Pristine Dataset

Frontier data sourcing · independent

The data that trains the next capability is not on the internet.

The web kept the finished answers and threw away the reasoning that produced them. We find the people who still hold that reasoning, verify they can do the work, and turn it into training data your model can learn from.

Brief us on a gap Supply rare data
Scarcity axis Select a sample

Rate bands are publicly reported market ranges for data work, from commodity labelling to board-certified generation. Context, not a quote.

What we supply

Three forms of the same scarce thing: a specialist’s judgment, written down.

Form 01

Expert trajectories

Full problem-solving traces from practitioners: the dead ends, the discarded hypothesis, the reason they rejected the plausible answer. Not the polished write-up the web already has.

Produced by
Verified practitioners working in the domain now
Signed
Every item traceable to the person who wrote it
Typical use
Mid-training, SFT, reasoning distillation
Form 02

Environments and rubrics

Interactive tasks an agent can attempt, plus the grading criteria that decide whether it actually succeeded. The environment is the cheap half. The rubric is what you are buying.

Produced by
Domain experts paired with environment engineers
Hardened for
Reward hacking, stale flows, ambiguous pass conditions
Typical use
RL, agent evaluation, capability regression suites
Form 03

Licensed proprietary corpora

Data that exists but was never published: internal decision records, QA histories, case files, workflow captures. We source the holder, negotiate the licence, and clear it before a single file moves.

Sourced from
Operating companies, practices, labs, agencies
Cleared for
Consent, PII, third-party rights, retention terms
Typical use
Domain pre-training, evaluation sets, cold-start RL

How a program runs

Five steps. You see a sample before there is volume.

  1. 01

    Scope the gap

    You name the capability and what counts as correct. If we cannot state the failure your model is making today, there is nothing worth collecting yet.

  2. 02

    Map the supply

    We work out who can actually produce and grade the work, how many exist, where they are, and what it costs to reach them. You get the map before you get an invoice.

  3. 03

    Vet on evidence

    Credentials checked against a source, work samples scored blind by a senior in the same field. A claim on a CV is not evidence.

  4. 04

    Build and grade

    Production runs with a second expert grading behind the first. Disagreements are kept, not averaged away, because the disagreement is often the signal.

  5. 05

    Deliver and re-cut

    You review a representative sample, run your own error analysis, and we re-cut the spec against it. Volume follows the sample, never the other way round.

Provenance

Training data is a liability until you can say where it came from.

Every item we deliver carries a record: who produced it, when, under what instruction, from which sources, and under which licence. If you cannot answer those five questions about a corpus, do not train on it.

Independent
No model-lab equity, no shared ownership, no parent company competing with you. Neutrality is the product.
Consent first
Licence and consent in writing before collection begins, with the intended training use stated plainly to whoever supplies it.
Segmented staffing
Experts on your program are not simultaneously on a competitor’s. Specs, prompts, and model behaviour stay inside one program.
Handling
PII removed at intake, isolated pipelines, access limited to the working team, retention on your terms, deletion honoured on request.

For labs and model teams

Brief us on a gap

Describe the capability your model is failing at. You get a written supply map back: who can produce the data, roughly how many of them exist, and what a first batch would take. No call required to get it.

For experts and companies

Get paid for what you know

Remote, part-time, asynchronous. Specialists write and grade work in their own field. Companies licence workflow data that never left the building. Published rates for expert data work run from roughly $40 to $300 an hour depending on the field and the verification required.