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

Verify a Contact Sheet After Reordering the Grid

The editor liked the frame with the empty bench. It was the second tile yesterday. Today, a different frame occupies that spot: the same bench, from almost…

The editor liked the frame with the empty bench. It was the second tile yesterday. Today, a different frame occupies that spot: the same bench, from almost the same angle, with a delivery cart just inside the doorway. The editor says, “Keep number two,” and two people hear different instructions. One remembers the empty bench; the other has today’s grid open. Both are looking at legitimate proofs. Neither can safely resolve the instruction from the number alone.

A contact sheet displays a selection of photographs together at reduced size. For an editorial team, it can also be the record that connects a conversation to particular images. Those jobs are compatible as long as the sheet does not pretend that position is identity. Positions help people find an item in a particular proof. Identities help them keep talking about that item after the proof changes. A filename may help an operator locate a copy, but a name can be reused and the same image may have several derivatives. The question to settle before release is not “Does this grid look like the old grid?” It is “Does this particular export contain the approved assets, in the approved order, with the approved presentation?”

The checks that answer that question are different, and combining them makes failures hide. A six-tile export can be the wrong six tiles. The correct six tiles can appear in the wrong sequence. The correct identifiers can be printed under the wrong crops. The order can be exactly right on screen and wrong in the PDF. Finally, every one of those details can be correct in the proof you opened while the file you intend to distribute belongs to a different revision. This guide builds a short evidence trail for each of those risks, using editorial examples rather than assumptions about particular software.

AI disclosure: This article was drafted with AI assistance. The image identifiers, proofs, comments, teams, and test results in the examples are fictional.

Decide what a tile actually represents

A working selection should distinguish the thing photographed, the particular photograph, and the image file that will be displayed. A photograph of a door taken at nine in the morning and another taken five seconds later might depict the same subject. They are still different photographs. A color-corrected derivative and a cropped detail might descend from a single capture. They may still warrant different editorial treatment even where the source identity is shared. The smallest useful identifier is the one that points uniquely to what the approval controls. If reviewers approve a particular crop, a code that identifies only the broad image family is not enough.

Suppose an archivist calls the source photograph Harbor 27. The proofs show a full-frame derivative and a detail of the bench. A worksheet can record Harbor 27 plus distinct rendition references. It need not create a second source photograph merely because the detail is used. Conversely, if two shutter releases are almost identical, it should not fold them into one record because their previews look alike. The distinction allows the asset history to remain truthful even when the editor only cares which composition appears.

A grid slot is a third kind of reference. It identifies a position in a particular layout revision: page two, row one, column three, for example. A printed number can deliberately be shorter, such as tile nine. That is fine if the proof also identifies the revision, but the number alone should never function as the primary key in a long-lived decision log. Imagine finding a note six months later saying the photographer approved tile nine. Without the proof revision, that claim has no unambiguous target. Even with the revision, it says nothing yet about whether the approval followed the source capture, the attached crop, or the entire composition.

Give each kind of identifier a job. The durable image reference follows the catalog subject through ordinary file renaming and layout edits. A rendition reference follows the exact approved visual treatment under local versioning rules. A slot gives human readers a quick way to navigate the proof. An export name makes a delivered file easy to find. These references can be connected, but none needs to masquerade as another. For a small team, a plain sheet with four separate columns can accomplish the same basic distinction as a larger system with relationships between several records.

The policy should also say what happens when something changes. Adding a border to a thumbnail can be a layout revision, while trimming visible content may require identifying a different crop. Re-encoding the same image bytes changes the stored file, but whether that creates a new editorial rendition depends on whether the displayed result remains equivalent under the specific check. Replacing a mistaken image with a different capture is not an innocent file refresh, even if the filename stays constant. Define those distinctions before anyone relies on an “approved” badge to make them on the spot.

Preserve the instruction before moving the pictures

A common sequence error begins before any export. Someone tells the operator to move the bench beside the window. That could mean put it immediately after the window in reading order, put it in the tile physically beside the window in the printed grid, or keep the two in the same two-image group while allowing a page break elsewhere. Each interpretation changes a different property. The smaller the team and the more often everyone points at a shared screen, the easier it is to miss that there is no durable statement of the intent.

