ApiaryActiveLive
Try: pause · settings · learn · wipe
← Community / Reading Room
CT
craft · 24 min read

Cite the Corner, Not the Screenshot: A Durable Reference for Details in Digitized Images

An online exhibit includes a crop of a handwritten mark in the corner of a notebook. The mark looks like 18. An editor thinks it reads 13. Both people can see…

An online exhibit includes a crop of a handwritten mark in the corner of a notebook. The mark looks like 18. An editor thinks it reads 13. Both people can see the cropped image in a slide deck, but neither can find the whole page again. The object title appears on the final slide, without a page number. The viewer link lands on the notebook's cover. Another website has the same crop with better contrast and no credit at all.

The dispute isn't mostly about penmanship anymore. The team lost the path from the little rectangle to where it came from. A screenshot can convey what a researcher saw; it cannot reliably convey which page, which digital copy, which context, or which transformations got the researcher there. Those are separate facts, and each belongs in an evidence packet small enough that someone will actually keep it.

A durable image citation links at least three levels: the collection's public item record, the particular digital surface within that item, and the location on that surface. Add a fourth when you edit the pixels: the derivation from the surface to what the reader sees. No technical standards are required to start. A standard such as the International Image Interoperability Framework, or IIIF, can make the location easier to save when an institution exposes it. The method still works on a basic catalog with a numbered set of scans.

The test is simple. Give your citation to someone who didn't make the crop. Could they find the full surface, put a box around the same feature, and make a different interpretation if they needed to? If they would have to trust the person who named Screenshot Final, the citation needs another component.

Start with the record, not the download

When an institution has an item page, save that page even if the viewer has a full-resolution download. The item record can carry the title, identifier, institution, collection, date description, credit line and reuse information. A file URL might only deliver the current technical representation. If a web designer changes how thumbnails are made, your download URL may change without changing what the institution considers the object.

Write down what the identifier identifies. It might refer to one photograph, one folder of sheets, one volume, a bound pair of volumes, or a description with multiple digital surrogates. A number exposed next to the catalog entry does not automatically tell you which scan showed the mark. Two departments at the same institution can use different number systems. Keeping the holding institution alongside the identifier costs little and helps users recover when an old link fails.

If your discovery began through a search overlay, preview embed or a syndicated record, follow the citation back to the institution before saving the primary link. The syndicated page may be useful as a search path; the holding institution is better positioned to explain what was digitized and which statement applies to it. Don't substitute a title search for the exact item URL when the exact link is available. And if a collection uses a named persistent identifier, keep it as a separate field, not as characters you plan to extract from whatever the browser routed through last.

An item title can change as cataloging improves. Preserve the title you saw and the date you checked it. If a title carries obsolete or harmful terminology needed to find the item, do not silently adopt that terminology as your own caption. Keep the source's catalog term and the exhibit's description visibly distinct where the distinction matters.

Identify the actual surface

Cite a page or side rather than a vague portion of an object. Keep the human label (“facing leaf after numbered page 7”), the digital sequence (“image 12 of 44 in the inspected viewer”), and the image or surface identifier if available. The three can disagree without one being wrong. Front matter, foldouts and empty backs are sometimes scanned; printed page numbers can repeat or disappear. A digital sequence may also change after a rescanning project. The overlap gives the next researcher several paths back to the same place.

The IIIF Presentation API 3.0 specification provides a useful vocabulary even if you never write a viewer. A Manifest describes the presentation of an item and can organize several Canvases. A Canvas is a reference frame used to position content: for a notebook it can correspond to a page surface with a particular width and height. Image content is associated with that frame. Save a genuine Canvas identifier when the provider exposes one. Do not infer a Canvas identifier from a predictable-looking URL or assume a page identifier always resolves to a directly downloadable image.

The Library of Congress APIs page distinguishes its public interfaces, including image services using IIIF conventions. Not every object on its website necessarily offers the same API artifacts. Follow the institution's documentation and the actual links available for an item. Where the data model supplies a separate image service, a region request to that service and a reference to a region of the Canvas are different kinds of references, even when your specific item happens to make them line up.

