Rights and contracts
Dataset exclusion transformation controls: trace every removed record
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Dataset exclusion transformation controls prove that a removed record, field or document stays out of every artifact made from the source: samples already shared, working copies, derived fields, summaries, embeddings, splits and the delivery manifest. Record the decision once, with who decided, the governing document and the evidence, then test each artifact and hold delivery until every one passes or is rebuilt.
License terms
Dataset exclusion
Key takeaways
- An exclusion is done only when the removed content is absent from every artifact that leaves your control, and you can show it.
- Content survives in eight places: shared samples, working copies, derived fields, quoted text in other records, generated summaries and labels, embeddings and indexes, splits and the delivery manifest.
- Rebuild anything computed from excluded content rather than deleting rows from its output; embeddings can be inverted to recover much of the original text.
- Redaction and filters carry out a rights decision but never make it; a contractually excluded record stays excluded however well it is masked.
- Privacy law and regulators already push removals to recipients and derived models, so the review log should name every recipient of an earlier copy.
When does a dataset exclusion count as done?
An exclusion is done when the removed records, fields or documents are absent from every artifact that leaves your control, not only from the main export, and you can show it. That covers samples already shared, working copies, every derived file and summary, any embeddings or search indexes built during preparation, and each file named in the delivery manifest.
Every exclusion starts as a permission question with an owner. Someone with authority over the records decides that a set of them may not be licensed for a requested use; a governing document says why, whether a customer contract, a notice, a confidentiality obligation or a third party's rights; and evidence shows exactly which records the decision covers. Engineering then has to carry that decision through each transformation without losing any of it.
The usual failure is not a bad filter. It is a good filter applied once, to one table, after copies and derivatives already exist. This guide traces where removed content survives, gives a permission matrix and a review log for recording the decision, and sets the conditions for holding, rebuilding or narrowing a delivery. It is general information: the rights questions themselves belong to your counsel.
Where removed content survives: eight places to test
Removed content survives wherever an artifact was made before the exclusion was decided, or was made from a copy the exclusion never reached. Test each place separately, because a clean main export says nothing about the others.
Two places deserve emphasis. Text duplicated inside other records, such as an email quoted in a reply or a ticket pasted into an internal note, escapes any rule keyed on record IDs: the excluded message is gone and its words remain. And anything computed from the content keeps information about it. Morris and colleagues showed that text embeddings can be inverted to recover much of the original text (paraphrase), so a stored embedding of an excluded document is not an exclusion.
The map, place by place:
| Where it survives | How the content gets there | Test that finds it | Evidence to keep |
|---|---|---|---|
| Samples already shared | A review sample sent before the exclusion was decided | Match the sample's record keys against the exclusion list | Sample manifest, recipient and a deletion confirmation |
| Working copies and caches | Staging tables, notebooks and exports made earlier | Inventory every copy by location and creation date | Copy inventory with a deletion or rebuild entry for each |
| Joined tables and derived fields | A join or aggregate computed before the rule ran | Recompute after the exclusion and compare counts | Before and after counts for each derived field |
| Text inside other records | Quoted replies, forwarded threads, pasted notes | Search kept records for excerpts of excluded ones | Hit count and the treatment of each hit |
| Summaries and labels made in preparation | Generated text or labels built from excluded content | Trace each derived row to its parent IDs | Parent-link coverage report |
| Embeddings and search indexes | Vectors or indexes computed from excluded content | Rebuild from the clean source and diff the item lists | Index build log tied to the source version |
| Splits and sample packs | Train, test or pilot files cut from an earlier version | Recut from the clean version and diff record keys | Split manifest naming its source version |
| Delivery manifest | A manifest listing files built before the exclusion | Check each file hash against the clean build | Signed manifest with hashes and the exclusion version |
Why a filter or redaction does not settle the permission question
Technical access, a working filter or a redaction pass shows what was removed, not whether what remains may be licensed. An exclusion records a rights decision; the transformation that carries it out cannot make that decision.
Privacy regulators draw the same line from the other side. The UK Information Commissioner's Office guidance on anonymisation treats pseudonymised data as personal data in the hands of anyone holding the extra information needed to re-identify people, and judges identifiability by what a recipient could reasonably use (paraphrase). Masking names in one customer's tickets does not turn that customer's contractually restricted records into licensable ones.
So keep two questions apart in the record. Whether a record is permitted for the requested use is a rights decision with an owner and a governing document. Whether the delivered data meets the agreed privacy standard is a processing question. A record can fail the first and pass the second, and no amount of redaction moves it from excluded to permitted.
The rights-to-records permission matrix
The matrix sets each record family against each requested use and records the permission state, the governing document, the evidence that settles it and how the exclusion is carried. Fill it before any sample leaves, and treat every cell that still needs review as excluded until counsel resolves it.