Turn positional language into an explicit instruction linked to the proof the editor used. “In revision four, use the empty bench, Harbor 27 full, currently tile two. Place it immediately before Harbor 41 doorway in the reading sequence. The two may cross a page break.” That instruction supports an independent check. It names both assets, anchors the prior position to a specific proof, and separates reading order from physical adjacency. If the pair actually needs to appear together on one spread, the instruction should say so rather than letting the layout tool choose.

This matters most when work overlaps. If one operator removes a photograph while another moves it, their edits cannot both be expressed as uncomplicated slot changes. If the editor’s note was issued after seeing only a PDF, the source project might have been further ahead. Preserve the PDF reference and identify what the editor was deciding, not what the project happens to display today. When the old proof and its source mapping cannot be found, treat a reference such as “the third one” as unresolved. Showing the editor the two likely candidates is faster than guessing and silently locking the wrong one into the next proof.

Before making a change, write down what must remain unchanged. A reorder request is often also an instruction not to recrop anything. A change to one caption does not authorize removing another image. A grid-width adjustment may permit row breaks to move but not alter the reading sequence. Those invariants give reviewers something concrete to look for. Without them, a polished export can feel automatically successful because it looks like the kind of thing the editor asked for, even when it contains an unrequested visual or selection change.

Create a short manifest the team can reconcile

A manifest here means a simple record of what a particular proof is supposed to contain. It can be a worksheet or a lightweight report produced by the layout system. Give it a proof identifier and a revision, then list the intended reading sequence with one row per occurrence. Include the durable image ID, rendition reference, short visible caption or caption revision, and any exceptional layout constraint. Record renderer settings only to the extent that they matter for the handoff: page size, reading direction, number of columns, view name, or whether a crop inset must show its border. Avoid treating a formatting record as permission to release images.

An occurrence is not always an asset. If the same detail is intentionally shown twice, say once at normal size and once enlarged for comparison, give the two occurrences separate references while preserving their common asset. That allows a duplicate check to distinguish deliberate repetition from a mistake. For example, two rows with Harbor 27 full may both be valid if one is the introductory tile and the other is a clearly marked reprise. Two indistinguishable rows with identical occurrence keys are different: they make it impossible to say which was inspected or approved.

A minimum useful snapshot should be possible to read without the layout application. It answers which proof is being reviewed, which photos appear, which appearance was selected for each one, and where each occurrence should be found. In an export with variable tile sizes or group headings, include a logical reading sequence even when logical positions cannot be inferred from sort coordinates. A row filled with pictures and one tall portrait can have at least two visually plausible reading paths. The operator who designed the layout should not be the only reader able to identify the intended one.

For a synthetic exhibition called Harbor Workroom, the starting order was A, B, C, D, E, F. B was the empty bench, and C was the cart entering. The editor asked for the bench before the doorway photograph F, wanted C moved to the end, and approved leaving all crops unchanged. The revised instructions resolve to A, D, E, B, F, C if the rest of the relative order is preserved. The resulting manifest registers B and C by durable IDs and attaches their existing rendition IDs. It makes no change to either image record’s capture date. This is a layout sequence, not a new chronology of the shoot.

The proof still needs a physical binding to this record. In one workflow that may be a clearly printed revision in every page footer plus an exported machine-readable list. In another it may be a file fingerprint recorded with the proof and the manifest in the same release folder. A fingerprint can establish that two files are byte-for-byte equivalent if the comparison uses the same method, but it does not establish that their contents satisfy the manifest. The team should create that binding after determining which export was actually reviewed, not generate it from the workbook and assume the PDF agrees.

Check membership before explaining the order

Compare the selected occurrences first. With a unique occurrence key per intended tile, line up the expected and observed keys and note missing, extra, and repeated keys. The question is not whether two lists are equally long. In Harbor Workroom, suppose the wrong job yields A, D, E, B, F, B. Everything through position five checks out and the count is still six. The second B is an unexpected repeat and C is missing. A count-only test misses both; a sequence-only screenshot comparison may spot the duplicate if the reviewer remembers it, but close images readily escape that kind of informal review.