For a non-IIIF case, a screenshot of the whole surface can be useful internally if retention and reuse allow it. Save the institution's source link, visual locator, original image dimensions if known, and when you checked them. If the full page cannot be downloaded or retained, write enough for someone to locate the feature: “lower inner margin, directly below the triangular puncture and next to the printed number 7.” Natural language is a retrieval method, not a last resort.

Move the rectangle back to source coordinates

Suppose you zoom a Canvas with a known width of 3,000 and a height of 4,000 reference units. You see the mark within a box whose upper-left corner lies 1,660 from the left and 2,270 from the top, with a width of 320 and height of 160, expressed in that same coordinate system. Store the source surface identifier and those four values together. For a IIIF Canvas spatial fragment, the Presentation standard describes a reference of the form: Canvas identifier, then #xywh=1660,2270,320,160. A prose description remains valuable because not all software will navigate or highlight the region when handed that reference.

Do not measure from a device screenshot and assume the numbers are Canvas coordinates. A browser may have scaled the image to fit, left padding at the sides, rotated it, or displayed only part of it. Device pixels and CSS pixels can also differ. The cleanest approach is to ask the viewer for a native region reference where offered, or measure against a displayed source representation whose relation to the Canvas can be verified. If correspondence is unknown, save your measured coordinates as belonging to a specifically identified derivative and keep the visual locator. Honest coordinates from the wrong frame, labeled as such, are less harmful than mislabeled “exact” Canvas values.

For a simple scaled case, write the conversion down. Imagine the full upright 3,000-by-4,000 image was resized to a 750-by-1,000 image, without cropping or padding. A box at (415, 567.5), 80 wide and 40 high, in displayed image-local coordinates corresponds to (1,660, 2,270), 320 wide and 160 high, in full-image coordinates. The factor is four because the two dimensions were scaled by the same factor. This arithmetic is valid only when that geometry is known. A view rotated by 90 degrees or an image painted into a subsection of a Canvas needs a different mapping. Don't make your newsroom debug a transformation that can be avoided by measuring on the right original frame.

If you resize or rotate the exhibit crop, write down the order. Cropping a region from a full surface and then rotating that region isn't the same as first rotating the entire surface and copying the original coordinate numbers. Human readers may not need this detail in a caption; the internal evidence packet should. Also keep a distinction between the literal region and the readable feature. A rotated image may make handwriting easier to follow, but the crop was taken from a specific physical location.

Worked example: two claims and one rectangle

The institution, object identifiers and specific images in this example are synthetic. It demonstrates a record structure and is not a claim about a real notebook or collection.

A town museum preparing an exhibit finds a notebook with a handwritten mark by a sketch of a bridge. Its provisional slide says “Inspection 18, work completed.” The notebook title at the catalog is “Bridge Committee Sketches, notebook B.” The digitized viewer includes a cover, blank endpaper and an image on which printed page number 7 is visible. On image 12 of 44, the bridge sketch has a mark in the inner margin. The online viewer exposes a manifest and a Canvas ID; for this example the location is expressed as xywh=1660,2270,320,160 on a 3,000-by-4,000 Canvas.

Write the evidence packet as separate facts:

FieldDemonstration value
Holding institutionDemonstration Town Museum
ItemBridge Committee Sketches, notebook B; example item identifier DEMO-NB-B
SurfaceViewer image 12 of 44 when inspected; printed page 7; keep the actual Canvas ID in a real packet
Source region1,660 left; 2,270 top; 320 wide; 160 high in Canvas units; lower inner margin
Exhibit derivativeCropped to the source region; rotated upright for display; brightness adjustment noted if applied
ObservationTwo adjacent handwritten shapes; the last may be a 3 or an 8
Inference“Inspection 18” is plausible but not established by the mark
Reuse checkStaff must check the item-specific terms for the intended exhibit; synthetic example does not assign a rights status

The original slide bundled together at least two claims. The number is one. Whether the notation is an inspection record, and whether that means work was completed, is another. Even if the number clearly says 18, the mark doesn't establish an event. The next editor should be able to see the ledger heading, any neighboring marks in the same hand and the figure caption. A public caption might read: “Sketch in the Bridge Committee notebook, with an ambiguous handwritten number at the lower margin.” If a separate dated inspection log later warrants a fuller caption, cite the log alongside the image rather than making the crop carry the new claim alone.

