Systems and records
Creative revision decision dataset: keeping the reason for each change
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Creative workflow data should store each revision as a link between the feedback that requested it, the version that answered it and a reviewer's decision on that version, all tied to the brief version in force. A rights screen for every stock, music, talent, font or client component then decides what may be delivered or reused. Keeping the reasons attached is what makes a creative revision decision dataset explain why the final work won.
License terms
Creative revision decision
Key takeaways
- A revision is only explainable when it links the feedback that asked for it, the version that answered it and the decision made on that version.
- Review tools attach comments to the version being criticized, so keeping only final files loses the reasons.
- Map every tool's approval labels to a short fixed list, and keep conditional approvals distinct from plain ones.
- Screen every stock, music, talent, font, client and generated component; unknown rights block reuse.
- When rights are unclear, a metadata-only case keeps the decision trail without moving third-party media.
- Validate each case with simple joins and date checks, and reject failures rather than patching them.
How should a creative revision decision dataset preserve the reasons behind revisions?
Store every revision as a link between three things: the feedback that asked for the change, the version that answered it, and the decision a named reviewer made about that version. The reason lives in the feedback text and the decision value, so neither can be dropped or merged into a free-text project note.
Most creative archives keep the files and lose the argument. A shared drive holds v1 through v7 of a launch video, the review tool holds timecoded comments, the project tracker holds the approval, and the brief sits in an email thread. Each system is complete on its own terms, yet none of them can answer the question a model, or a new art director, actually needs answered: what changed between v4 and v5, who asked for it, against which version of the brief, and was the result accepted.
A creative revision decision dataset reassembles that chain into cases. One case is one deliverable from first brief to final approved scope, with every feedback round, revision and decision joined by stable identifiers. The rights screen sits inside the same case, because a revision that swaps a stock image or a music cue can change what may be delivered or reused even when nobody mentions it in the comments.
Why the final file alone cannot explain why the work won
A final asset shows the outcome but hides the choices that produced it. The rejected directions, the client note that killed a headline and the legal comment that forced a disclaimer are the parts that teach judgment, and they exist only in the review trail.
Review tools treat feedback as belonging to a specific version. Frame.io's help center, for example, notes that on its legacy platform comments are not carried over within version stacks, and recommends copying them to a new version by hand. In practice that means the reason for a change sits on the version that was criticized, not on the one that fixed it. A dataset that keeps only the latest version keeps the fix and loses the request.
Provenance standards do not fill the gap either. The C2PA specification describes an actions assertion that records edits and other actions taken on an asset, but it states that the order of actions in that list is unspecified, and the standard reason values it defines are focused on redaction. Content credentials can tell you that an image was cropped or color-adjusted; they do not record that the client asked for warmer skin tones in round two. Treat them as useful evidence about what happened to a file, not as the decision record.
What the rights screened brief, feedback and revision linkage schema includes
The schema has seven linked record types, each with a required identifier that the next record points back to. The table below is the asset: field groups, what each one proves, and the check that must pass before a case is accepted. Field names are suggestions; the joins are what matter.
- Keep reviewer identities as roles (client brand lead, legal reviewer, creative director) unless the agreement with each person allows names.
- Store comment text verbatim and add the feedback type as a label; do not summarize the comment in place of the original.
- Version the schema itself (for example, schema version 1.0 dated 2026-10-18) so later changes to a field are traceable.
| Record | Required fields | What it proves | Validation check |
|---|---|---|---|
| Brief version | brief_id, brief_version, issued_at, objective, audience, mandatory elements, constraints, territories, channels | What the work was asked to achieve at a given moment | Every revision cites exactly one brief_version that existed before the revision date |
| Asset version | asset_id, version_number, parent_version, created_at, file hash, tool or source system | Which file a reviewer actually saw | parent_version points to an earlier version of the same asset_id; hashes are unique |
| Feedback item | feedback_id, asset_version, reviewer_role, timecode or region, comment text, feedback type (fix, preference, legal, brand) | What was asked for, by whom and where in the work | Each feedback item attaches to a real asset_version and has a reviewer role, not a personal name |
| Revision | revision_id, from_version, to_version, addressed_feedback_ids, change summary | Which request each change answered | addressed_feedback_ids are all on from_version or an earlier round; to_version is the direct child of from_version |
| Decision | decision_id, asset_version, decision value, decided_at, decider role, next steps text | Whether the version advanced, and on what condition | Decision value comes from a fixed list; a conditional approval must carry next steps text |
| Rights screen | component_id, component type (stock, music, talent, font, client asset, generated), license or release reference, permitted uses, territories, term end, AI training permitted (yes, no, unknown) | What may be delivered and what may be reused | Every component in the final version has a screen record; unknown blocks reuse |
| Final scope and outcome | final_version, approved channels, approved territories, launch date, excluded components | What actually shipped and under which limits | Final version has an approving decision and no component whose screen fails |
Which decision values should the schema allow?