Sometimes a whole selection is wrong while its numerical labels are consistently unique. Imagine a proof for the west exhibit mistakenly exported from a tab named East Gallery. Each tab has six images and the same template. The labels run from one to six on both. There are no duplicates within either export. The manifest still catches the error because labels one through six are not the keys used for membership. The observed keys belong to the wrong selection, and that is a larger failure than a few slots being swapped. No amount of changing the reading order can repair it.

Check the occurrences you actually observe. If the renderer outputs a manifest automatically, ask whether that manifest describes what it intended to render or what was preserved in the resulting file. A render job's log can assert that B was scheduled to paint while the output shows C under B's caption because the file reference was stale. The scheduled keys are still useful evidence for finding the boundary at which the error emerged. They are not independent verification of the exported image.

When evidence is incomplete, label it precisely. A PDF whose text layer shows all six expected identifiers has passed a label-membership check. It has not necessarily passed visual membership or image-to-label mapping. A scan of the print without visible durable references may prove that there are six tiles and let the editor recognize familiar compositions, but it is weaker on the two nearly identical frames. An honest “indeterminate for two tiles until compared at full size” is more actionable than granting a whole-sheet pass based on the clearest four.

Verify order as its own property

Once membership is correct, check the sequence represented by the deliverable. For a simple left-to-right printed proof, compare the reading order row by row and then page by page. For a right-to-left print or a sheet with paired images or extra columns, first record the intended reading direction and any group structure. If a grid simultaneously exposes a visual order and an order for keyboard or assistive reading, decide which is part of the deliverable and check them separately. They can diverge even when every card is present.

Harbor Workroom can pass membership and still fail sequence. A swapped pair could produce A, D, E, F, B, C. This contains the same six occurrences, yet it reverses the editor’s intended bench-to-doorway comparison and places the emphasis on the doorway first. The operator should record the expected and observed positions rather than saying the overall proof is “misaligned.” A specific difference identifies a small, repairable scope and helps the team notice if the layout engine reintroduces it after the next proof.

A particularly useful test case puts the meaningful move across a page boundary. If putting B before F moves it from the bottom of page one to the beginning of page two, inspect both pages. A graphic editor's artboard, a multi-page PDF, and a printed booklet may use different variables for page numbering and ordering. Checking page two in isolation may find B at its correct new location without catching an obsolete copy still present at the end of page one. Checking only a thumbnail overview can also hide whether the sequence is left to right or intended to descend in columns. Use the same viewing mode the recipient will use, and note the convention in the review record.

An explicit reversed or rotated sequence deserves special attention. If an editor drags tiles in descending column order while a template enumerates them horizontally, the first and last positions might both look plausible but the middle comparison pairs change. A reviewer who knows the intended beginning and ending can still miss the wrong traversal. Record one informative middle adjacency in addition to the endpoints and inspect its physical placement. For a small proof, a full manual sequence comparison is practical. For a large proof, combine a structured sequence check with independent visual checks around meaningful moves, page boundaries, and positions that look ambiguous.

Compare the displayed pixels and the overlay together

A correct identifier is necessary but not sufficient. Pick a witness feature that separates close candidates under the approved crop. For Harbor 27 full, the empty doorway could be decisive. For Harbor 27 detail, a hinge mark might be diagnostic if the doorway is intentionally outside the frame. A proof reviewer should name the witness feature and compare the expected rendition with the actual displayed tile. The feature should identify the decision, not just an arbitrary obvious object in the photograph. If every frame shows the same bench, the bench is not helpful.

Do not mistake the evidence for a second representation of the same failure. If the screen view and the exported contact sheet get thumbnails from the same cached preview, confirming that those two versions match may only confirm that the cache is consistently wrong. Compare against the approved rendition or an independent resolved asset view. In a small workflow, open the approved asset as its own file outside the template and make sure it carries the expected source identity. For a bulk workflow, compare a reliable resolved asset identifier as well as a meaningful visual sample. The existence of a lookup table is not proof that the output used it.

