[0:30] Thank you, and thank you to the organisers for having me. My name is Julien Raemy, and I am now at the Walter Benjamin Kolleg, in Digital Humanities, at the University of Bern. I wrote this abstract based on a project that Murielle Cornut and I carried out while we were both PhD candidates in Basel. Murielle is now a Project Collaborator at the Photothek of the Kunsthistorisches Institut in Florenz, Max-Planck-Institut, and the work is very much a collaboration between the two of us. My talk revolves around one question: how do we model an object whose pages exist in several material states at once?

[1:15] Our case comes from Cultural Anthropology Switzerland in Basel. In 1991 it acquired the collection of the Kreis family, a local family of physicians and printers, and a good example of urban bourgeois culture. That is around 20,000 photographs, and about a quarter of them are arranged in 93 albums. During PIA, Participatory Knowledge Practices in Analogue and Digital Image Archives, all 93 albums were restored, catalogued and digitised. The Swiss National Science Foundation funded the research; separate grants funded the digitisation. The one I will focus on is Album 31. You can see its blue textile cover, with two holes where a cord binding used to be. It opens with a family trip through northern Italy in 1927: Milan, Venice, Pisa, Florence. The later pages follow the youngest son, Walter, on his student trips to Grenoble, Arles, Lyon, Strasbourg and the Valais. Photo albums are notoriously "bulky" objects, as Stiegler and Yacavone put it, and Stiegler speaks of their "aesthetic, epistemic, and praxeological restlessness". They simply do not read in a straight line.

[1:30] In the archive, the album is catalogued with a classic "is part of" logic: the album, then its pages, then the photographs on them. The album gets one identifier, the cover is unit 01, the pages follow, and each of the 197 photographs is linked to its page and to the album. That is consistent, but a lot stays implicit. The photograph numbers do not follow the travel story, for instance. And the usual digital surrogate gives you one stable image per page. The problem is that a page is not one thing. Some photographs are loose: one on page 6, two on page 14, one on page 22. The conservators deliberately left them unattached, because nobody could say for sure where they originally belonged. On page 17 you can still see the corners of photographs that are gone. And every page has its parchment interleave, which you can leave down or lift. In numbers: 128 captures for 84 page views, for a single album. So which of these is "the" page? The screenshot on the right already hints at our answer. Mirador shows the whole spread here, so you see one layer for page 4 and the four states of page 5.

[1:30] Why did we have to do this at all? Most systems that receive the files after digitisation expect one image per page, so the information about which capture belongs to which state gets lost. IIIF lets us remodel this after the fact. With a good digital asset management system that already holds such structure, you could plug it in even more easily. So we use IIIF not as a display technology, but as a modelling layer. The album becomes a Manifest, each page a Canvas, and the scans are painted onto it through annotations. The key piece is the Choice: all the captures of a page become alternatives within one annotation, on one Canvas. No separate record per state, nothing thrown away. On the right is the actual Canvas for page 5, abridged: a Choice of four images. The default capture comes from Memobase through Image API 2.1, the others through 3.0. The Choice does not mind; that is interoperability at work. Context goes on top with the other two LOUD specifications. Web Annotations add knowledge to regions of a page. And since this week, Linked Art is in place too. IIIF says how the album is presented, Linked Art says what the album is. The record, linked at the bottom of the slide, describes the album as a physical object, classified as a photo album with the Getty AAT, produced between 1926 and 1928. The two point to each other: the Manifest links to the record with seeAlso, which you see at the top of the excerpt, and the record links to all three Manifests with subject_of. Two-way links across two different APIs: that is interoperability in practice. And we did not write this JSON by hand: we used iiif-prezi3, the Python library maintained by the IIIF community, so the output follows the specification.