Now suppose a second catalog link offers a sharper image on which the last stroke looks like an 8. Before revising the caption, confirm it is the same leaf and digitization or identify it as a different representation. A shifted scratch, different handwriting elsewhere, or an opposite facing leaf can expose a mistaken match. Changes in contrast can reveal faint marks or make scratches seem continuous. Keep the earlier observation and document why the new view changes it. Do not replace the underlying source reference with the edited export that appears to support your preferred answer.

Keep citation and publication rights distinct

A source record can be public even when an image is not cleared for your specific reuse. A visible image may carry its own rights statement; metadata and media can have different statuses even on the same platform. The Smithsonian Open Access FAQ and its developer tools describe cases where metadata is available without an accompanying media file, as well as CC0 assets provided by the program. Its terms of use also distinguish copyright notices from other rights questions. They are useful examples of why you should record the source's statement for the particular content you want to use, not guess from the institution's brand.

RightsStatements.org's Copyright Undetermined statement is a statement about uncertainty, not a license. If an item carries an undetermined notice, preserve the statement accurately and check the planned use rather than recasting it as open permission. This guide is an editorial workflow, not legal advice; when a display requires a decision beyond your ability to verify, ask the institution or the responsible editor. A stable citation can still be supplied without republishing the image crop.

Public availability is not an instruction to harvest every frame, either. For a small exhibit, link the record and use only media you have a reason to use. Respect the institution's access terms and documented service constraints. If you don't need clearer images to support the claim, don't go looking for alternate delivery endpoints. A useful evidence packet is small, auditable and careful about who should see any retained materials.

Check whether the Canvas has the same geometry as the image

A Canvas is the coordinate space used to present a view. An image service describes the delivered representation of an image. They can have the same dimensions and an obvious correspondence, but their agreement is an implementation detail to check, not a property to infer from a pretty viewer. An image may paint only a rectangle within the Canvas. A collection may display more than one resolved representation of a leaf. Some viewers paint an open volume as a single two-page surface, then allow a reader to isolate the right page. The crop locator must name both the space in which the rectangle was recorded and the particular surface it identifies.

Read the painting annotation's target and content before swapping coordinates. The IIIF Presentation API distinguishes the content from the Canvas on which it is presented. When a viewer hides that information, refer to the described Canvas and say what you can actually locate. The IIIF Image API 3.0 specification describes image service information and the separate order of region, output size, rotation, quality and format in an image request. These are not instructions for changing the location of an item in the collection. A rotated derivative can make a handwritten checkmark clearer, but rotation is an output operation distinct from the original coordinates locating the crop.

A synthetic arithmetic case makes the distinction concrete. Suppose the full Canvas is 5,000 by 4,000 units. An unrotated full image, 2,000 by 3,000 service pixels, is painted one-to-one into the Canvas rectangle starting at x=2,400, y=100. The questioned detail is at canvas x=2,550, y=320 with width 300 and height 120. Under those explicitly stated assumptions, subtract the placement offset. In image service coordinates the region starts at x=150, y=220, still 300 by 120. Do not put the Canvas's x=2,550 into an image request; it would start beyond the bounds of that representation. Save both coordinate tuples and the target rectangle used to derive the relationship.

If the whole image instead fills a rectangle half as wide and half as tall as its native dimensions, subtraction is only the first step: a Canvas offset of 150 units inside the inset becomes 300 pixels on the original image. That scaling step should be applied to the width and height as well as the left and top. Round outwards for a pixel request where exact fractional bounds are not available and say how rounding changed the rectangle. If any of the assumptions fails, such as a rotation, a source crop inside the painted image or a non-obvious transformation, don't reuse this arithmetic. Test with at least two distinguishable corner features or retain the Canvas locator plus a human description until the mapping is documented.

Rendered output width is also a separate fact. Requesting a 600-pixel presentation of a 300-pixel feature does not create missing manuscript detail. An API may not support a chosen combination of output operations or sizes. Where the server declines an image request, don't modify the evidence locator merely to get an error-free thumbnail. The image service can fail while the evidence record remains a usable locator. Try the repository's documented viewer workflow, preserve the intended rectangle, and name the relationship to the full source that you were able to verify.

