AI disclosure: Drafted with AI assistance for Apiary. Every speaker, business, amount, and media asset in the example is fictional. This is an editorial workflow example, not a compliance policy or claim about a real source.
A nonprofit posts a public video explaining how to join its tool library. The presenter says, “Borrowing is free on Wednesday.” The transcript says “Thursday.” An editor watches the recording, fixes the text, and calls it done. Two days later a member finds the video through an embed. That player still has the older caption file and shows Thursday. The main site now has the correct transcript, but there are two answers attached to apparently the same video.
Then a volunteer reports a harder problem: the presenter actually got an unrelated detail wrong in another section. The text editor silently replaces that spoken sentence with the correct policy. People who read and people who listen now get different information again, this time by design. One edit repaired a representation error. The other hid a source error. They deserve different handling.
The method in this guide is to tie corrections to exact assets and label the kind of change before choosing a remedy. That makes it possible to preserve an honest record without turning every typo into a public notice or every correction into a video re-edit.
Decide whether the words, the source, or the cut is wrong
A transcript error occurs when the text misrepresents the source. In the library example, the audio clearly says Wednesday, and the editor verifies that against the recording. Correct the equivalent sentence in each published text representation, including the caption track if affected. The video does not need replacement just to correct a copy error. If the mistake materially affected what the public could do, a concise note on the page may nevertheless be appropriate.
A spoken error occurs when the text accurately reproduces the recording but the statement itself is wrong. Suppose the video says returns are accepted until 9 p.m., but the published schedule closes the return window at 8 p.m. The accurate transcript should not silently pretend the presenter said eight. Keep the quoted speech true to the video and attach a correction where readers see the wrong statement. On the player page, place that correction near the relevant content. If the incorrect point is essential, dangerous, or likely to drive actions before viewers see an editorial note, reconsider continued distribution of that version and prepare a corrected cut or replacement.
A source change occurs when the publisher trims, replaces, or otherwise changes the media referenced by downstream text. Web caption cues and transcript time links belong to the exact edit. Even a small cut near the opening can make every later link point to the wrong moment. A title and thumbnail staying the same aren't good reasons to retain old timestamps.
Finally, a later fact change can make an accurate historical statement stale without making it an original mistake. If the library really closed at 9 p.m. when the film was made and the hours moved to eight this year, add an as-of date or current policy notice. Don't treat a dated statement as a transcription error, and don't tell the public it was false all along. A quote can remain a true record of a past claim while a page separately tells people what to do now.
| Type | What actually went wrong | Likely affected assets |
|---|---|---|
| Transcription | Text differs from source | Transcript, captions, extracted quotation, downstream copies |
| Spoken factual error | Source is inaccurate | Correction note, transcript context, maybe revised media and captions |
| Recut | Source timing or content changed | Cue track, transcript timestamps, chapter times, shares or embeds |
| Changed policy | Source may be historically correct | Context note, date, current policy links, possibly publication status |
These categories aren't an excuse to keep harmful old copies online. They are a diagnostic step that keeps the chosen action truthful.
Use two questions to classify reports with more than one problem
Ask what the audience could reasonably understand from the rendition they used, then ask what the publication knows to be accurate. A report like “the workshop says registration is fifty dollars” could mean the user heard a different amount, read it in the captions, or saw it on a briefly displayed slide. The figure could be a draft correction to a program that was recently repriced. A report is a pointer to investigate, not an invitation to choose whichever price resolves the email fastest.
In a small newsroom, write the received wording and route separately from your conclusion. Then listen to the appropriate part of the source, read the relevant text versions, and check the organization's authorized detail for the period in question. If the recording mentions one amount, the slide another, and the current schedule yet another, do not force the case into a text-only correction. An authentic transcript should preserve the speech; a descriptive version should also accurately describe the slide where relevant. The publisher may need a plainly labeled clarification, a new release, or another scoped decision. Record the conflict at the smallest useful level before deciding how to address it across different experiences.
