AI data market
What does media production workflow data add beyond the finished cut?
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Media production workflow data adds three kinds of evidence a finished cut cannot carry: why each revision happened, who approved which version, and the rights attached to each element. Researchers already go to lengths to recreate the first, generating or crowdsourcing before-instruction-after edit sets. Yet no open standard records the why, and the 2026 writers' agreement adds a notice duty for training licenses. Treat a buyer utility pilot as the threshold.
License terms
Media production workflow
Key takeaways
- Production histories add three kinds of evidence finished content lacks: revision reasons, version approvals and element-level rights.
- Researchers already recreate the before-instruction-after shape: MagicBrush crowdsourced over 10,000 triplets, and CommitPack gathered almost 4 TB of Git commits.
- C2PA 2.5 lists 33 current actions; only one records a review, and standard reason values exist only for redactions.
- The WGA's 2026 agreement requires written notice to the Guild when writers' material is licensed to train a commercial generative AI system.
- No paired test of histories against finished assets is public; make that pilot the diligence threshold for any supplier.
Is a production history a separate asset from the content it produced?
It is, when the history carries evidence the content cannot: the reason for each revision, the approval of each version and the rights attached to each element. A finished film, commercial or podcast episode shows the result. The workflow records around it show the choices, and choices are what an editing assistant, a review agent or a clearance tool has to learn. For an investor the question is not whether media matters to AI, but whether those records form a supply with its own buyers, its own rights and its own preparation cost.
Here, media production workflow data means the versions of an asset, the notes that requested each change, the status and approval events on each version, the editorial timeline that assembled the cut and the clearance records for its elements. Finished content means the delivered masters and their published metadata.
Judging the category takes comparable evidence on five tests: a buyer task that depends on the history, a consistent way to capture it, rights that allow licensing, a preparation burden someone can price, and measured utility over finished assets alone. Public sources answer the first three in part. Nobody has published the last two, and this article marks those gaps instead of filling them with estimates.
Where demand for revision evidence already shows up
The clearest signal is that researchers go to real lengths to manufacture the before-instruction-after shape that production histories hold naturally. InstructPix2Pix (Brooks, Holynski and Efros, 2022) obtained its training data by combining a language model, GPT-3, with a text-to-image model, Stable Diffusion, to generate a large dataset of image-editing examples. MagicBrush (Zhang and colleagues, 2023) then used crowd workers to build over 10,000 manually annotated source, instruction and target triplets on real images, arguing that earlier methods were zero-shot or trained on automatically synthesized data with a high volume of noise.
Software shows what happens when such histories are public. The OctoPack paper (Muennighoff and colleagues, 2023) assembled CommitPack, almost 4 terabytes of Git commits across 350 programming languages, because a commit pairs a change with a human-written description of it, and then filtered it to a 2-gigabyte instruction-tuning set covering 277 languages. A version with its review note is the media equivalent of a commit with its message. No open media counterpart to CommitPack turned up in this research, and studio version histories sit in private trackers.
Production itself is adopting AI tools that would need this evidence to improve. Netflix told shareholders in its second-quarter 2026 letter that GenAI workflows had been used in roughly 300 of its titles in 2026, with the largest concentration of work in post-production. That is a statement about tool use, not about data purchases, and it should be read narrowly. No public statement found for this article names production histories as a data category a model developer is buying.
What production tools capture, and what no open standard records
What changed in a revision is increasingly recorded in standard form; why it changed is not. OpenTimelineIO, an interchange format hosted by the Academy Software Foundation, describes itself as a modern edit decision list for editorial cut information: clips, timing, tracks, transitions, markers and metadata, with video and audio referenced externally rather than embedded. Its architecture notes say every object carries a metadata dictionary for blind data, which is where any reason for a cut would have to sit, in whatever shape a studio chose.
Content provenance stops at the same boundary. We counted the pre-defined actions in version 2.5 of the C2PA technical specification at repository commit 3d91c65 (October 25, 2026). Its Table 4 lists 36 actions, 3 of them deprecated; of the 33 current actions, exactly one, c2pa.reviewed, records that content was assessed without being changed. An action may carry a reason field, but the specification notes that only redaction-focused values are defined so far, four in all, such as c2pa.PII.present. It also says the order of actions in the array does not imply the order in which they were performed.
Approvals live in vendor systems. In Autodesk's Flow Production Tracking, formerly ShotGrid, a Version tracks an iteration submitted for review and moves through statuses from Pending Review to Final Approved. Those statuses belong to one vendor's data model rather than to an open standard, so approval histories exported from different tools need mapping before anyone can compare them, and the mapping is preparation work a buyer will ask about.
