Capture context before cropping
Photograph the elevation, its surroundings and relevant details where lawful and appropriate. Note the capture date, viewing direction, obstructions and whether the image shows an exterior or an interior. A tightly cropped capital can support a feature observation but cannot reveal how that element relates to the structure. Keep the original image and document every crop used by a model.
Lighting, perspective and repair materials affect what a classifier sees. Scaffolding may hide a roof; shadows can resemble openings; replacement stone can resemble a construction boundary. Record these conditions as limitations. Avoid image enhancement that invents architectural detail. If an image is rectified or contrast-adjusted, preserve the unmodified file and explain the transformation so another reviewer can repeat or reject it.
Read dates as claims about events
A date on a drawing may refer to its production, approval, copying or later annotation. A building permit does not establish that the proposed work was completed. An alteration file may concern a shopfront rather than the whole facade. Store the date, the event it describes and the confidence in that interpretation as separate fields. An interval or qualified date is often more defensible than a single year.
Notre-Dame de Paris and Dresden’s Frauenkirche illustrate why visible appearance must be interpreted through histories of repair and reconstruction. Their circumstances are different, but both caution against equating what a camera records today with a single original construction moment. The same principle applies to ordinary buildings whose changes are less publicly documented.
Preserve text uncertainty
Optical character recognition can make archival documents searchable, but its output is not a transcription authority. Historic typefaces, handwritten additions, damaged paper and column layouts can introduce errors. Keep page references and image coordinates for quoted passages. Mark uncertain readings rather than silently replacing them with a plausible architectural term. Check names and dates against the original document whenever they support a classification decision.
Multilingual sources add another layer. A normalized English category may be useful for search, yet the original wording may carry a narrower or different historical meaning. Preserve both. Europeana, Gallica and the Deutsche Digitale Bibliothek can help locate material, but an aggregator record should lead back to the holding institution when possible. Record the item’s rights statement before downloading, sharing or using it for model training.
Combine modalities without hiding disagreement
A multimodal system can compare image features with document text or combine separate model outputs. The benefit is complementary evidence; the danger is that one input merely repeats the other’s assumptions. A caption containing the target style name can leak the answer into an evaluation. Keep descriptive text distinct from interpretive labels, and test whether performance survives removal of obvious label-bearing fields.
Do not merge contradictory sources into a smooth narrative too early. If a survey describes one phase and a photograph appears to show another, retain both claims with their dates and scope. The discrepancy may indicate an alteration, a mistaken identity or a document that was never implemented. It is a review task, not a reason to delete an inconvenient source.