Then inspect the overlay and its association. Picture A with A's number and B's caption could improperly describe the right photograph. Conversely, B’s number over A’s photograph can misdirect even when the text is accurate for B. To test association, pick a tile whose caption has a characteristic phrase and verify that it appears in the visible caption under the correct pixels, not just somewhere in search results. If text is also selectable or exposed separately, test whether it associates with the same tile there when that is within scope. A text layer can omit what the printed page shows, or preserve a hidden version that should not be reused as proof.

A helpful negative case substitutes a plausible wrong crop rather than a dramatically unrelated scene. Suppose the publisher approved a wide crop because the picture was meant to show both a counter and the passage beside it. A tight replacement highlights the same counter and may even look cleaner at thumbnail size. If the manifest tracks only the original photograph ID, a routine membership check passes. Inspecting the approved framing shows that the passage disappeared. That is a rendition failure, not a failure to select the right capture. Recording the problem at the right level tells the operator to restore the agreed crop rather than hunt for a completely different photograph.

For physical proofs, account for cropping introduced after the layout stage. A desktop view can show the full rectangle but the print driver may shrink, center, or clip the page to fit paper. Inspect the nearest-to-edge tile and a text line near the bottom of the page. A successful content check at full PDF zoom is not the same as a successful print check. If the job is delivered as a PDF only, record that you checked the PDF and do not imply the print process was also verified.

A second worked case: same photos, different promises

The fictional North Arcade team produces a sheet for a community exhibit. Its photographer offers two versions of a storefront image: a full view and a deliberate crop that excludes a temporary menu board. An editor wants both versions in the working proof for comparison, but only the crop in the go-forward selection. A reviewer writes “Storefront good” after the designer moves the full view onto the last page and brings the crop to the top. A producer assumes the approval applies to both because they share one source record. The issue is not that anyone duplicated a file incorrectly. The issue is that the decision lacks the target and level of precision necessary for the requested use.

In a reopened review, the editor refers to source Arcade 12, crop R3, and says it may be included in the go-forward selection. The full view remains in the internal comparison proof only. The new manifest therefore has two clearly labelled deliverables. Its comparison sheet includes two occurrences of Arcade 12 with different rendition references; its go-forward proof contains one occurrence with R3. The IDs printed beneath the comparison photos can share the source prefix without implying that the two crops are interchangeable. Looking up the source record still reveals that both are derived from the same photograph.

The export error in this case is different from Harbor Workroom. The membership, order, and caption all match the go-forward manifest. But the renderer resolved R3 using an old thumbnail that was created before the crop was changed. The menu board, which should be excluded, is still visible in the pixels. That finding requires inspecting the image, not comparing the expected list against the proof labels. A new export selects the correct approved derivative and a reviewer checks it against the designated source. They note the crop witness explicitly: the right-hand menu board is outside the frame; the main storefront inscription remains visible. The reviewer does not need to reproduce or guess the board text.

One decision still needs to be kept separate from the proof: whether either rendition may be used for a particular publishing channel. The contact sheet confirms which rendition was approved for an editorial selection. It does not by itself establish a rights decision, permission, a release policy, or approval for a later change to the image. If the broader workflow has a clearance record, link to that decision rather than producing a second version of its contents inside the sheet. The link should not imply it was checked unless the reviewer actually checked it.

Approve the artifact you intend to hand off

At the point of review, name what the approval covers. A complete proof approval concerns a particular export and may include layout, labels, membership, rendition and order. An image selection approval concerns a durable source or a specific rendition and may remain meaningful even if the tile moves. A narrow correction approval might cover one updated caption and verify nothing about the rest of the proof. Do not expand a narrow approval simply because a general approval field is easier to display on the next version.

For an export approval, record a simple receipt: reviewer role, proof revision, exact export identity, scope reviewed, exceptions, and the result. A date and version identifier are usually more useful than a bare stamp. Where teams use file fingerprints, record one for the exact file inspected. Avoid naming a proof “final” in lieu of recording what it represents. When the image record, manifest, or final file changes, a clear receipt makes it possible to compare before trusting the previous approval.

There are legitimate narrow carry-forwards. If the editor approved a photo as a source selection without specifying placement, the old selection note can still apply after the same source moves. If the reviewer approved that rendition only on the first page beside a reference image, moving it may invalidate the contextual comparison even when the rendition is unchanged. An approval for an exported file cannot be transferred to a different byte sequence just because its filename is unchanged. If the purpose does not call for treating re-encoded copies as materially different, define an equivalence test and still record which actual copy was inspected.