Rights that travel separately from the pixels
Production histories carry rights that finished content does not, because drafts, notes and outtakes involve people whose work never shipped. The Writers Guild of America says it reserves the right to assert that using writers' material to train AI is prohibited by its contract or other law. Its summary of the 2026 Minimum Basic Agreement adds a concrete duty: companies must give the Guild written notice when they license writers' material to train a commercial generative AI system to create outputs, and the Guild can ask to discuss the license, including any remuneration. The summary says this also reaches licenses of produced projects based on scripts.
That makes script drafts and writers' revision notes a rights layer of their own. Owning a draft therefore does not end the analysis: under the 2026 terms, licensing it to train such a system still calls for notice to the Guild. Performer likeness, music cues, stock footage and client-owned brand assets add further layers, and each one can remove records from an otherwise usable history.
Machine-readable permission does not settle the question either. Version 2.0 of the C2PA specification, dated January 2024 in its change log, removed the Training and Data Mining assertion from the core standard, so a manifest built only on the current core specification does not say whether an asset may train a model. For production histories, permission has to be established from contracts and clearance records, element by element.
Sector opportunity scorecard: five evidence layers, four tests
The scorecard rates each layer of a production archive on four tests, using only the public evidence cited above, and it rates no company. Buyer evidence is strong when a cited training set or paper uses the layer's shape and none when the sources reviewed show no such use. Capture is partial when an open format records the structure but not the reasons, and weak when only vendor tools or free text hold it. The rights path names the party or term that decides a license. Utility counts only a published paired test against finished assets.
Read across a row to see where a supplier's claim needs proof; the last two columns are where a diligence request should start. The scorecard:
| Evidence layer | Public buyer evidence | Open capture standard | Rights path | Utility measured over masters |
|---|---|---|---|---|
| Finished masters | Baseline: not rated here | Baseline: not rated here | Mixed: talent, music, stock and client layers | Baseline arm of the pilot |
| Editorial timeline | None: not found in the sources reviewed | Partial: OpenTimelineIO, reasons left to blind metadata | Contract: producer agreements decide | None: no public paired test |
| Versions and notes | Strong: edit triples and commit histories | Weak: vendor trackers and free-text notes | Notice: Guild duty for writers' material | None: no public paired test |
| Approval events | None: not found in the sources reviewed | Weak: vendor statuses; one C2PA review action | Contract: producer agreements decide | None: no public paired test |
| Clearance records | None: not found in the sources reviewed | Weak: free-form; no training flag in C2PA core | Mixed: depends on each license's terms | None: no public paired test |
The buyer utility pilot that would settle the question
Only a paired pilot can show whether histories add utility beyond finished assets, and no results from such a pilot are yet published, by SourceX or, as far as this research found, anyone else. The design below is what an investor should ask any supplier to run or to show.
At a glance
The buyer utility pilot that would settle the question
A pilot that reports only the history arm, or chooses projects after seeing them, should not move an investor. Its design has six parts:
- Population: a fixed list of completed projects, chosen before any records are examined, with every exclusion logged by reason: rights, missing links or damaged files.
- Arms: A holds finished masters with their published metadata; B holds the same projects plus linked versions, notes, approvals and timelines.
- Tasks: name the instruction behind the change from one version to the next, predict whether a version was approved, and flag elements a clearance record blocks.
- Metrics: accuracy per arm on held-out projects, with the project as the unit of analysis and confidence intervals from resampling projects under a published random seed.
- Preparation cost: staff hours per project to link notes to versions and clear rights, plus the share of elements excluded.
- Decision rule, fixed in advance: the history counts as a distinct asset only if arm B beats arm A, after exclusions, on a task a named buyer has asked for.
Does a bigger archive mean more licensable supply?
No. Archive size counts files, while licensable supply counts linked and cleared histories, and the two can diverge sharply. A producer with decades of finished work may hold versions without notes, notes in email, approvals in a retired tracker and drafts that carry Guild obligations, leaving few projects where every layer survives.
Three ratios separate the two quickly: the share of versions with a linked note, the share with an approval event tied to the exact version, and the share of elements with a clearance record. Ask for each by project, with the project count as the denominator. Generative tools used inside production add a fourth check: when AI-generated material sits in a history, its provenance has to be recorded too, or a buyer may end up training on another model's output without knowing it.
What could make the production-history thesis wrong
Four developments would weaken the thesis, and each comes with a signal an investor can track:
- Synthetic edits become good enough. InstructPix2Pix showed that edit examples can be generated at scale from pretrained models; if generated triplets match real histories on buyers' acceptance tests, the premium for real revision records shrinks. MagicBrush's complaint about noise in synthesized data is the counterweight to watch.
- Rights friction outweighs value. The Guild's 2026 notice-and-discussion duty and its reserved position on training add steps to every license that touches writers' material; if comparable terms spread to other crafts, preparation costs rise for every supplier.