Separate uncertainty about what was said from uncertainty about truth. A person can hear clear speech that makes an unverified claim. A second person can confirm the policy from a current schedule and still be unable to tell exactly which term the speaker used under a dropout. Neither piece of evidence resolves the other. Put a second verification row beside the recording review. “Speech confirmed; policy basis unresolved” and “policy confirmed; phrase partly inaudible” suggest different next actions. Avoid a dashboard field that compresses both to “verify complete,” even if a single editor usually closes both.
The most useful severity question is what a user might do differently because of the problem. An incorrect line about where to return a tool could cause an unnecessary trip; a mistaken period in the transcript may not change an action. The editor should weigh audience and placement as well as the raw number of changed characters. A one-word correction can warrant more visible treatment than a long paragraph altered for readability. Keep the rationale in the record without framing every minor maintenance action as a public incident.
Name the assets separately
Represent the clip and its related assets by stable identifiers and explicit revisions. A guide can be a single editorial item while its video, transcript, caption track, poster, and page have different update histories. A workable minimal mapping could look like this:
content: library-membership-guide
video: membership-r3.mp4
english caption track: membership-r3.en.vtt, text revision 5
descriptive transcript: membership transcript, revision 7
published page: /guides/library-membership
partner player: embed mounted by the neighboring shop
The video asset is a distinct reference for timing. Caption text can be corrected without shuffling the video revision. A transcript can incorporate new explanatory context even if it represents unchanged speech. The ability to say “media r3 with caption text 5 and transcript 7” makes it easy to check a combination. Version filenames or a manifest are one implementation; choose the smallest one that keeps unrelated revisions from drifting.
W3C's media transcript guidance distinguishes a transcript that conveys audio information from a descriptive transcript that also represents visual information needed to understand the media. A corrected transcript may therefore contain a spoken quote, a note about what's visible on screen, and an editorial correction. Make those roles visibly distinct. A plain label like “Editor's note” is more useful than putting revised information inside the original quotation. WebVTT covers timed text tracks, which can require related but separate corrections.
Keep the original recorded material available to authorized editors when policy permits, not automatically public. The W3C PROV overview explains the general entities, activities, and agents model for provenance; a small publication doesn't need to implement that full model to benefit from knowing which artifact derived from which and who made a revision. A timestamp and brief reason for the change may be enough. More formal retention rules depend on the organization and the content.
Track the dependency that made each copy separately vulnerable
The publishing item is a parent, not the only source of truth for every route. A player can use timestamps aligned to its own playback master. A news card can use manually copied excerpt text. An audio-only release can have a spoken segment similar to the video with different cue timing. A social graphic may quote a phrase without carrying any machine-readable version pointer. Finding the clip by title identifies a topic, not its downstream publishing graph.
Start from how the reported route is assembled. If a public page renders captions from a named asset, record that relationship and inspect the actual file and enabled track in play. If the snippet is copied into a separate editor field, mark the dependency as manual. An unchanged video may keep the same identifier while captions, download and snippets receive narrowly scoped revisions. The smallest useful inventory lists where the disputed information can appear, who controls it, and whether it references a shared file or stores a copy. It makes an editor's “done on the page” a readable statement rather than a claim about every possible surface.
Shared assets deserve a fan-out check before replacing bytes. An approved caption file might serve a first-party article and a mirrored copy of the same video. A second player might reference the file but line up to a different cut because it accidentally kept an outdated association. The same text can remain appropriate in both contexts while the time cues only fit one. Decide whether to introduce a corrected immutable reference for the appropriate media instead of changing a shared resource with unclear dependencies. The aim is not to standardize on a particular CDN strategy. It is to know whether the edited file corresponds to the video that a particular screen actually presents.
Give persistent navigation a separate pointer when possible. A stable topic link can lead users to the right explanatory section through several related versions even if the current exact recording is different. In the archive, document which captions belonged to an older video. In the live guide, use the visible default actually checked by the editor. A topic identifier stays stable to support readers; a revision identifier changes to preserve provenance. If you conflate them, a user may reach a valid guide but be unable to tell which quote or cues they received.