The easiest approval trap is copying an old receipt into a new draft. In that draft, the receipt looks neat and complete, but it describes the wrong file. Keep approval records immutable as historical evidence; add a new entry that describes the change and whether approval was checked again instead of silently replacing the old fields. A new review should still be quick when the change is narrow. The point is not to re-run a whole week’s work for a single moved tile. It is to preserve enough evidence to justify a limited recheck rather than assume the old approval magically covers it.

Reconcile the delivered file without rerunning the whole project

Open the proof from the location or package the recipient will use. In an ordinary editorial workflow there may be several plausible artifacts: one opened by the layout application, one copied to a review folder, and an older copy already in a message draft. All three can have the same friendly filename. If the destination is a shared folder the team controls, verify the folder copy after replacement, then record its revision and file identity. If distribution is not part of the review, stop at the designated candidate and state that the recipient copy has not yet been verified. Sending is not necessary to test a local artifact.

A practical test follows one sample occurrence through all layers. Start with a manifest row that crossed a page boundary. Resolve the approved asset and rendition. Read the visible tile reference in the candidate, inspect the separating visual detail, check its immediate neighbor, and verify the proof revision in a footer or handoff record. The useful question is not whether each of those facts exists somewhere in the project, but whether they can be traced together in the reviewed export. For a small six-tile sheet it is often faster to repeat this for every occurrence than to design a statistical sampling policy. For a very large sheet, compare the entire exported sequence where possible and concentrate direct inspection on risky transitions and near matches.

If the received file cannot be checked independently, say what was checked instead. An application preview can establish what the current screen presented in that environment. A file list can establish which filename was selected, not its complete pixels. A fingerprint match can establish correspondence to an inspected copy even if the operator does not manually reopen a second identical copy. A page screenshot can prove only what was visible in the captured area. None should be described as a universal substitute for another; choose evidence strong enough for the specific question and keep its limits with the receipt.

Consider how different outputs represent the same revision. An eight-image two-page PDF and a single scrolling mobile image may both contain the approved sequence but differ in line wraps and crop readability. If the mobile export omits the printed tile numbers by design, require a different reliable mapping, such as visually legible compact references or a separately bound mapping that reviewers can access. Do not treat a pass on the PDF as a pass on the mobile derivative where meaning depends on details below the mobile tile resolution. Color, compression, and small labels can all change the practical interpretability of a proof without changing the selection list.

Make a deliberate failure prove the review is doing real work

A review rehearsal can use a disposable fictional duplicate while preserving the real accepted file. Start with a correct six-tile case whose records everyone can resolve. Make one intentionally wrong change to a separate export. For example, preserve all IDs and positions but put the old caption under the correct door photograph. Ask the reviewer to record which check fails and which checks still pass. If the response is simply “the preview looks finished,” improve the check before depending on it. A useful method localizes the fault instead of giving a blanket red mark without an explanation.

The next rehearsal can change only the underlying image file, leaving the label and position alone. This tests whether reviewers inspect an independent rendition or merely read overlaid text. Use two crops close enough that a genuine witness matters but not so close that verification depends on guesswork. In a third rehearsal, export the entire correct sheet with a stale revision footer. The expected resolution may be to regenerate a clean artifact and rebind the approval rather than repaint any tile. Those failures exercise different boundaries even though each might be described casually as the wrong proof.

Plan the independent evidence before correction. If a worksheet spells out the expected sequence and a human draws the export from an unrelated old file, the worksheet is useful as the expected side of the comparison. If a program derives both the expected and actual lists from the same incorrect source data, agreement does not prove independence. One low-cost separation is for the editor to save the intended changes as an explicit ordered selection and the reviewer to derive the observed sequence from the actual export. The two need not employ different brands of software; they need to represent different sides of the question.