Give an image reading room for a genuine disagreement

Distinguish what a detail reviewer sees from what the article wants it to mean. A slash, black dot or short stroke may be ink, damage to a microfilm copy, a scratched glass plate or something introduced during processing. The full view and adjacent surfaces can sometimes help separate those possibilities. One other person's opinion after seeing your prediction is rarely an independent reading. For a consequential detail, invite two readers to write down the shapes, whether they seem joined, and what would change their reading before revealing the favored transcription. Give both the full surface and a clearly noted derivative so they can refer to the same thing.

Don't combine disagreement into a falsely tidy average. One reader can say the mark appears to close at the lower edge; the second can say that edge is clipping from the scan. The public conclusion can leave the number unread or report that two interpretations remain plausible. Keep alternate literal readings with the explanation that supports each. An editorial reviewer can then resolve a consequential issue by obtaining a better rendition or by finding evidence that a particular interpretation fits the notebook's system of notation. They do not have to reopen all transcriptions because one judgment was lost during layout.

Record transformations that have evidentiary effects. Normalizing a slight tilt can reposition the bounding box; increasing contrast can hide weak pen strokes; sharpening can introduce apparent closed shapes along an edge. Record what was done, not a vague stamp of approval that the derivative is “enhanced.” If the crop is used because it makes comparison easier, say that. If the source does not reveal the relevant distinction at available quality, say that too. The reasoning record should not depend on recovering the precise sliders of one manipulation program, but it should preserve whether the claim came from the source at normal view, a visibly different derivative, the presence of the same mark elsewhere, or a separate record.

For comparison pairs, make the direction explicit. Do the two crops refer to two different physical examples, two captures of the same source, or two different processed views from one capture? The first may support repeated use of a sign; the second may help rule out a scanner artifact under some conditions; the third does not create independent capture evidence. If the statement is that an inscription appears on two objects, two derivatives of one object are not enough. Store intermediate crops as aids rather than letting them multiply silently in a count of examples.

A second worked example: the missing half of a map foldout

This case is entirely synthetic. A conservation volunteer is trying to cite an embossed lettering fragment near the fold of a map of an old park. The map is part of a physical file with a narrow stitched flap. One electronic presentation shows the full map with the flap folded in; another digitization gives it a separate surface. A newsletter graphic attached to the draft combines the two. The editor initially thinks the composite is a scan of one unfolded map.

The collection page in the demonstration tells the editor the common physical item identifier. The first Canvas shows the printed body and the folded area at the right. It is useful for locating the crease in relation to the path and labeled pond. The second representation places the flap on a separate Canvas. A region there shows letters on the flap without an accessible view of the original alignment of that section against the map body. These are not two competing crop-coordinate systems for the same original photograph. They are two observations of different configurations. Assign each surface its own reference; do not carry a numeric rectangle from one to the other.

On the newsletter image, a designer inserted part of the flap's scan over the folded map and rotated it slightly. An overlap touches the end of a thin road line. That alignment is an editorial choice, not evidence that the road extended to the letters in the original unfolded map. The public illustration might still be useful if clearly labeled “composite of two digitized views,” with a seam and two controlling citations. But if the article's claim depends on whether the road actually ended by the lettering, the composite does not support the claim. A documented unfolded view or examination by staff could provide evidence the editor does not currently hold.

The editor saves three independent statements: the fold is visible next to the illustrated pond in the first view; three partial letters are visible separately on the flap in the second; orientation and contact of the flap with the road line are not established. This is more precise than “a map shows the path at the entrance,” and it gives a specialist a concrete request. A human with the item in front of them knows which flap and road line to inspect. They do not need to guess what the newsletter's pixel offsets mean.

Now imagine a fresh scan from the repository years later. It omits the entry cover, adds an unnumbered handling note and changes the index of the unfolded map. An editor recording only “image 16” would likely jump to the wrong position. The item identifier, description of the foldout and two visual landmarks still point to the needed surface. If the new Canvas identifier resolves to a different object or the old identifier disappears, keep the older citation in the record and add the new mapping with a date. Do not write a retrospective statement implying the older image showed what the newer evidence first made visible. This is the difference between preserving a claim and pretending the presentation never changed.

