Expert data

Data from people who know the work.

Two blind expert solves and a sealed review on every task, and construction drawings labeled by a concrete estimator.

What experts produce

Judgment a model can't fake.

  • Blind solves

    Two experts solve each task without the answer or each other's work. When they disagree, a third, senior reviewer decides.

  • Sealed reviews

    Each verdict is committed before any is revealed, so no reviewer can follow another.

  • Labeled drawings

    Construction takeoffs labeled by hand: footings, slabs and walls, measured from the drawings.

    74,750 sq ft hand-labeled by our founder

  • Provenance on every task

    Each solve is recorded with the expert's consent. The task, each recording and each verdict are hashed into the task's manifest.

    2480e8…44ae6a

How it works

From a blind solve to a sealed verdict.

  1. 01

    Two experts solve it blind

    Each works in a recorded browser, from the prompt alone, without the reference answer.

  2. 02

    Reviews stay sealed

    Each reviewer commits a verdict before anyone's is revealed. Agreement decides; a disagreement goes to a senior reviewer.

  3. 03

    It ships with its manifest

    Checks, solvers, reviews and the hashes of the task and each recording, in one file per task.

Who solves

Native speakers and domain experts.

Native speakers of languages AI labs under-cover, starting with Polish.

Construction estimators, starting in the US.

Our expert waitlist is open. Expert opportunities

Tell us what your models get wrong.