Rights and contracts
Dataset deletion evidence: what a buyer can show and what to ask
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
A buyer can support a dataset deletion confirmation with a signed statement scoped to the delivered files, records from each storage location, the dates soft-deleted copies, key deletions and backups stop being recoverable, and specific answers to verification questions. No artifact proves no copy exists anywhere, so judge whether the evidence covers every location and shows completed deletion rather than a step toward it.
License terms
Dataset deletion evidence
Key takeaways
- Tie every piece of evidence to the delivery IDs and file hashes signed at receipt.
- A delete on a versioned S3 bucket, a soft-deleted object or a key pending deletion can all still be recoverable.
- Date the deletion when recoverability ends, not when the delete action ran.
- Key destruction makes encrypted copies unreadable but not decrypted working copies.
- Strong answers name a location, method, date and person; weak ones restate the obligation.
What dataset deletion evidence can a buyer realistically provide?
A buyer can provide four kinds of dataset deletion evidence: a signed statement scoped to the delivered files, system records from each place the data was stored, the dates on which recoverable copies such as soft-deleted objects and backups expire, and answers to specific verification questions. None of these proves that no copy exists anywhere. Together they show that named locations were dealt with, by whom and when.
The supplier's job is to judge whether the evidence covers every location and whether each item actually shows deletion rather than a step toward it. That second point is where most packets fall short. A storage console that no longer lists a file, a key marked for deletion, or a delete request on a versioned bucket can all look final while the data is still recoverable.
This page is about evaluating what the buyer sends. The contract clauses that require it and the shape of a destruction certificate are covered on the pages linked at the end; here the question is what each artifact can and cannot show.
Scale the review to the package. A de-identified set of order exceptions held in one bucket needs a short packet and a quick read; a delivery that spread across a training cluster, an annotation vendor and several notebooks needs every row of the matrix below. Decide which level applies when the deletion request goes out, so the buyer knows what to prepare.
Which evidence does each storage location produce?
Each place a copy can live produces its own kind of record, and each has a trap that makes a deletion look complete before it is. The table is the core of the guide: for every location, the artifact to ask for, what it shows and the reason it might not mean what it seems.
The platform behaviors cited come from each vendor's documentation as checked on the date of this article. Settings vary by account, so ask the buyer for the configuration in force, not the default.
| Copy location | Artifact to ask for | What it shows | Trap to check |
|---|---|---|---|
| Amazon S3 bucket with versioning | Delete requests with version IDs, or lifecycle rules expiring noncurrent versions, plus a listing of versions after deletion | Specific object versions were permanently removed | A delete without a version ID only adds a delete marker; earlier versions remain |
| Any S3 bucket | CloudTrail data event records for the delete operations | Who deleted which objects and when | Data events are not logged unless configured, so the log may simply not exist |
| Google Cloud Storage bucket | Bucket soft delete policy and the date the retention period ends | When deleted objects stop being restorable | Soft delete is on by default for 7 days and can be set up to 90 |
| Azure Blob storage | Soft delete retention setting and deletion timestamps | When deleted blobs stop being recoverable | Retention can be anywhere from 1 to 365 days |
| Data encrypted under an AWS KMS key | ScheduleKeyDeletion record and the completion date | The key, and so the encrypted copies, become unusable | Deletion waits 7 to 30 days and can be cancelled in that window |
| Data encrypted under a Cloud KMS key | Key version state history | The key version moved to destroyed | Default 30 days scheduled for destruction, restorable until then |
| Drives, laptops and on-premises disks | Sanitization record naming device, method and operator | The method used, such as clear, purge or destroy | A clear-level overwrite is weaker than purge or destroy |
| Derived files: tokenized sets, indexes, notebooks | Inventory by delivery ID and a signed statement | The buyer looked for derived copies, not only originals | Derived files do not share the original hashes |
| Backups | Backup rotation schedule and the last date a backup containing the data expires | When the last restorable copy ages out | Backups are often excluded from immediate deletion |
How do you tie the evidence to what was actually delivered?
Tie every piece of evidence to the delivery record: the delivery ID, the dataset version and the list of files with their hashes that both sides signed at receipt. A deletion statement that says all data from the agreement has been deleted is far weaker than one that lists the delivery IDs and attaches the file list it covers.
Originals are the easy part, because their hashes are known. Ask the buyer to hash any surviving files in its staging and archive locations against the delivered list before deletion, which shows the originals were found where they were expected. Then ask for the derived copies by name: converted formats, tokenized or chunked versions, embeddings and vector indexes, evaluation subsets, and files in experiment-tracking tools. These have different hashes, so the only way to find them is an inventory the buyer keeps by delivery ID or dataset tag.
If the buyer cannot produce that inventory, treat the deletion statement as covering originals only, and say so in writing. That is not an accusation; it is an accurate record of what the evidence supports.