The entries are illustrative, written for a hypothetical services company; they show how precise a cell must be, not the answer for any real dataset:
| Record family and requested use | Permission state | Governing document | Evidence that settles it | How the exclusion is carried |
|---|---|---|---|---|
| Support tickets, general accounts: training | Permitted, with de-identification | Customer terms version in force at collection | Terms text and the accounts each version covers | No exclusion; de-identification step logged |
| Support tickets, one enterprise account: training | Excluded | That customer's master agreement confidentiality clause | Clause reference and the account's identifiers | Account ID list plus a text search for its name and domains |
| Call recordings: any use | Needs review; excluded until resolved | Recording notices and the consent rules that apply | Notice versions by date and channel | Whole family kept out of samples and derivatives |
| Internal notes quoting customer emails: training | Permitted once quoted text from excluded accounts is removed | The quoted email's own account terms | Excerpt search results | Excerpts removed before summaries or embeddings are built |
| Contracts with third-party schedules: evaluation | Schedules excluded; main terms need review | The third party's confidentiality terms | Schedule identifiers and counsel's note | Document-level exclusion with a page map |
The review log: one entry per exclusion decision
The review log is the decision record that travels with the data. Write one entry when each exclusion is decided and update it as each artifact passes its test, so anyone can see who decided, on what authority, what the decision covers and where it has been shown to hold.
At a glance
The review log: one entry per exclusion decision
It differs from an exclusion summary prepared for a buyer: this is the supplier's and counsel's working record, and its last fields only fill in as the transformations are tested. Each entry holds:
- Decision ID and version, with the date decided and the date of any amendment.
- Decided by: the person with authority over the records and the role that gives them that authority.
- Governing document: the clause, notice version or obligation that requires the exclusion, cited by reference rather than quoted.
- Scope unit: record family, account, record, field, document or passage, with the identifiers that define it.
- Matching rule: the exact key list, field condition or text search that finds covered content, with its own version number.
- Artifact status: for each of the eight places on the map, pass, rebuilt, deleted or not applicable, with the test run and its date.
- Residual risk: what the tests cannot rule out, such as paraphrased mentions, and who accepted it.
- Delivery link: the manifest version and the exclusion version the delivered files were built from.
- Downstream notices: each recipient of an earlier sample, and confirmation that its copy was deleted.
How an exclusion changes over time without breaking the chain
Treat the exclusion list as versioned data: every amendment gets a new version, every artifact records the version it was built under, and any artifact built under an older version is rebuilt or retested before it ships.
Exclusions change for ordinary reasons. Counsel narrows a record family after reading a contract more closely, a customer's renewal adds a confidentiality clause, or an improved matching rule finds mentions the old one missed. Each change quietly invalidates artifacts that were correct a week earlier, and without a version stamp on each artifact nobody can tell which ones.
The rule that works is mechanical. At packaging time, compare each artifact's exclusion version with the current one. If they match, the earlier tests stand. If the artifact's version is older, rebuild it from the clean source or rerun all eight tests, and record which. A widening amendment, one that excludes more, always triggers this. A narrowing amendment, one that releases records back into scope, needs counsel's sign-off before anything is added back, because releasing content is a rights decision too.
Keep superseded versions. When a buyer, an auditor or your own counsel later asks why a record was or was not in a delivery, the answer is the exclusion version named in that delivery's manifest and the log entry that produced it.
What laws and regulators already expect downstream
Several regimes already require removals to follow data to its recipients, which is why the review log names every recipient of an earlier copy. None of this is legal advice; whether a rule applies depends on the data, the parties and the jurisdiction, and the summaries below are paraphrases of texts your counsel should read in full.
Under the EU General Data Protection Regulation, Article 19 requires a controller to tell each recipient of personal data about a rectification, erasure or restriction carried out under the relevant articles, unless that proves impossible or involves disproportionate effort (paraphrase). California Civil Code section 1798.105(c) requires a business honoring a verified deletion request to delete the information, notify its service providers and contractors to delete it, and notify third parties to whom it sold or shared the information, with the same impossibility exception (paraphrase; check the current code, since amendments can change the text).
Regulators have also reached past the data into what was built from it. In December 2023 the FTC announced an order requiring Rite Aid to delete photos and videos collected by its facial recognition system, along with data, models and algorithms derived from them, and to instruct third parties that received such images to delete them as well (paraphrase of the FTC's announcement and order summary).
Research on withdrawn datasets shows what happens without that discipline. Peng, Mathur and Narayanan found that after the DukeMTMC and MS-Celeb-1M datasets were retracted, their data remained widely available and kept being used in research, which they call runaway data (paraphrase). A withdrawal that does not reach copies and derivatives withdraws very little.
Worked trace (illustrative): one excluded account through a support-ticket export
Follow one excluded enterprise account through a hypothetical export of support tickets to see where each test bites. All counts are invented for illustration and describe no real company.
The decision comes first. Counsel concludes that the account's master agreement bars outside use, so every record tied to it is excluded for every use. The matching rule combines the account ID with a text search for the customer's name and email domains, because the account also appears in other customers' tickets when its staff are copied in.