Work a correction from report to evidence
The fictional library receives a note from a member: at 00:47 the captions say Thursday but the audio says Wednesday. The reviewer opens the exact published video revision for that channel. They watch the whole sentence and a few seconds before and after, listening for whether the video includes a qualifier such as “this week only.” They then check the authoritative membership page before deciding whether the word is merely mis-transcribed or the whole message requires correction.
In this case the recording clearly says Wednesday and the membership site confirms Wednesday. Finding the erroneous word in only one paragraph does not complete the inventory. Search the actual caption track, the descriptive transcript, the excerpt in the news listing, and the manually added pull quote if there is one. Watch the partner embed or obtain confirmation of the different asset version if that route is outside your direct control. A global text replacement would be careless if the same video elsewhere mentions a real Thursday event.
Apply the correction specifically. Bump the text revisions, leave the video binary alone, and preserve approved descriptions of the speaker’s visible demonstration if they are unaffected. The primary page and first-party embed should fetch the corrected caption track. If a hosted third-party copy can't be updated with the same process, its status remains unknown or pending even after your own page is corrected. Don't mark it verified based on a promise to update later.
For the second scenario, the presenter definitely says 9 p.m. and the policy in force at recording time was 8. The editor retains “9 p.m.” in accurate quotations with a correction next to the passage: “Correction: despite the 9 p.m. time stated in the recording, the return window closes at 8 p.m.” A public explainer with a prominent time claim may warrant replacing the misleading cut entirely. If a replacement is approved, audit all related assets against the new video and note the revision; don't just switch the MP4 out under the old timestamps.
Another case: the audio is right and the graphic is wrong
Suppose an independent makers' club records a short video about reserving a classroom. The presenter says the first reservation uses form C. As they speak, a filmed monitor clearly shows form D. The public text transcript reproduces C correctly. It contains no descriptive note about the onscreen detail. A director initially proposes replacing C with D in the transcript to make the proof image easier to understand. They are trying to fix a continuity problem in the recording by changing the only rendition that was faithful to the speech.
Instead, verify the approved form from the club's policy. Suppose it is indeed C. Classify the conflicting graphic separately: one audiovisual asset gives two viewers different signals, even though the captions match the speech. The editor can preserve the speech, describe the conflicting graphic if needed to convey the state of the published clip, and place a clear direction about the correct form near the relevant passage and player. If the confusion is consequential or persistent, prepare an authorized corrected visual edit and determine whether captions need only a new attachment or resynchronization. Publishing a replacement is its own action, not an accidental effect of changing a text field.
An interim guide update can say, “Use form C. The recording accurately says C but briefly shows D on screen. The shown form is not the one to submit.” This note doesn't introduce a silent contradiction for a user relying primarily on transcript, audio, or graphics. It also preserves the factual basis that justified the correction. If the issue affects a downloadable instruction sheet, add that surface to the ticket and state whether it displays a proper form link of its own. A snapshot shared by an attendee is not under the same publishing control merely because the organization made the original recording.
If the captured graphic contains personal information that can't remain public under the publisher's policy, ordinary correction and archival preferences may not determine the appropriate handling. Route the concern to the designated owner and avoid copying the personal information into the ledger. It can identify the range and issue type without repeating the removed data. This doesn't make the correction unverifiable: authorized review evidence may still be kept and assessed under the relevant rules. The public explanation need only describe what readers can safely know and the action actually taken.
Preserve links when timestamps move
A transcript with jump links to important moments is convenient. It is also a dependency. If a six-second opening is removed, an old link for 00:47 now lands at 00:47 in the new cut, where someone could be answering another question entirely. Simple constant offsets only work if the rest of the timeline is unchanged. If there are cuts later, use an explicit segment map or recalculate relevant anchors by reviewing the new master.
