Provenance helps collection holders make visual collections easier to find and use while source records stay authoritative.
Start with one collection, one representative sample, and one defined problem.
Commercial street, tram rails, wet paving, c. 1930
In a Provenance workflow, a model can propose a caption, labels, and a confidence score from each source record. A curator can review those fields before they are reused. Approved records can then support semantic search and related-image browsing.
Schematic. Record identifiers, captions, and cluster counts are illustrative. They are not live data.MTL ArchivesA collection people can search.
A public search and browsing system built from Montréal’s open photographic datasets while keeping the source record intact.
in the production collection
June 2026 audit
Keep the source intact.
Generated fields can support search and research without replacing the source record. Each layer stays separate, reviewable, and traceable.
- 01
// Source
Original files, identifiers, and metadata remain intact.
- 02
// Enrich
OCR, captions, and search details are added beside the source record.
- 03
// Review
People can inspect, correct, accept, or reject generated information.
- 04
// Use
Approved information can support search, research, interfaces, and exports.
Start with a bounded pilot.
Before a larger implementation, we work with a representative slice of the collection and one clearly defined problem.
Led by Wielfried Zouantcha, creator of MTL Archives.
- 01
One collection and one defined use case
- 02
A representative sample of up to 2,000 records, subject to data review
- 03
Two weeks after data access and scope agreement
- 04
A documented assessment and a working prototype or structured output