Avoid introducing meaningful fictional failures into a shared proof by accident. Use unmistakable internal rehearsal names, prevent the faulty copies from entering the delivery folder, and preserve the sample instructions where the next reviewer can repeat them. The desired result is a stronger review process, not a stockpile of increasingly convincing false finals. A short record noting the seeded problem, what the reviewer saw, and how the test was closed is usually more useful than keeping all the temporary assets indefinitely.

What the standards support, and what remains an editorial choice

The FADGI still image technical guidelines describe institutional digitization best practices and connect image conformance work to appropriate targets, software, monitoring, and evaluation. Their useful lesson here is to separate measured or inspected image qualities from assumptions made because a file appears usable on screen. This proof workflow does not claim to determine whether an image meets a FADGI tier. A contact sheet thumbnail cannot replace the tests specified for a digitization program.

The IPTC Photo Metadata Standard distinguishes properties including a Digital Image GUID and supplier-related identifiers from descriptive properties such as a caption. That distinction helps explain why a system can preserve a reference while changing words or placement. It does not mean every incoming file contains an embedded identifier, that identifier values can be trusted without checking how they were assigned, or that file metadata automatically survives an export. The local manifest remains a useful place to preserve the verified relationship, especially when some assets arrive with sparse metadata.

The IIIF Presentation API explicitly represents ordered presentation structures. Its alternative page sequences recipe illustrates how an alternative order can reference the same canvases without creating a new underlying object every time its position changes. That is a helpful analogy for distinguishing image identity and editorial order. It does not prescribe how a contact sheet must label its tiles or settle which order an editor intended. If a workflow actually uses IIIF, consult the relevant structure and viewer behavior; do not assume an exported PDF inherits those behaviors automatically.

A derivative can also preserve the right number of occurrences while swapping the role of two versions. In the North Arcade comparison, the full view and the cropped detail may appear beside each other with a shared source prefix. If captions refer to them as left and right, a switch can make a technically accurate caption misleading after export. Use an explicit role such as full view or detail and name the comparison in the manifest. A second reviewer should be able to read the two roles without relying on a memory of who first arranged them. Inspecting that pair verifies the comparison as presented, not just the presence of two legitimate files.

The smallest defensible handoff

The handoff does not need to become a database migration. Start with the exact export and a short statement of which old proof gave rise to the instructions. Attach or link the ordered occurrence record. Note any important differences in layout, crop, or caption. Check the actual file against its expectations and name the remaining uncertainty if any of those checks could not be done. Then preserve the exact reference to the file that passed, so someone later can distinguish it from a near-identical candidate.

If an operator has only thirty seconds to back out of a risky exchange, stopping the send is better than making the best guess about tile two. The team can resolve the reference using the editor's old proof or two clearly named alternatives. If membership is correct but the order is still under review, label the artifact accordingly. If revision, membership, order, and presentation all pass, say so for that artifact rather than for every future derivative. This keeps the claim narrow enough to be reliable and useful enough that the next person can act without repeating the whole conversation.

Frequently asked
What is Verify a Contact Sheet After Reordering the Grid about?
The editor liked the frame with the empty bench. It was the second tile yesterday. Today, a different frame occupies that spot: the same bench, from almost…
What should you know about decide what a tile actually represents?
A working selection should distinguish the thing photographed, the particular photograph, and the image file that will be displayed. A photograph of a door taken at nine in the morning and another taken five seconds later might depict the same subject. They are still different photographs. A color-corrected…
What should you know about preserve the instruction before moving the pictures?
A common sequence error begins before any export. Someone tells the operator to move the bench beside the window. That could mean put it immediately after the window in reading order, put it in the tile physically beside the window in the printed grid, or keep the two in the same two-image group while allowing a page…
What should you know about create a short manifest the team can reconcile?
A manifest here means a simple record of what a particular proof is supposed to contain. It can be a worksheet or a lightweight report produced by the layout system. Give it a proof identifier and a revision, then list the intended reading sequence with one row per occurrence. Include the durable image ID, rendition…
What should you know about check membership before explaining the order?
Compare the selected occurrences first. With a unique occurrence key per intended tile, line up the expected and observed keys and note missing, extra, and repeated keys. The question is not whether two lists are equally long. In Harbor Workroom, suppose the wrong job yields A, D, E, B, F, B. Everything through…
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