A small publisher could use stable named anchors for topics in the transcript, then map those anchors to timecodes in each video revision. “Return deadline” is easier to preserve across media revisions than a raw seconds parameter. This does not guarantee an external player will understand your custom name. The page that owns the mapping can provide an explanatory fallback when somebody follows an old raw time link and no reliable equivalence exists.
Test a link before and after each edited segment. Test the boundary nearest the corrected statement, long enough to hear or read the surrounding context. If the corrected video omits a statement, don't pretend an old cue moved to a new time. Record that passage as removed, and give a stable text reference to the intended updated information where appropriate.
The same concern applies to language editions. An English transcript with time links and a translated caption file may be made by different people. In a rush, the English material gets corrected but the secondary version is still wrong. Decide whether the translation conveys the exact source statement, an explained correction, or an approved revision. Then check each representation. Don't silently replace a translator's rendered context just because a timecode in a different file moved.
Map retained ranges, replacement passages, and removals independently
Imagine an edited two-minute original. The new master removes the interval from zero through six seconds, then removes twelve seconds starting at old time 48 and inserts a new eight-second explanation. On a retained frame from old time 35, the corresponding new time is 29. A retained frame from old time 70 maps to 60 after the cumulative net ten seconds removed. An old link to 52 seconds is inside the replaced segment. There is no exact equivalent frame at a new time if the new material is different. Direct that link to an appropriate named topic or explain why the original passage was replaced, rather than declaring “52 becomes 42” across the whole recording.
Write the cutover convention down before testing. If segments are expressed as start-inclusive, end-exclusive intervals, an old timestamp of exactly 48 is part of the replaced block and exactly 60 starts the next retained block in the example. The behavior a visitor sees also depends on player behavior and media execution, so these intervals are a planning map rather than evidence of frame-exact seek performance. Test near the boundaries as a reader would. A chapter jump can land in a relevant place and still fail to give essential lead-in context if it puts a new audience right at an unexplained acronym or partial instruction.
Caption spans can cross a cut boundary. Moving the whole cue by an offset may leave removed speech visible against the new sequence or hide the opening of a retained sentence. Review the source immediately before and after the change, then repair captions according to what people now hear and need, not according to an earlier master’s parsing. New narration can introduce sound cues or other information requiring a fresh look. Carry the same review to the descriptive transcript if relevant visual context changed. Nothing about the production arithmetic certifies that a courtesy trim preserves the same information.
A modest segment map can live as a few readable lines with start, stop, status, new reference, and reviewer. Persist the references for reasonably expected users, especially when a guide is cited by partner instructions. If there is no official mechanism for redirecting external time links, say so at handoff. It is better to keep one stable known route to the relevant answer than to produce fictional exact mappings for every externally distributed link.
Build a receipt and keep it smaller than the work
The correction ledger doesn't need the entire transcript. Its job is to identify the evidence, affected range, approved action, and verification. For example:
| Field | Worked example |
|---|---|
| Ticket | LIB-021 |
| Class | Transcription error |
| Video reference | membership-r3 |
| Evidence | At 00:47–00:49, recording says Wednesday; verified against schedule |
| Changes | Transcript 6→7, English caption text 4→5, corrected excerpt |
| Media bytes | No change |
| Verified | Primary page and first-party embed, signed out |
| Pending | Neighboring shop embed, outside direct control |
A useful receipt separates “changed in source” from “observed on public surface.” A screenshot can show the new transcript on your laptop, but it can't prove what the embed sees. Record the URL, asset or version, when it was fetched, and the action tested. Include a fresh session and a viewer path that resembles the reported problem, such as a link that starts at the affected timestamp. For changed closed captions, actually enable the relevant track and play the range rather than confirming the file exists.
Don’t confuse edit history with publicity. Tiny punctuation corrections can be logged internally without an intrusive public banner. Changes affecting a claim that members might rely on should be visible in proportion to their consequence. Apply the publisher's policy about major corrections, archived versions, permission to publish contact details, and handling source material. If the disputed text includes private information or a safety issue, plan the response in the relevant editorial process. A public ledger should not become a permanent copy of sensitive wording the team meant to remove.
