[ Human-authored training data ]

Human expertise,
parsed for machines.

Datasets, reasoning traces, RL environments and evals — written from scratch by senior engineers, calibrated to your model’s pass rate.

Request a free sample → See a redacted task ↓
>NDA · lab-calibrated · never resold
  • 01Fully human-authored
  • 02Calibrated to your pass rate
  • 03Multi-layer QA
  • 04Delivered under NDA
//01Calibration[ target pass@1 · 0.10–0.30 ]

Hard enough to teach.
Solvable enough to learn from.

Every task is calibrated for an optimal 10–30% first-pass success rate on your target models, set where public benchmarks stop being useful. Too easy teaches nothing. Unsolvable teaches noise.

Calibration run · Batch 07pass@1 · k = 32 · vs ██████
Accepted / authored
Yield
36.6%
Band
0.10–0.30
Model
██████

    Live histogram of authored tasks by frontier-model pass rate. Tasks between 10% and 30% pass@1 are accepted; above 80% they are rejected as too easy; at 0% as not solvable; everything else goes back to its author. 142 of 388 tasks accepted in this batch.

    //02What we author[ SFT · RL · Evals ]

    Four kinds of data. All written by hand.

    Every datapoint is created from scratch by senior practitioners: primarily senior software engineers, alongside vetted domain specialists. Never scraped, never synthetic.

    //A

    Datasets

    Real-world engineering tasks with reliable ground truth, written from scratch. Never scraped.

    >SFT>Evals

    //B

    Reasoning traces

    Step-by-step solutions from senior engineers, dead ends included, so a model learns how the answer was found.

    >SFT>Process rewards

    //C

    RL environments

    Executable tasks with verifiable rewards and rubrics, tuned to your agent’s pass rate.

    >RL>Verifiable rewards

    //D

    Expert evaluations

    Private, expert-graded benchmarks for the places where public ones saturate.

    >Evals>Rubrics

    //03Inspect a task[ redacted ]

    This is what you buy.

    A redacted task, the expert’s reasoning, and the QA it passed. Redacted, because yours will be too.

    t-0413 / redacted.view
    Drafting
    Task
    Rust · Concurrency

    Shutdown deadlocks a connection pool under load

    Make Pool::shutdown() in the service wait for every borrowed connection, without holding the lock across an .await.

    Reasoning trace
    1. Reproduce: shutdown() hangs in 1 of 4 runs.
    2. Cause: drain() awaits while holding the lock.
    3. Revised after review: bind PooledConn to the pool lifetime; release in Drop.
    4. Verify: loom finds 0 deadlocks; 17 of 17 tests pass.
    1. Automated checks
    2. Two independent reviews
    3. Expert solve
    4. pass@1 0.18, in band
    //04How an engagement works[ pilot → scale ]

    Start with a pilot batch.

    Typically 10–50 tasks to align on research goals, target difficulty and format, and to validate clarity, correctness and model pass rates before scaling.

    1. S1

      Pilot

      10–50 tasks

    2. S2

      Specify

      schema · targets · QA bars

    3. S3

      Author

      senior engineers · NDA

    4. S4

      QA & calibrate

      pass@1 → 0.10–0.30

    5. S5

      Deliver

      JSONL · Parquet · envs

    //05Guarantees[ in writing ]

    What we put in the contract.

    //06Why it matters[ memo ]

    Models learn what humans write down.

    Scraped text runs out at the frontier, and synthetic data teaches a model its own mistakes. So every task we ship is written from scratch by a senior engineer, reviewed twice, solved independently by a domain expert, and calibrated against your model before it leaves our hands.