Which verification questions separate strong answers from weak ones?
Use verification questions to test the packet, and judge the answers by how specific they are. A strong answer names a location, a method, a date and a person. A weak answer restates the obligation. The guide below lists the questions worth asking every time, with what each kind of answer looks like.
Send the questions with the deletion request, not after the packet arrives. Buyers prepare better evidence when they know what will be asked, and the answers then arrive with the packet rather than as a second round.
Read the answers as a set. A buyer that lists five storage locations but describes deletion for only three has told you where to look next. A buyer that answers every question in one sentence each, with no names or dates, has usually not yet done the inventory, and a short follow-up asking for it is fairer than a dispute.
| Question | Strong answer | Weak answer | Follow-up if weak |
|---|---|---|---|
| Where did copies of these deliveries live? | A list of buckets, servers, drives and tools by delivery ID | Our secure environment | Ask for the inventory or the access log of the landing location |
| How were versioned or soft-deleted copies handled? | Versions deleted by ID or expired by rule; soft delete end date stated | Files were deleted | Ask for the bucket versioning and soft delete settings |
| Were any copies encrypted under a key you control? | Key ID, deletion scheduled date and completion date | Everything is encrypted | Ask whether decrypted working copies existed and where |
| Which derived files were made, and what happened to them? | Named derived sets with locations and deletion dates | Only the original files were used | Ask about tokenized sets, indexes and evaluation subsets |
| When will the last backup containing the data expire? | A date from the rotation schedule | Backups are handled by IT | Ask for the retention period of the relevant backup system |
| Who performed and who verified the deletion? | Two named people with roles | The team | Ask for a named accountable signer |
| Is anything retained, and on what basis? | Specific items, the legal basis and the retention end date | Nothing | Ask directly about legal holds and records retention duties |
| Were models trained on the data, and what does the agreement say about them? | Named models and the clause that governs them | Not applicable | Check the agreement; model rights are a separate term |
Why is deletion a process with waiting periods, not a moment?
Deletion on most platforms is a process with a waiting period, so the date a buyer clicks delete is rarely the date the data stops being recoverable. Ask for both dates, and treat the later one as the deletion date for your records.
The windows are not small. Cloud Storage soft delete is on by default with 7 days of retention and can be set as long as 90. Azure Blob soft delete retention can be set from 1 to 365 days. An AWS KMS key scheduled for deletion waits 7 to 30 days, 30 by default, and the deletion can be cancelled during that time. A Cloud KMS key version is scheduled for destruction for 30 days by default, and Google notes key material can remain in its systems for up to 45 days after the scheduled destruction time.
This has a practical consequence for the certificate. Either date it after the last window closes, or have it state both the action date and the date recoverability ends for each location. A certificate signed on the day of the delete request, with a 30-day key window still open, certifies an intention rather than a result.
The windows also cut the other way. A buyer that deletes the moment a license ends, with soft delete switched off, loses the chance to recover from a mistaken deletion of unrelated data. Asking for the configured windows rather than for instant deletion keeps the request reasonable and the evidence honest.
What does cryptographic erase add, and where does it stop?
Cryptographic erase adds a clean way to make every encrypted copy unreadable at once, and it stops at any copy that was decrypted. NIST defines cryptographic erase as a purge technique in which the keys protecting the encrypted data are sanitized, and SP 800-88 Revision 2 notes that for virtual or logical storage there may be no other purge option.
For licensed data that sits in cloud storage, that makes key destruction the strongest technical evidence available, provided the key protected the data and no copy existed outside it. Ask for the key ID, show that the delivered data was encrypted under it, and confirm the destruction completed after any waiting window.
Its limit is the working copy. Training pipelines, notebooks and preprocessing jobs usually read decrypted data and may write new files that are encrypted under different keys or not at all. Those copies need their own line in the evidence. The strongest position of all is when the supplier controls the key from the start, which is covered in the page on verifying deletion yourself.
Physical media follow the same logic with different tools. For a returned or wiped drive, ask which SP 800-88 method was applied: clear, purge or destroy. A clear-level overwrite suits low-risk media reused inside the same organization; for drives that held licensed records and are leaving the buyer's control, purge or destroy is the level to expect.
Illustrative worked example: reviewing a close-out packet
Illustrative example, not a real company or transaction. An evaluation license with an AI developer ends, and the supplier receives a close-out packet covering one delivery of support tickets. The supplier assesses each item against the questions above.