- Capture never standardizes. With reasons for cuts left to blind metadata and approvals held in vendor statuses, every archive needs custom mapping; if that cost stays high, usable supply stays small.
- Producers keep histories for their own tools. Netflix's reported use of GenAI workflows across roughly 300 titles suggests large producers are investing in production AI themselves, and a producer building in-house may prefer to keep its history rather than license it.
Questions for a production-data supplier, SourceX included
The useful questions test the layer where a supplier claims its advantage, using definitions the investor sets rather than ones the supplier chooses. SourceX is one company they apply to. For businesses licensing operational records to AI developers, it finds and qualifies the companies, reviews rights, writes the contracts, prepares and delivers the data and handles payment. It does not publish operating or commercial figures in articles and shares company metrics with investors directly. Anyone weighing SourceX as an investment can ask for its materials on the investors page; the company is raising only from accredited investors, under Rule 506(c).
Five questions to start with:
- Which evidence layers does each project in your sample keep, and what share of versions carry both a linked note and an exact-version approval?
- How were rights cleared for drafts, notes and outtakes, and which projects required notice to a guild?
- What share of elements was excluded, and for which reasons?
- Has anyone run a paired utility test against finished assets alone, with projects fixed in advance?
- How are approval statuses from different tracking tools mapped to one vocabulary, and who checked the mapping?
How the scorecard and the C2PA count were built
Research for this article was completed on October 27, 2026. The C2PA counts come from a short script that parses version 2.5 of the specification at repository commit 3d91c65: rows of Table 4 that begin with c2pa. are counted as actions, rows whose meaning begins with DEPRECATED as deprecated, and the reason values are those listed in section 16.4.1.4.2. OpenTimelineIO statements come from its documentation at commit 94e1ad6. The paper figures, the Netflix statement and the Guild terms were checked on their publishers' own sites.
The scorecard ratings are editorial judgments that apply the rules stated with the table to those sources; they rate evidence layers, not companies. Not measured: how many archives keep linked notes or approvals, preparation hours, rights exclusion rates, prices, and the utility of histories over finished assets. Those need the pilot above, and its results are not yet published.
Frequently asked questions
What is media production workflow data?
It is the record of how a piece of media was made: the versions of each asset, the notes that requested changes, the approval events on each version, the editorial timeline and the clearance records for each element. Finished content is only the delivered result. The workflow data shows the decisions behind it.
Why would an AI developer want production histories instead of finished content?
Editing and review tools learn from pairs of before and after with the instruction in between, which finished content does not contain. Researchers have generated such examples synthetically (InstructPix2Pix) or used crowd workers to create them (MagicBrush); real histories hold the same shape from actual productions, though their added utility has not been measured publicly.
Do union agreements restrict licensing script drafts for AI training?
They add obligations. The Writers Guild of America reserves the right to assert that training on writers' material is prohibited, and its 2026 agreement requires written notice to the Guild when writers' material is licensed to train a commercial generative AI system, with a right to discuss terms. This is general information; counsel should review any specific license.
Does the C2PA standard record why an edit was made?
Mostly not. Version 2.5 records what actions were taken, optionally when, and defines standard reason values only for redactions. A free-text description can be attached to an action, but the reasons behind creative revisions have no standard vocabulary.
Does SourceX publish results or metrics for production-history datasets?
No. SourceX does not publish operating or commercial figures in articles, and no utility pilot results are published. It shares company metrics with investors directly; materials can be requested at sourcex.si/investors, and any offering is made only to verified accredited investors.
Sources
- InstructPix2Pix combined GPT-3 and Stable Diffusion to generate a large dataset of image-editing examples for training. Source
- MagicBrush built over 10,000 manually annotated source, instruction and target triplets for real-image editing, noting earlier methods trained on noisy synthesized data. Source
- OctoPack's CommitPack holds almost 4 TB of Git commits across 350 languages, filtered to a 2 GB, 277-language instruction set; commits pair code changes with human instructions. Source
- Netflix's Q2 2026 letter: GenAI workflows used in roughly 300 titles in 2026, concentrated in post-production. Source
- OpenTimelineIO is an interchange format for editorial cut information, a modern EDL, supporting clips, timing, tracks, transitions, markers and metadata with media referenced externally. Source
- In OpenTimelineIO, all objects have a metadata dictionary for blind data. Source
- C2PA 2.5: 36 pre-defined actions, 3 deprecated, one review action; standard reason values only for redaction; action order not implied; Training and Data Mining assertion removed in 2.0. Source
- In Flow Production Tracking, a Version tracks an iteration submitted for review, with statuses from Pending Review to Final Approved. Source
- The WGA reserves the right to assert that exploitation of writers' material to train AI is prohibited by the MBA or other law. Source
- The 2026 WGA MBA requires written notice to the Guild when writers' material is licensed to train a commercial GAI system, with a right to discuss the license including remuneration. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.