Every unit of training data has a person behind it. Someone wrote the demonstration, graded the preference pair, red-teamed the response. The entire value of expert human data rests on one question: was that person who the vendor said they were?
The industry's standard answer is a resume, a profile, and sometimes a webcam selfie at sign-up. None of these establishes the fact that matters. A resume is a claim. A profile is a claim. Even a passed qualification test only proves that somebody passed it once — not who, and not who showed up for the work afterward.
The four facts that matter
When we say an expert is verified at source, we mean four specific properties have been established before their first task, and travel with every unit of work they produce:
- Identity — the person doing the work is who they claim to be, confirmed against authoritative records, not a profile photo.
- Employer — where they actually work, verified at source rather than read off a resume.
- Role — what they actually do: the real function, not the title typed into a form.
- Seniority — how senior they actually are, established against systems of record, not self-attestation.
Each of the four closes a different fraud path. Identity alone stops account sharing, but not credential inflation. Employer alone stops the fictional hospital, but not the receptionist presented as a physician. Role and seniority are what turn "works in healthcare" into "practicing ICU nurse, eight years" — and that difference is exactly the difference buyers are paying for.
Claims, documents, and systems of record
It helps to think of vetting methods as a ladder with three rungs. The bottom rung is self-report: whatever the contributor typed into a form. The middle rung is document collection: ask for a diploma, a license PDF, an employment letter — then take the documents at face value. The top rung is the system of record: go to the registry, the directory, the authority that would know, and confirm the fact where it lives.
Most of the industry operates on the bottom two rungs, and the middle rung is weaker than it looks. Documents are supplied by the person being vetted — which means the check inherits every incentive problem it was meant to solve. A system-of-record check inverts the direction of trust: the fact is confirmed by a party with no stake in the contributor getting hired.
A resume is a claim. A system of record is a fact.
Systems of record exist for more professions than most people expect. Medical licenses live in public registries. Many jurisdictions register engineers, electricians, aviation technicians, lawyers, and accountants. Employment and role can be confirmed at the employer itself. The work is operationally heavy — it does not scale like a sign-up form — but nothing about it is impossible. It is simply a cost most vendors have chosen not to pay.
Verification is an event. Provenance is a property.
A one-time check answers who signed up. It does not answer who did the work last Tuesday. That second question is provenance, and it is where verification either becomes real or quietly evaporates.
Provenance means every unit of data carries a link back to a named, verified expert — with the role and seniority they were verified under, and the QA trail behind the work. When provenance is attached per unit, an auditor can pick any row in the dataset and walk it back to a person. When it is attached per batch, or per project, or not at all, the verification happened somewhere — but the data can no longer prove it benefited.
Why this is suddenly urgent
Frontier models have exhausted what generalist raters can teach them. The work moving through human-data pipelines now is clinical rubrics, financial-model audits, legal citation review — tasks where the grader's actual expertise is the whole product, and where a confident non-expert produces labels that look clean and teach the model something wrong.
At the same time, remote work has made impersonation cheap. The person on the qualification call and the person doing the tasks need not be the same person. Accounts are shared, resold, and subcontracted. None of this is exotic; it is the default failure mode of any marketplace that pays for expertise it cannot see.
The conclusion we drew is the one this company is built on: verification cannot be a screening step at the front of the funnel. It has to be the spine of the pipeline — done at source, attached to every unit, and standing behind every dataset we ship. The next post in this series turns that standard into a checklist any buyer can put in front of any vendor, including us.
Expert human data, verified at source.