The result is not a rejection. It is a short list of items the buyer can supply within days, after which the supplier records the deletion as complete with the later recoverability date. The exception in this example is the legal hold: the buyer retains one copy of a subset for a stated legal reason, which the supplier records with its end date rather than disputing it, after checking with counsel.
| Item received | What it shows | Assessment | Action |
|---|---|---|---|
| Signed statement listing delivery DLV-2026-014 and its file list | Accountable signer, scoped to the delivery | Strong | Accept |
| Listing of the S3 landing bucket showing no objects | Current versions are gone | Incomplete: bucket is versioned | Ask for version listing or lifecycle rule |
| KMS key deletion scheduled on day 0 with a 30-day window | Encrypted copies will be unreadable on day 30 | Strong once complete | Record day 30 as the date |
| Note that a tokenized training set was deleted from a shared drive | A derived copy was found and removed | Weak: no location or date | Ask for path, date and who deleted it |
| Backup rotation of 35 days | Last backup expires on day 35 | Acceptable if dated | Record day 35 as the final date |
| One subset retained under a legal hold | A known retained copy with a basis | Acceptable with end date | Record the hold, basis and review date |
Which evidence gaps belong with counsel rather than the data team?
Stop and bring in counsel, privacy or security review when the evidence gap is about rights or law rather than paperwork. The deletion request itself rests on the agreement, so an argument about scope belongs with the people who negotiated it.
Where the buyer acted as a processor under the GDPR, the Commission's standard contractual clauses show the expected form: at the controller's choice, delete all personal data and certify that it has been done, unless law requires storage. Many data licenses make the buyer a controller instead, so whether that model applies is a question for counsel.
- The buyer refuses to name storage locations or the people who performed deletion.
- Personal data was in the delivery and the buyer's role as controller or processor is unclear.
- The buyer claims a legal hold or retention duty without stating its basis or end date.
- Trained models, fine-tuned adapters or vector indexes are involved and the agreement is silent on them.
- Evidence suggests copies reached third parties such as annotation vendors or inference providers.
Writing deletion evidence into a SourceX license
Deletion and retention expectations belong in the signed agreement, since that document also bounds what SourceX may do with the de-identified dataset afterward, sublicensing and distribution included, and the record owner signs off on its scope and terms before any delivery. The verification questions above are how those expectations get checked later. Across the deal, SourceX handles sourcing, rights review, contracting, preparation, delivery and payment.
The opening step is a short company assessment; no data has to be uploaded to learn whether the records might qualify.
Frequently asked questions
What evidence can support a buyer's dataset deletion confirmation?
A signed statement scoped to the delivered files, system records from each storage location, the dates soft-deleted copies and backups stop being recoverable, and specific answers to verification questions. Together they show that named locations were handled; none proves no copy exists anywhere.
Is a deleted S3 object really gone?
Not always. In a versioning-enabled bucket, a delete request without a version ID only adds a delete marker, and earlier versions remain. Ask for evidence that versions were deleted by ID or expired by a lifecycle rule.
When should a deletion certificate be dated?
After the last waiting window closes, or with both dates shown for each location. Soft delete periods, key deletion waiting periods and backup rotation all keep data recoverable after the delete action.
Does destroying the encryption key count as deletion?
NIST treats cryptographic erase as a purge technique, so destroying the key makes data encrypted under it unreadable. It does not reach copies that were decrypted and stored elsewhere, which still need their own deletion evidence.
Can a buyer keep some data after a deletion request?
Sometimes, for example under a legal hold or a records retention duty, if the agreement allows it. Ask for the specific items, the basis and the end date, and have counsel review whether the retention is permitted.
Sources
- NIST SP 800-88 Revision 2 (September 2025) sets out media sanitization methods, including clear, purge and destroy, and treats cryptographic erase as a purge technique. Source
- NIST defines cryptographic erase as a purge technique in which the keys protecting encrypted data are sanitized. Source
- AWS KMS key deletion is scheduled with a waiting period of 7 to 30 days (30 by default), can be cancelled during that period, and the key cannot be used while deletion is pending. Source
- In a versioning-enabled S3 bucket, a delete request without a version ID only adds a delete marker; earlier versions remain until deleted by version ID or by a lifecycle rule. Source
- Cloud Storage soft delete is on by default with a 7-day retention, configurable from 7 to 90 days, and soft-deleted objects can be restored during that period. Source
- Azure Blob soft delete keeps deleted data for a retention period of 1 to 365 days, during which it can be recovered. Source
- Cloud KMS key versions are first scheduled for destruction, 30 days by default, and can be restored in that period; key material can remain in Google systems for up to 45 days after the scheduled destruction time. Source
- AWS CloudTrail does not log data events, such as S3 object-level operations, unless they are added to a trail or event data store. Source
- The European Commission's standard contractual clauses for controllers and processors (Decision 2021/915) require the processor, at the controller's choice, to delete all personal data and certify that it has done so, unless law requires storage. Source
See if your company qualifies
A short company assessment. No data uploads are needed.