Verification
What we check, and what stays declared.
Verification is the product. A dataset is not usable because someone says it is good. It is usable when its origin, rights, consent, structure and defects are known, and when the parts that remain unknown are written down instead of omitted.
Six dimensions
Each is reviewed separately, and each produces two answers: what was established, and what was not.
Provenance
Where the data came from, through every hand it passed.
What we review
Collection route, originating party, date range, and whether the chain can be evidenced rather than asserted.
What stays unverified
Who collected it and under what arrangement, where the supplier cannot produce records.
Rights
Who holds them and what they permit.
What we review
Ownership or licensing authority, permitted use, territory, duration, exclusivity, prior licensing, and whether rights extend to commercial AI training and to the licence the buyer needs.
What stays unverified
Whether an upstream agreement a supplier describes but cannot show actually says what they say it says.
Consent
What the people recorded were told and agreed to.
What we review
The consent instrument and version, the scopes it grants, whether it covers the intended use, and whether withdrawal and deletion can be traced.
What stays unverified
Whether a contributor understood the instrument, and anything a supplier asserts without producing the record.
Metadata
What is labelled and how consistently.
What we review
Field coverage, label consistency across the set, whether labels mean the same thing throughout, and whether metadata travels with the data through delivery.
What stays unverified
Accuracy of demographic self-declaration, and any field recorded only as free text.
Quality control
How defects are found before delivery.
What we review
Technical integrity, format and checksum verification, language confirmation, speaker structure, transcript-to-audio review, and what happens to material that fails.
What stays unverified
Anything outside the reviewed sample. A review of six files is evidence about six files.
Use-case fit
Whether it suits your specific workflow.
What we review
Suitability assessed separately for training, evaluation, benchmarking, ASR, TTS and voice agents. Never one score.
What stays unverified
Fit against a specification we have not been given.
Speaker structure, reviewed as its own axis
“Conversational speech” is not one thing. A set that suits a single-speaker recogniser can be the wrong shape for a two-party voice agent, and the difference is structural rather than acoustic. Where the requirement turns on it, these are established separately and reported separately.
— Number of speakers per item, established from the material rather than from the label.
— Speaker attribution: whether turns are attributed, and by what method.
— Turn structure: how turns are segmented and timed.
— Overlap: whether simultaneous speech occurs, and whether it is marked.
— Channel topology: mixed single channel, or separated per speaker.
— Diarisation status: present, absent, or asserted but not evidenced.
— Transcript attribution: whether the transcript carries the speaker labels the audio structure implies.
Diarisation and channel separation are reported only where the exact material carries them. Neither is inferred from a dataset being described as conversational.
Evidence classes
Every finding carries the class of evidence it rests on. A dataset does not get one state; each dimension gets the state its evidence actually supports, and they are reported separately.
Review scope must match the claim.
Supplier declared
Stated by the holder. No evidence reviewed. Recorded as a declaration and treated as one.
Document reviewed
A contract, consent instrument or record was produced and read. What it says is reported; what it does not cover is named.
Sample tested
Established by inspecting delivered material. The finding is evidence about the sample, and the sample size is stated with it.
Independently established
Confirmed against a source that does not depend on the holder's account, such as re-measuring files or re-hashing against a manifest.
Conditional or missing evidence
Review opened, evidence outstanding or incomplete. Named as outstanding rather than rounded up or left out.
Not reviewed
Outside the scope that was reviewed. Reported as not reviewed, not as satisfactory.
Review outcomes
The evidence classes above roll up into one outcome per dataset. Only the first two states can be presented to a buyer as reviewed supply, and any limitations travel with the offer.
Verified — full scope
Reviewed across all six dimensions with evidence produced, and the verification scope stated with the dataset. Can be presented as reviewed supply.
Verified with limitations
Reviewed and usable with named gaps. The limitations are disclosed with the offer, not on request. Can be presented as reviewed supply.
Conditional
Review incomplete or evidence outstanding. Not offered while it stands.
Not reviewed
No review has taken place. Declared information only.
Rejected
Failed review. Not offered, and not offered later without a new review.
How supplier review works
Sonexis sources only from organisations that hold or can lawfully license human data. Supplier applications are declarations until evidence is reviewed. Organisation review and dataset review are separate gates, and clearing the first does not clear the second.
Applying is not verification. An approved supplier is an organisation we are willing to review datasets from — approval of the organisation does not approve any dataset it holds. Each dataset is reviewed on its own rights, provenance, consent, metadata and quality control before it can be used.
Everything an applicant writes is a declaration. Until evidence has been reviewed, it stays a declaration and is treated as one internally.
Apply to supply dataConsent, where Sonexis collects
For collection run under a Sonexis specification, whether directly or through an approved collection partner, consent runs against a versioned instrument.
— Consent is task-specific, not a blanket platform agreement.
— The scopes granted are snapshotted onto the recording at the moment of submission, so later edits to the consent text cannot retroactively widen what a past contributor agreed to.
— Export is gated against those snapshotted scopes and fails closed.
— Withdrawal and deletion requests are tracked, and the account of what could and could not be deleted is written.
— Sensitive actions are recorded in an append-only audit log.
This describes Sonexis-controlled collection. Supplier-owned data is governed by the supplier’s own consent instrument, which is reviewed separately and reported as reviewed, not as ours.
See what this produces
A Dataset Passport records what was reviewed, what was established, and what was not. One illustrative specimen is published in full.