Distinguish a version in the build from a copy you checked in playback
It helps to separate several verbs: prepared, released, observed, and acknowledged. An editor can prepare revised caption text and an assistant can verify its parseability. A deploy can release it. A reviewer can observe it on the target route. A partner can acknowledge the request to update a copy. These labels tell us different facts, and they describe different responsibilities. “Partner received email” doesn't prove a partner clip changed, just as “new file in build output” doesn't prove every old CDN response disappeared.
When reviewing a rollout, check the user's path without accidentally testing only your editor session. A preview logged into a CMS may resolve a new manifest while a public page resolves the old one. A detached, signed-out view where applicable can provide evidence for the public rendition delivered at that moment. Verify the visible copy, the selected track, and the affected range in context. If a power user opens a direct download that the site links separately, include that path if it carries the same material. A single link cached locally is not a general freshness assertion; do not overclaim from one observation.
If the issue is visible only in a partner rendition, distinguish a confirmed stale instance from a merely plausible one. Open the exact known route where authorized, collect only the information helpful for correction, and route it through the right owner. If you cannot inspect that rendition, leave it unverified and explain what was or wasn't confirmed. Avoid implying that an archive and a public partner player have the same responsibilities. Their public audience, editing rights and retention constraints may differ.
What automated checks can contribute
A release test can compare the media revision named in the caption manifest with the media revision the player references. It can verify the corrected word exists at the intended anchor, that timecodes are parseable, and that each referenced file is present in the build output. It can detect if the published page reverted to the prior caption asset ID. These are useful mechanical gates, especially when the same material enters search, downloads, and embeds from different pipelines.
That test can't know whether the speaker actually said Wednesday. It cannot establish that a jump link provides enough introductory context. Machine transcription confidence doesn't verify a policy statement. An independent review should check at least the disputed range against the published recording, and should look up the authoritative factual reference if the correction changes what a member should do. Use two steps: confirm what was recorded and confirm what is true. Preserve the difference.
On a text-only change, a narrow review and recheck are often enough; the whole video doesn't need to be remastered. On a new cut, rerun synchronization and accessibility checks according to the changed surfaces. A lightweight inventory can track which clips share the old assets so the team doesn't have to hunt through every page manually. When a hosting platform caches captions independently of video, a versioned track reference avoids relying on everybody clearing their local browser or CDN cache. Test the observed result rather than assuming a new URL proves the correct contents.
Avoid turning an unresolved disagreement into a confident correction
A recording can be unclear for reasons that have nothing to do with editing. A notification tone may mask the first syllable of Wednesday. Two people may speak at once. A recording from a remote call may have a dropout. The written script, prepared before the session, tells you what the presenter intended to say, not necessarily what the video contains. If the source is truly unclear, keep the uncertainty clear where the transcript represents it and use a separately verified factual note for the practical instruction.
Some teams use alternate camera audio or an unedited master to resolve indistinct words. Treat those as evidence and check they correspond to the same moment. A clearer answer from a different take may be a valid replacement if a new edit is authorized; it is not proof of what the current upload says. A timestamp, media reference, and brief rationale is usually more useful than a detailed speculation about why the speaker sounded unclear.
Triage related discoveries instead of letting the first report become an unbounded cleanup. The ticket for Wednesday can close when the affected answer is corrected and verified on the relevant surfaces, with the partner copy tracked separately. An unrelated ambiguous name discovered during review deserves its own check. That keeps the team from either delaying an easy material correction or quietly changing other parts of the recording without evidence.
Rehearse how an honest correction handles both speed and scope
A small publishing team can run a desk simulation before a busy release. Give the reviewer three fictional findings from the same guide: an obvious text mismatch, an ambiguous audio phrase whose expected wording is clear from a script, and a literal but outdated policy claim. Ask the reviewer to assign an owner, identify the necessary checks, and write three small proposed public explanations. Look for the person who can keep the script as evidence of intent without mistaking it for proof of delivered words. Look too for a concrete channel inventory, not a promise to “fix wherever needed.”
