What is expert data?
Glossary · Updated November 16, 2026
Expert data is the training and evaluation material that domain experts produce for AI models: answers, labels, rubrics, graded outputs and worked tasks from people qualified in a field, rather than crowd labels. Mercor, the largest marketplace for it, pays contractors with domain expertise in fields such as physics and finance to answer questions and create specialized data used to train models (Dealroom, 6 July 2026), and its 10 November 2026 round priced the company at $20 billion. In healthcare the experts are clinicians in the loop. What expert data does not carry is the provenance of the record the expert judged: a label is only as good as the record it was made on, and the Data Trust Index (DTI) scores that record 0 to 100 before anyone labels it.
The phrase comes from the trade press around Mercor: "Mercor doubles to $2B gross revenue run rate as AI labs buy expert data" (Dealroom, 6 July 2026). Mercor describes the work as evaluation and post-training, with environments where agents learn complex workflows (Mercor, 10 November 2026). MDisrupt, a health expert network, argues healthcare needs its own version built around clinicians and health experts in the loop (11 November 2026).
The record layer under expert data is SuperTruth's part. In DTI, provenance is the heaviest of the eight dimensions at 25 percent of the default weight, so a record that cannot show where it came from scores at most 75 even when everything else about it is perfect: Silver, not Gold. An expert's label on that record inherits the same ceiling. The guide Clinicians in the loop: which records need one covers which records should reach an expert and what they carry when they do.
What counts as expert data
- Answers and worked tasks from people qualified in a field, used to train a model after its first training run (post-training).
- Evaluations: an expert grades a model's output against a rubric, often one the experts wrote.
- Labels on records: in healthcare, a clinician marks what a note, a result or an image shows, and which cases a model gets wrong.
- Environments: tasks and tools in which an agent is tested on a real workflow.
The categories are drawn from Mercor's own description of its work (10 November 2026) and Dealroom's report (6 July 2026), both read 16 November 2026. Mercor and MDisrupt are named so you can place the term; SuperTruth is not affiliated with either and neither has endorsed it.
What expert data lacks, and what the record layer adds
An expert judgment has an author, a credential and a date. The record it was made on may have none of those: no documented source, no consent for the use, a value from 2019, two sources that disagree. The judgment is then precise about an input nobody can vouch for. DTI scores the record first (provenance, consent, recency, quality, concordance, validation, breadth, stability) and keeps the score with it, so an expert label can say what record it was made on and how much that record could be trusted at the time.
A worked example in DTI's own terms: provenance carries 25 of 100 points, so a record with no provenance at all scores no higher than 75, Silver, however clean its values. A label on it is a label on a Silver record. The tiers are SuperTruth's names (DTI white paper, DOI 10.5281/zenodo.19601616, 2026).
Questions people ask
Is expert data the same as data labeling?
No. Labeling is one kind of it. Expert data also covers evaluations, rubrics, worked tasks and agent environments produced by people with a credential in the field, which is why its buyers are AI labs doing post-training rather than teams assembling a dataset.
Does expert data carry provenance?
The judgment does; the record under it usually does not. A clinician's label says who labeled and when, not where the record came from or whether consent covers the use. The Data Trust Index scores the record for that before anyone labels it.
Where this runs