Expert network
Experts, verified and re-tested.
Practising physicians, staff engineers, chartered accountants and commercial counsel author training data and grade model output. Under 2% of applicants are admitted, and every expert is re-calibrated weekly against seeded gold sets.
Last admissions cycle · illustrative
01 · Vetting
Four stages of vetting
A CV gets an applicant to the first stage. Everything after it is confirmed with an issuer, timed, or paid for, and the last stage continues for as long as the expert stays.
Verified
With the issuing body
Licences, registrations and degrees are confirmed with the body that granted them, not read off a profile. Most applicants stop here.
Examined
Timed · set by the network
A timed exam written and marked by practitioners already in the network, pitched at the level of a senior colleague rather than a textbook.
Trialled
Paid · real tasks · blind-graded
Candidates do real work on a live project and are paid for it. Two existing experts grade the output blind against gold sets.
Re-calibrated
Weekly · for as long as they stay
Every expert grades a seeded gold set each week. Drift below threshold triggers retraining; persistent drift means removal from the network.
02 · What they produce
Author the training data, then evaluate the model
The same experts work at both ends of the pipeline, so the rubric a model is judged against was written by someone who also wrote the training data.
Author
Training data written by practitioners, with the reasoning recorded alongside the answer.
Evaluate
Professional judgement applied where string matching cannot separate correct from plausible.
03 · Coverage
Six verticals, staffed by practising specialists
Roles, not names. Everyone in the network practises now, or was recently senior, in the field they grade.
04 · Engagement
Three engagement models
Dedicated team
Named experts · your tooling · months
A named group of experts assigned to your lab, working inside your annotation and evaluation tooling under your rubrics, with a DataOrigin lead.
Sprint pod
One deliverable · weeks
A gold set, a rubric, a red-team report. We staff the pod, run calibration, and hand over a finished, audited deliverable.
Grading bench
On demand · graded runs · days
Submit model outputs or environment runs. A calibrated panel grades them and returns scores with written reasoning per item.