Historical Data Methodology / Editorial draft
Beyond the scan: preserving context in archival AI
A proposed collection pilot method covering scope, digitization, structured records, AI-assisted extraction and expert review under agreed access conditions.
Prepared for editorial review · Proposed approach · No empirical results reported
Collection scope and record structure
A scan makes a page accessible as an image. Structured interpretation requires additional choices: what the document represents, how it is described and which uses the collection holder permits. Those choices belong at the start of a collaboration.
The proposed first step is a bounded collection sample. Together with the institution, define the languages, record types, image condition, available descriptions and intended research question. Record exclusions as carefully as inclusions.
Assess existing images before deciding whether further digitization is needed. Agree responsibility for scanning, image requirements and record identifiers with the collection holder. Document metadata and passage references provide the structure for tracing extracted data back to the source.
AI-assisted extraction and expert review
Before processing, agree who can access the material, what derived data may be shared and who will review readings. Sensitive or restricted records should be addressed through the institution’s own access requirements and an agreed scope.
Within the agreed scope, AI may assist transcription, interpretation and entity extraction. These outputs are proposed readings for assessment, not historical conclusions. The review sample should include difficult pages as well as legible ones. Record unresolved readings, omissions and corrections, together with the method's limitations.
Dataset handover and knowledge transfer
The proposed handover is a structured, searchable and source-linked historical dataset with reviewed transcriptions, candidate entity links, review decisions and unresolved questions. Export formats, method documentation and limitations should be agreed with the institution. Knowledge transfer should enable its researchers to inspect how each interpretation was produced.
These are proposed collaboration principles, not a claim that a pilot has been completed or that a named archive endorses Geyserich. Publication of any collection examples or results would require the permissions agreed for that collaboration.