Make migrations additive and check the fallback

Treat a move from a portal viewer to a new exhibition platform as a small migration. Choose a handful of difficult examples to test: the first numbered page after two unnumbered surfaces, a detail at the join of a two-page spread, an upside-down image rotated for readability, and a region where the source record is accessible but the display is not approved. For each, ask someone outside the production team to reproduce the locator using the public or appropriate internal role. It isn't a successful handoff if the workflow works only with the editor's private session.

Keep the stable item information alongside a current resolvable URL. If you save the manifest or published metadata for internal reproducibility and the institution allows it, label the time of capture and reference used. It is not reasonable to promise to mirror all associated images. The main aim is preserving the minimum evidence needed to recover the surface: item title and ID, Canvas or page identity, plain-language region, exact numeric region when stable, and how the display changed it. A temporary asset URL may be excellent to locate renditions on the day of production and a poor URL to cite on a printed sign. An old cached printout is evidence of what the editor saw then, not an always-accurate declaration of current rights or availability.

A migration result can be one of several meaningful states. “Same locator tested in replacement viewer” is good where an exposed Canvas remains resolvable. “Item preserved; rectangle recomputed after a recapture” needs both tuples and the source observations used to confirm the mapping. “Old rendition unavailable; exact placement unrecoverable” is not a failure to be hidden by pointing at the nearest-looking leaf. Use the source image only for the parts of the claim still observable, or remove the tightly dependent illustration from the presentation if it can no longer be checked.

If the link opens but the crop no longer lines up, verify a visual guard first: one place immediately inside the left edge and one landmark near the bottom or right edge. This can catch accidentally viewing a resized thumbnail or a mirrored grip decades later even when coordinates look numerically reasonable. Store the plain-language guard near the rectangle, not buried in unrelated catalogue notes. A comparison that fails a guard is a reason to recompute from the documented source, not to nudge the rectangle toward the feature the author hoped would appear.

Prepare a public caption and a private audit note

A public caption can identify the item, surface and factual detail in a sentence understandable without technical vocabulary. Include orientation or synthetic composition when it matters. The evidence note can describe coordinates, access dates and quality in more detail. Never expect a visitor to decode an annotation target to understand a basic exhibit claim. Conversely, don't flatten a carefully grounded observation into the more attractive but unsupported “the signed drawing proves the bridge opened that year.” The public-facing copy should match the evidence note at the level of certainty, not at the level of engineering detail.

A simple audit note names the exact object, representation, surface, region axis order, specimen dimensions, rough visual landmarks, edits applied, evidence text and any reason a later editor might have to revisit the claim. Include the expected outcome of a recovery attempt: “left side of crease enters region; only three letters legible,” for example. The other reader can compare that expected outcome to what arrives, rather than taking an HTTP response as proof the original recommendation survived. When a high-status graphic in the exhibition is embellished with a better illustration later, log whether it still depicts the same evidence or has become a decorative substitute. Replace the caption if the evidentiary role changes.

For large projects, resist building full secondary image archives merely in order to save small citations. Collect a plain record for each consequential crop before preserving media. Preserve derivatives under the policies appropriate to those images and that use. If the archive no longer confirms a public asset, the conservative response is to keep the citation and pause reproduction for review rather than making a fresh hidden copy available. On the next editorial pass, the needed question is not “where did that JPEG go?” but “which record, view and claim did that JPEG support?” Keeping those separate is the preparation that makes a citation durable.

Use the boundaries of a region as a visibility statement

Coordinates can cut a contextual connection. A rectangle that contains a writer's initials and loses the line assigning them to a different diagram is technically correct as an image locator but misleading as an illustration. Save an evidence crop and a context range when the article combines information from distinct locations. If the wider range is too heavy for a display, cite the context separately and use a caption that points to the relationship. A human reviewer can then check whether the relationship is present on the full surface rather than assuming every apparent adjacency in a montage was present before editing.

On a manuscript opening, this may require several simple zones rather than one huge rectangle. Suppose the left page contains a legend, the right page contains a symbol and an adjacent inserted note uses a different pen. The reader cannot conclude the note is part of the legend merely because one screen shows all three. Give each evidence region its own locator plus a reason to inspect it, and distinguish physical adjacency from editorial assertion. If the catalogue is undecided about the inserted note's association, call it a note in the digitized object rather than attributing its placement to the manuscript's original author.