Replay with a different constraint. Suppose the video host supports a readily replaceable caption track but not a correction next to the old embedded player. In a training setup, the right choice may be to intercept the site route with a clear contextual note, migrate an authorized embed, or explicitly mark the third-party copy pending. The exercise should not instruct someone to quietly encode current facts as though they were originally spoken. It should make participants name which source matters, what viewers are currently likely to see, and who can make the next change. That is the skill transferable to another publisher.
Track one metric that measures audience benefit. For a handful of important logged cases, check how many verified viewer routes still had a material disagreement after the first release, and note why the team missed the route. A useful pattern might be download text linked by a guide that wasn't in a caption incident checklist. Add that route to future checks, without counting harmless punctuation as a major defect. Do not make editorial courage dependent on a graph of closed tickets. Sometimes a visible correction and explicit outstanding copy are more accurate than a rapid green dashboard.
Decide how an export should behave when it leaves the player
A plain-text transcript can outlive the media on the site. Someone may save it for a meeting, quote it in a newsletter, or read it when playback isn't available. If the only correction is a bubble that appears for three seconds on the current player, ordinary downstream copies of the text will still lack an essential distinction. When scope warrants, place the correction in the transcript itself as named editorial context, not as replacement speech. Provide enough identity for someone to see which edition that downloaded text describes when it isn't embedded in the site.
An export might omit formatting such as background color or iconography. If the difference between a quote and a correction is expressed only by typography, a simple text export can turn a correct edit into ambiguous reconstructed speech. Use plain visible labels and keep meaningful labels when exporting. Try the actual plain-text or printed form if your publication offers one. A quick review can check what the file says without depending on the surrounding player, and confirm that descriptive context still reads as context rather than narration.
Version information shouldn't bury the practical answer. A compact note might identify the guide, media revision and transcript edit date near the top, while placing the specific correction near the disputed passage. A user following a named topic can find “Use form C” where it matters; an editor can still determine whether the cited discussion was from r2 or r3. An older copy may remain on a reader's device indefinitely. No file layout ensures perfect replacement. Explicit attribution and date make it easier to notice when an old copy is being asked to answer a new question.
For translation, preserve the role of a correction even if wording necessarily changes. A translator might use one phrase for a literal quote and another for current policy in a way that reads naturally to the audience. Supply the two pieces as separate fields and share which claims are authoritative. Test a deliberately awkward case on review: the transcript quote says the previous amount, the note supplies the verified amount, and an excerpt generator is configured to select the shortest sentence. Confirm that the public excerpt does not inadvertently promote the old amount alone. An excerpt may be better targeted at a stable link than at any extracted quote.
If a corrected export needs to be regenerated, verify the actual audience-facing download and its name. That link may be edited in a different tool from the library entry. Ensure that someone with the relevant role can identify which copy was observed, whether the superseded export is still intentionally available, and what would need to change if the next policy revision modifies a related passage. For a single affected document this can fit into one row. Its scope matters more than the ceremony around it.
Leave one clear answer where people need it
The goal isn't identical typography on every copy. It is that someone reading the transcript and someone listening to the clip can tell what was said, what the current information is, and which revision they are seeing. That matters most when an old quotation remains available alongside a new factual correction.
A simple release statement might say, “Corrected Wednesday in the transcript and captions; the video itself was already accurate. Verified on the guide and our embed; a partner copy is still pending.” For a factual correction, it might read, “The recording says 9 p.m. The verified time is 8 p.m. The page now provides a correction adjacent to the recording; a revised cut has not been published.” Neither statement requires investigators to guess what was changed. Both make disagreement visible enough to resolve, which is the point of the ledger.
Start with the next smallest correction you actually receive. Write down whether it is a text error, a spoken error, a recut, or a later fact change. Find the copies that people use. Make only the edit that solves the type of error in front of you. Then check the exact viewer path that brought the problem to light. The result should be a more accurate publication rather than a tidier internal transcript that readers may never see.