[2:00] Let me show you what this looks like in Mirador. (Start the video.) Here we are on page 5. First I switch on the interleave, "mit Schutz" in German. Now the loose photograph, "mit Zusatz". And now both together. Four states, one page, one Canvas. Across the album there are three patterns. On the cover, you can switch between its current state and the remnants of the original cord binding. Most pages alternate between the bare page and the page with its interleave. Page 5 is the most complex case, with four alternatives. Simpler surfaces, like the pastedowns, just have a single image. In total, the layers Manifest has 86 Canvases, and 40 of them carry a Choice. Since yesterday the layers are back in the published Manifest. 124 of its 128 images display, so you can open it in any IIIF viewer yourself. If asked why this is a recording and not live: it is simply safer in an online talk. Our former image server is no longer publicly reachable, so the Manifest was rebuilt under the same URL. Memobase holds only the default capture of pages with photographs; the page variants and the pages without photographs are tiled from the files delivered by CAS and served from hostiiing. If asked about the 4 missing images: 029_mit_Schutz, 039, 055_mit_Schutz and the verso of the loose photograph 09861 have not been delivered yet, so they stay at their pending address and do not display. That is why the Manifest still carries an "Under construction" notice. If asked why 84 page views on slide 3 but 86 Canvases here: the two extra Canvases are the recto and verso of a loose photograph (09861), which the paged view leaves out. The verso is one of the 4 images not delivered yet.

[1:30] This brings us to a paradox. The digital album shows more than you could ever see in the reading room: states that never coexist when you hold the album. In that sense it is "more than real". And the Manifest does not pick a winner: the ambiguity is part of the structure. The price is fragmentation. Every state is a separate act of handling: the cover was captured at 3662 by 2730 pixels in one state and 3663 by 2729 in the other, because the album moved between shots. Hosting leaves traces too: Memobase holds only the default capture, at 2400 pixels, so on page 5 it now sits next to sharper variants of about 3,700 pixels from another host. Page 29, on the slide, is the extreme case: its interleave layer has a recorded width of 12583 pixels, but that file has not been delivered yet. The bulkiness Stiegler describes does not disappear; it becomes a multitude of images and choices. From a science and technology studies perspective, this is mediation, not substitution. In Latour's terms, the Manifest is a new actor in the album's network, alongside the archive, the scanner, the image server and the viewer. And as Valerie Johnson reminds us, researchers still seek the physical encounter with personal objects.

[1:45] We call this a "boutique" Manifest, a term borrowed from colleagues at the Getty, who made a custom Manifest for their Bayard Album. Care has a cost: it took conservators, archivists, anthropologists and technicians. Can it scale to the other 92 albums? Not by hand. So we built iiif-photoalbum, and the diagram shows how it splits the work in three. First, the human part: curators fill in a spreadsheet, and images that share a page become its layers. That is where judgement is needed. Second, the automated part, run on every change and built on iiif-prezi3. It finds each image, reads its size, checks that the layers match, builds three Manifests per album and, optionally, a Linked Art record: more structured data for machines, linked both ways to the Manifests. Then it validates them, with the IIIF validator and by checking that every image answers. What needs attention goes back to the curators as a report. Third, publication under stable URLs, open in any IIIF viewer. This is a proof of concept, and it works: two albums, 200 images, six Manifests and two Linked Art records. And it is generic: a spreadsheet and images already online, so other institutions can use it beyond CAS and PIA, for instance from a DAM export. My takeaway: digitisation is interpretive modelling, not mass reproduction. Thank you, I am happy to take your questions. If asked about the figures: 195 of the 200 images are served: 124 of 128 for Album 31, 71 of 72 for Album 50. If asked about the checks: a layer is flagged when its aspect ratio differs by more than two percent from the default capture, or its resolution by more than a factor of two. If asked who defines quality: technical validity is defined by the IIIF specification and checked by its validator. Whether a layer is meaningful is a curatorial decision, made by curators together with conservators and researchers. If asked about long-term preservation: the Manifest points to image services by URI, so every migration means updating identifiers. Our image server moved from SIPI to Cantaloupe but kept its host name so the URIs would keep working. It is now no longer public, so the tool rebuilt the Manifests on Memobase and hostiiing without changing their own identifiers, which is what the published URLs rely on. They keep an "Under construction" notice until the last images are delivered. If asked about users: archivists, family members or automated systems such as handwriting recognition or object detection can all add annotations without touching the resource, and competing identifications can coexist.