Use a short fixed list of decision values and map every tool's labels into it, because the difference between a conditional and an unconditional approval is the most valuable signal in the record. Free-text statuses such as "looks good, small tweak" cannot be compared across projects.
Adobe Workfront's proofing documentation is a useful reference point. It separates Approved with changes, where the proof needs some changes but the reviewer does not need to see the revision before it moves on, from Changes required, where the reviewer must see another revision first. That distinction tells a model whether a reviewer trusted the team to finish the change unsupervised, which is exactly the kind of judgment the dataset is meant to capture.
A workable list has five values: approved, approved with conditions, changes required, rejected, and not applicable to this reviewer. Keep the tool's original label in a separate field so a reviewer can audit the mapping, and record administrator customizations, since some tools let account owners rename or hide decision options.
Which approval, rights condition or revision changes what may be delivered or reused?
Three kinds of event change the deliverable scope rather than just the creative: a rights condition attached to a component, a decision that carries conditions, and a revision that introduces or removes a third-party element. The schema must treat each as a scope event, not as an ordinary comment.
Ownership comes first. Under U.S. law, a commissioned work is a work made for hire only if it falls within one of the statutory categories and the parties signed a written agreement that expressly says so; the U.S. Copyright Office's Circular 30 sets out those criteria and states that a work failing any of them is not a work made for hire. An agency's archive can therefore contain work it created but does not own, work the client owns, and work where ownership turns on a contract nobody has re-read. Record the agreement reference in the case, not a conclusion about who owns what.
Component rights come next. The IPTC Photo Metadata Standard includes fields such as Model Release Status, which summarizes the availability and scope of model releases, and Web Statement of Rights, a link to a statement of copyright ownership and usage rights. When those fields exist in your files, import them into the rights screen; when they do not, the screen record has to be filled from the license or release itself.
- A stock image licensed for one campaign: deliverable for that campaign, excluded from reuse unless the license says otherwise.
- A talent release limited to certain territories or a term: the final scope inherits the narrower limit, and reuse after the term end is blocked.
- A client-supplied asset such as a logo or product shot: owned by the client; the case can describe how it was used but should not include the asset file without client permission.
- A conditional approval from legal (for example, add a disclaimer): the next version must show the change, and the condition stays attached to the final scope.
- A revision that replaces a licensed music cue with a library track: rescreen the new component before the version can be approved.
A campaign cut that passes validation, and one that does not
The worked case below is illustrative: a fictional mid-size agency, a fictional beverage client and invented identifiers. It is not SourceX customer data. It shows one case that passes every check and one variant that fails, so a reviewer can test a validator against both.
| Step | Valid record (illustrative) | Failing variant (illustrative) | Check that catches it |
|---|---|---|---|
| Brief | BR-12 v2 issued day 3: 15-second social cut, U.S. only, must show the can within 3 seconds | Revision cites BR-12 v3, which was issued after the revision date | Brief version must predate the revision |
| Round 1 feedback | FB-31 on asset A-7 v1 at 00:04, client brand lead: can appears too late (fix) | FB-31 attached to A-7 v2, a version that did not exist yet | Feedback must attach to an existing version |
| Revision | RV-9 from A-7 v1 to v2, addresses FB-31: product shot moved to 00:02 | RV-9 from v1 to v3, skipping v2 | to_version must be the direct child of from_version |
| Decision | Legal reviewer on v2: approved with conditions, next steps: add age disclaimer | Approved with conditions and empty next steps | Conditional decisions require next steps text |
| Rights screen | Stock clip S-4: licensed for campaign, U.S., 12 months, AI training not permitted | Music cue M-2 in final version with no screen record | Every final component needs a screen record |
| Final scope | v3 approved by client brand lead; U.S. social only; S-4 excluded from reuse | v3 marked final with no approving decision | Final version must carry an approving decision |
How to validate a case before it enters the dataset
Run the checks as code, in order, and reject the whole case on the first structural failure rather than repairing it silently. A rejected case with a reason is a better record than a patched one whose history no longer matches the source systems.
The checks are simple joins and date comparisons, so a short script over exported tables is enough; no special tooling is needed. Keep the validator's output beside the dataset so the person reviewing the delivery can rerun it.
Checklist · 6 items
- Identity: every asset_version, feedback_id, revision_id and decision_id is unique within the case.
- Lineage: each version except the first has exactly one parent of the same asset; no cycles. W3C PROV-O offers a vocabulary for this: wasRevisionOf is the specific form of wasDerivedFrom for a revised version that keeps substantial content from the original.
- Chronology: brief version, feedback, revision and decision timestamps fall in a consistent order for each round.
