A fashion object can carry several histories at once: the circumstances of its making, the lives of its wearers, its movement through dealers or families, and its later interpretation by a collecting institution. An image rarely reveals all of them. When a system groups garments by appearance, the record must prevent that convenient grouping from flattening these different histories into a single trend label. Begin by deciding which statements describe the physical object, which report documentary evidence, and which offer an interpretation.
Anchor identity before enrichment
Assign a persistent local identifier to the object, then distinguish it from identifiers for photographs, catalogue entries and digital files. A front photograph and a conservation image may depict the same garment without being interchangeable. Record the institution responsible for the object and retain its published accession number exactly as supplied. A future database migration should not require a researcher to reconstruct identity from a filename, a cropped thumbnail or a fashionable title that has since changed.
Standards help teams describe these relationships consistently. CIDOC CRM offers a conceptual framework for cultural heritage information, particularly events and relationships. Spectrum, maintained by Collections Trust, addresses collections management procedures. Neither automatically repairs a weak record. Their value lies in making the team specify what happened, to which object, under whose responsibility, and on what evidence. Choose a manageable profile rather than copying every possible field.
Treat claims as attributed assertions
“Made in Paris” should not occupy the same evidential position as “possibly made in Paris.” Preserve the source, the person or process making the claim, the date of entry and any qualification. If an image model proposes “silk,” store that as a proposed visual classification, not a verified material identification. Fibre analysis, a maker's invoice and a photograph offer different kinds of evidence. One should not silently overwrite another simply because the latest software returns a cleaner label.
Record gaps without inventing continuity
Provenance may include documented transfers, family accounts, sales, loans and periods for which no evidence survives. Mark an unknown interval explicitly. Do not infer uninterrupted ownership because two known owners share a surname, or turn an acquisition date into a production date. Sensitive histories may also require restricted access or community consultation. An AI-generated narrative can sound smooth precisely where the archive is fragmented; that fluency is a reason for closer checking, not a substitute for missing documentation.
Keep the source image accountable
For each digital image, retain a source reference, rights statement, capture information where available, and the relationship to the object. The International Image Interoperability Framework can support interoperable image presentation, while LIDO supports the exchange of descriptive museum information. Neither a IIIF endpoint nor a downloadable image is blanket permission for model training. Consult the institution's actual licence, contractual conditions and any restrictions associated with people or culturally sensitive material.
Design a useful minimum record
For a small pilot, a strong minimum is often better than an ambitious but empty schema. Include identity, object type, a qualified date or date range, material statements, measurements with units, relevant places, provenance notes, sources and image rights. Add a separate area for machine-derived descriptors. Make absent values meaningful: “not examined,” “not recorded” and “unknown after research” describe different situations. Collection teams can then prioritize work without mistaking a blank field for evidence that a feature does not exist.
- Documented
- A statement supported by an identified source.
- Proposed
- An interpretation awaiting appropriate review.
- Unresolved
- A question retained without a forced answer.
Make revision part of the archive
Every accepted machine suggestion should leave a review trail. Retain the model or workflow version, the input image reference, the original output, the reviewer's decision and the reason for a correction. Do not keep unnecessary personal information merely to make the log look complete. Periodic sampling can reveal whether the system repeatedly confuses repairs with original construction or modern mounts with historical silhouettes. A record designed for revision lets discovery improve while leaving earlier reasoning available for scrutiny.
Public catalogues at the Victoria and Albert Museum and The Metropolitan Museum of Art's Costume Institute are useful places to examine how object descriptions, attribution and collection histories meet. Read their individual records and rights information directly. Our discussion does not imply an institutional partnership or permission to reuse their holdings.