Archive scans can also exclude a relevant extent. A part physically cropped off, hidden by a binding or obscured by damage cannot be made visible by widening the coordinate request. Write this as a visibility limit, not a missing mark. For example, “no mark visible along the surviving edge in this scan” supports much less than “the original sheet never had a mark there.” If the edge matters, show the full image and identify which part is cut off. Keep this limitation near the caption so it does not vanish when a designer copies the crop onto another slide. Evidence is generally stronger when its known edges are visible or accounted for, not when an attractive cut looks artificially complete.

In written notes, a phrase such as “lettering partly visible, abrasion runs into the lower margin” can also serve an accessible role for a reader who cannot inspect a low-contrast crop. Describe what you can see without completing the uncertain reading. That one sentence does double work: it explains what the cited region contains and lets an independent editor judge whether the public copy overclaims it.

Test the handoff with a person who wasn't there

Ask a second person to recreate the crop. Give them the packet and five minutes, not your browser tab. They should be able to open the item page, locate the surface, identify the rectangle and compare at least one surrounding feature. If the page preview is behind an access prompt in their browser, the test has shown a retrieval problem before the exhibit goes live. If the stable identifier still finds the item but the old region link doesn't highlight, the durable parts have done their job: relink the current viewer, retain the old identifier in the notes, and check the rectangle.

Consider three failure drills. Replace the screenshot with a blank and see whether the citation stands alone. Open the citation on a phone and check that it is not just a pointer to an enormous unmarked volume. Read the caption without the image and ask whether it overstates its evidence. A citation can be mechanically healthy but editorially misleading. Solving only the link test is not enough when the central claim remains weak.

For a series, give each region a short local ID such as NOTEBOOK-B-07-R1. Use the ID in artwork notes, caption review and approval records. Let the evidence packet link outward to the item and inwards to your derived crop. If you revise the caption, keep the previous note and the reason; do not have two different files quietly circulating as equally official evidence. You can keep the data in a simple table or per-image text file. The size of the system matters less than whether someone other than the original author can reproduce it.

Keep the first independent successful reconstruction as a short note: who compared which presentation, on what date, and whether the location or only the broader page could be confirmed. A later failure can then be investigated as a change to the catalog, scan, tool or evidence packet instead of assumed to be a mistaken first reading.

Finally, decide what the reader needs. A public exhibit might show an ordinary source link, page label and concise disclosure of rotation or color adjustment, while staff keep the Canvas information and rights status separately. A digital scholarship appendix can expose the region link and both visual context and precise geometry. Both kinds of publication can be good. What cannot survive is a screenshot promoted into a claim with no way back to its source.

AI disclosure: Drafted with AI assistance for Apiary. The worked object and evidence packet are synthetic. The official specifications and institutional documentation describe actual interfaces or reuse models, not the synthetic catalog record.

Source notes

Frequently asked
What is Cite the Corner, Not the Screenshot: A Durable Reference for Details in Digitized Images about?
An online exhibit includes a crop of a handwritten mark in the corner of a notebook. The mark looks like 18. An editor thinks it reads 13. Both people can see…
What should you know about start with the record, not the download?
When an institution has an item page, save that page even if the viewer has a full-resolution download. The item record can carry the title, identifier, institution, collection, date description, credit line and reuse information. A file URL might only deliver the current technical representation. If a web designer…
What should you know about identify the actual surface?
Cite a page or side rather than a vague portion of an object. Keep the human label (“facing leaf after numbered page 7”), the digital sequence (“image 12 of 44 in the inspected viewer”), and the image or surface identifier if available. The three can disagree without one being wrong. Front matter, foldouts and empty…
What should you know about move the rectangle back to source coordinates?
Suppose you zoom a Canvas with a known width of 3,000 and a height of 4,000 reference units. You see the mark within a box whose upper-left corner lies 1,660 from the left and 2,270 from the top, with a width of 320 and height of 160, expressed in that same coordinate system. Store the source surface identifier and…
What should you know about worked example: two claims and one rectangle?
The institution, object identifiers and specific images in this example are synthetic. It demonstrates a record structure and is not a claim about a real notebook or collection.
References & sources
  1. Apiary Reading Room — Open, cited knowledge base — funded to keep bee & practical research free.
From the Apiary Reading Room. Opinion & editorial — not financial advice. We don't overclaim.
More from the Reading Room