- Coverage: every revision addresses at least one feedback item, and every feedback item marked fix is addressed by a later revision or explicitly closed with a decision.
- Rights: every component in the final version has a screen record, and any component with AI training marked no or unknown is excluded from the training copy.
- Privacy: comment text is scanned for personal names, emails and phone numbers of reviewers and talent before export, and reviewer fields hold roles only.
0 of 6 ticked. Tick each item as you go.
Where does each part of a case live, and how do you join it?
Plan for five sources and one join key per pair, because no single creative system holds a whole case. The brief usually lives in a project tracker or a shared document, the versions and comments in a review or proofing tool, rights metadata in a digital asset manager, licenses and releases in a contract folder, and a surprising share of client feedback in email.
The hardest join is between files and comments. Review tools identify versions by their own internal IDs, while the shared drive identifies them by filename. Hash each exported file and store the hash on both sides, so a version in the review tool and a file on disk can be matched without trusting naming conventions that change from project to project.
- Project tracker or brief document: brief versions, issue dates, mandatory elements and territory limits.
- Review or proofing tool: asset versions, timecoded or region-pinned comments, reviewer roles and decision values.
- Digital asset manager: component metadata, embedded rights fields and usage restrictions.
- Contract and release folder: stock licenses, music licenses, talent and property releases, client agreements.
- Email and chat: client feedback given outside the review tool; log each item as a feedback record linked to the version it refers to, or mark it unlinked.
What to leave out when the rights screen is incomplete
Leave out the asset files and keep the decision trail whenever a component's rights are unknown. The comments, revision summaries and decisions describe your team's process; the pixels and audio may belong to a photographer, a stock library, a musician or the client.
This split is often the difference between a case that can be offered and one that cannot. A text-and-metadata case (brief, feedback, change summaries, decisions, scope) still teaches how feedback becomes a revision, and it avoids moving third-party material you have no right to license. Mark such cases clearly as metadata-only so a buyer does not assume the media is included.
Also exclude any case where the client contract forbids disclosure of campaign materials or strategy, and cases about unreleased work that is still under embargo. Keeping access to a review tool, holding a file on your drive and having a license to redistribute it are three different things, and removing names does not create a right to license.
Where this fits if you consider licensing creative workflow records
If an agency or in-house studio is weighing whether its review history could become a licensed dataset, the schema above doubles as a readiness test: cases that pass the validator and the rights screen are the ones worth discussing. At that stage SourceX can turn selected review records into a licensed dataset for AI, running the rights review, contracting and data preparation itself.
The opening step is a short company assessment that needs no data upload. Nothing is delivered until the company has approved the scope and terms and signed the agreement, so a metadata-only scope, or one that excludes certain clients entirely, can be set before any record leaves your systems.
Frequently asked questions
What is a creative revision decision dataset?
It is a set of cases, one per deliverable, that links each brief version, feedback item, revision and reviewer decision with stable identifiers, plus a rights screen for every third-party component. The point is to keep the reason for each change attached to the change, so the record explains why the final work was approved.
Do we need the original image, video or audio files?
Not always. A metadata-only case with the brief, verbatim feedback, change summaries and decisions still shows how feedback turns into revisions. Media files should be included only when every component has a cleared rights screen.
Can content credentials replace a revision log?
No. The C2PA standard records actions taken on an asset, but it states that the order of actions in its list is unspecified, and its defined reason values focus on redaction. Use content credentials as supporting evidence of file changes, not as the record of who asked for a change and why.
How should reviewer names be handled?
Store roles such as client brand lead or legal reviewer instead of names, and scan comment text for names, emails and phone numbers before export. Keep a private mapping only if you need it for audits, and never ship it with the dataset.
Who owns work an agency made for a client?
It depends on employment status and the contract. The U.S. Copyright Office explains that commissioned work counts as a work made for hire only in specific categories and with a signed written agreement saying so. Record the agreement reference in each case and have counsel confirm ownership before anything is licensed.
Sources
- Adobe Workfront separates Approved with changes (no need to see the revision before it advances) from Changes required (another revision must be seen), and administrators can customize decision options. Source
- Frame.io's help center says comments are not carried over within version stacks on its legacy platform and recommends copying them to the new version. Source
- The U.S. Copyright Office explains that a commissioned work is a work made for hire only if it is in a listed category and covered by a signed written agreement expressly saying so. Source
- The IPTC Photo Metadata Standard defines Model Release Status and Web Statement of Rights fields for release scope and usage rights. Source
- The C2PA specification's actions assertion records edits taken on an asset, states that the order of actions is unspecified, and defines only redaction-focused reason values. Source
- W3C PROV-O defines wasRevisionOf as a specific form of wasDerivedFrom for a revised entity that keeps substantial content from the original. Source
See if your company qualifies
A short company assessment. No data uploads are needed.