Then the trace. The main export drops the account's 1,200 tickets by ID. The text search finds 85 kept tickets quoting its staff, and those passages are removed. A 500-ticket review sample sent to a prospective buyer two weeks earlier contains 9 of the account's tickets, so the recipient is asked to delete the sample and confirm in writing. Summaries generated for a pilot were built before the decision; 40 trace to excluded parents, and the whole summary file is regenerated rather than trimmed. The embedding index is rebuilt from the clean source instead of patched. Only then is the manifest signed, naming the exclusion version every file was built from.
The order is the lesson. Decision and matching rule first, each copy and derivative tested against them one at a time, and the manifest last, as the proof that the earlier steps happened rather than a substitute for them.
When to hold delivery, rebuild an artifact or narrow the scope
Hold delivery while any place on the map is untested or failing, rebuild an artifact rather than patch it when it was computed from excluded content, and narrow the scope when the matching rule cannot reliably find covered content. The conditions in practice:
At a glance
When to hold delivery, rebuild an artifact or narrow the scope
- Hold: a sample recipient has not confirmed deletion, or a cell in the permission matrix still reads needs review.
- Rebuild: summaries, labels, embeddings, indexes or splits made before the exclusion; regenerate them from the clean source instead of deleting rows from the output.
- Narrow: excluded content appears in free text that no search can bound, such as paraphrases in internal notes; drop the affected record family rather than ship a known gap.
- Escalate to counsel: the governing document is ambiguous, the exclusion would remove most of a requested task, or a recipient of earlier material cannot be reached.
- Proceed: every place on the map passes, the manifest carries the exclusion version, and the residual risk has a named owner.
Carrying exclusions into a SourceX licensing deal
In a SourceX deal, exclusions need to be settled before anything is delivered, because the scope is fixed in writing first. SourceX runs a rights check after completing the data inventory with the company, and each deal's Dataset Deal Confirmation fixes the exact dataset, the payout, the rights and the processing route before the opportunity is presented externally; a material change after that needs a signed amendment.
Two further terms on SourceX's how-it-works page make early exclusions matter. Exclusivity for AI training covers only the selected data, not future or unselected records, so a clean boundary around the selection protects both sides. And once a deal has closed and the fee is paid, the license is permanent and models already built from the data are not untrained or deleted, so content found in a delivery afterward cannot be pulled back out of a trained model. The agreement also asks the supplier to confirm the data has not already been licensed to anyone for AI training.
A first check involves no records at all: the company assessment asks about systems and history without any data upload, and its result is not an offer or a promise of payment.
Frequently asked questions
What are dataset exclusion transformation controls?
They are the tests and records that prove an excluded record, field or document is absent from every artifact made from the source, including shared samples, derived fields, summaries, embeddings, splits and the delivery manifest. They pair a documented rights decision with a check at each transformation.
Does deleting excluded records from the export remove them from embeddings?
No. An embedding or index computed before the exclusion still carries information about the excluded content, and research has shown embeddings can be inverted to recover much of the original text. Rebuild indexes from the clean source and compare item lists rather than patching them.
Who decides what must be excluded from a licensed dataset?
A person with authority over the records decides, on the basis of a governing document such as a customer contract, notice or confidentiality obligation, with counsel resolving any rights question. Engineering carries the decision out; it should not make it.
What should happen to a sample shared before an exclusion was decided?
Match its record keys against the exclusion list. If it holds covered records, ask the recipient to delete it, get written confirmation, and record both in the review log. Replace it only with a sample cut from the clean version.
Can redaction replace an exclusion?
No. Redaction changes what a record reveals; it does not change whether the record may be licensed for the requested use. A record excluded under a contract or a third party's rights stays excluded however thoroughly it is masked.
Sources
- ICO anonymisation guidance: pseudonymised data remains personal data for anyone holding the additional information needed to re-identify people, and identifiability depends on the context, including what the recipient could reasonably use. Source
- GDPR Article 19 requires a controller to communicate rectification, erasure or restriction under Articles 16, 17(1) and 18 to each recipient of the personal data, unless impossible or involving disproportionate effort. Source
- California Civil Code 1798.105(c) requires a business that receives a verifiable deletion request to delete the information, notify service providers and contractors to delete it, and notify third parties to whom it sold or shared it, unless impossible or involving disproportionate effort. Source
- The FTC's December 2023 Rite Aid order requires deleting photos and videos collected by the facial recognition system and data, models or algorithms derived from them, and instructing third parties that received them to delete them. Source
- Peng, Mathur and Narayanan (2021) found that after DukeMTMC and MS-Celeb-1M were retracted, their data remained widely available and continued to be used, which they call runaway data. Source
- Morris et al. (2023) showed that text embeddings can be inverted to reconstruct much of the input text with an iterative correction method. Source
- SourceX's how-it-works page: rights check after the inventory; the Dataset Deal Confirmation fixes the exact dataset, payout, rights and processing route; material changes need a signed amendment; exclusivity covers only selected data; licenses are permanent after Closing and trained models are not untrained; the agreement asks the supplier to confirm no prior AI training license. Source
See if your company qualifies
A short company assessment. No data uploads are needed.