process is active. “Reviewed” means a reviewer has recorded a decision; it does not necessarily mean the proposal was accepted. Store acceptance, revision, rejection or deferral separately from workflow status. This prevents a reassuring badge from concealing a negative or unresolved conclusion.

Identify the reviewer’s role and the evidence they inspected. An architectural historian reviewing style is not automatically verifying cadastral identity or legal protection. If several people contribute, separate their responsibilities. Record disagreements and the reason for any final editorial choice. Where anonymity is necessary, maintain an accountable internal identifier rather than falsely implying that a decision has no author.

Keep legal status in its own field

A research classification should never inherit official status because it resembles an entry in a designated-building dataset. Record the competent authority, official identifier, scope of protection, source link and verification date where such information is included. A designation may concern a component, a site or a wider ensemble. It may not correspond to the unit that the model classified.

The ICOMOS Venice Charter and the Nara Document on Authenticity are important reference points for conservation thinking, but neither is a substitute for the law governing a particular property. Likewise, inclusion in an international or thematic collection does not automatically answer a local planning question. Keep interpretive context and enforceable designation distinct in both data and public-facing text.

Recommended fields for a reusable classification record
Field group	Minimum content	Common omission
Identity and scope	Building ID, component, category definition	Treating a detail photograph as a whole building
Evidence	Source ID, role, date interpretation, rights	Keeping a thumbnail but losing the reference
Model contribution	Version, inputs, output and tested scope	Publishing a score without its context
Human decision	Reviewer, action, rationale and unresolved checks	Using “reviewed” as a synonym for accepted
Revision history	Previous state, change, author and reason	Overwriting a disputed label silently

Version changes instead of erasing them

A revised attribution is not a data-management failure. It may reflect a newly located drawing, a corrected address match or a better category definition. Preserve the previous claim and link it to the revision. State whether the change affects one building, a source relationship or an entire category. If a model is retrained, do not silently replace reviewed human decisions with new predictions.

Export formats should preserve relationships, not merely appearance. A CSV may be convenient for tabular fields, while JSON can represent linked claims and nested evidence. A readable report remains useful for reviewers. Where practical, use UTF-8, explicit field definitions and stable identifiers across formats. A visually polished PDF alone is difficult to compare, validate or update systematically.

Publish only what should be public

Evidence may contain personal information, restricted archive images or sensitive details about vulnerable sites. Separate public reporting from internal research material. Explain access restrictions and provide a reference when redistribution is not permitted. Accessibility also matters: describe images, expand unfamiliar abbreviations and do not encode review status through color alone.

W3C PROV offers a vocabulary for describing entities, activities and agents in provenance chains. CIDOC CRM can support richer cultural-heritage relationships. Adopt standards in proportion to the project’s needs, but keep their central discipline: make the origin of a claim inspectable. The best final record allows another person to understand the decision, challenge it and improve it without starting again.

Prepare a research question Check common reporting questions