Datasets
Real-world engineering tasks, each with a ground-truth answer.
>Fine-tuning (SFT)>Evals
[ Human-authored training data ]
Human expertise,
parsed for machines.
Custom datasets, reasoning traces, RL environments and evals, written from scratch by senior engineers and domain specialists, and tuned so your model solves only 10–30% of them.
- 01Fully human-authored
- 02Tuned to your model
- 03Multi-layer QA
- 04Delivered under NDA
We make tasks at just the right difficulty for your model to learn from.
Every task is tuned until your model solves it only 10–30% of the time. A task your model always solves teaches it nothing, and one it never solves gives it no signal.
Example run · simulated
Each column counts tasks by how often your model solves them. Only the green ones ship to you.
Ready to ship / written
142 of 388
36.6% of written tasks ship. The rest go back for a rewrite or are dropped.
Example histogram of tasks by how often your model solves them. Tasks it solves 10–30% of the time ship. At 0%, or at 80% and above, they are dropped. Everything else goes back to its author for a rewrite. 142 of 388 tasks ship in this example.
What we author
Training and eval data, written by hand.
Every datapoint is created from scratch by senior engineers and domain specialists. Never scraped, never synthetic.
Reasoning traces
Step-by-step solutions from senior engineers, dead ends included, so a model learns how the answer was found.
>Fine-tuning (SFT)>Step-level rewards (PRM)
RL environments
Runnable tasks that score the agent’s work automatically, with rubrics where a test can’t decide.
>Reinforcement learning>Verifiable rewards
Expert evaluations
Private benchmarks, graded by experts, for places where public benchmarks are saturated.
>Private benchmarks>Grading rubrics
Quality control
This is what you buy.
One example task, tested on a frontier model.
T-0413
✓Accepted17 of 17 tests pass2 reviews
Fix a Rust connection pool that deadlocks on shutdown
Solve rate
This task50 tries18%in target (target 10–30%)
Typical public taskExample90%too easy
Every task passes three layers of QA
Automated checks
The task is fair to the agent, and your model solves it only 10–30% of the time on the first try.
Two independent reviews
Two senior engineers or domain specialists review it separately for clarity, correctness, edge cases, originality and difficulty. Both must approve.
Final acceptance test
The last step before delivery. Only a task that passes it is delivered.
How an engagement works
Start with a pilot batch.
Usually 10–50 tasks. We use it to settle goals, format and difficulty before scaling up.
-
S1
Scope
Scoped on what you’re training for and what model and difficulty to calibrate against.
-
S2
Specify
We settle the schema and QA thresholds with your team.
-
S3
Author
Senior engineers and domain specialists write each task from scratch, under NDA and work-for-hire agreements.
-
S4
QA & calibrate
The three QA layers described above, run against your model.
-
S5
Deliver
In your schema (JSONL, Parquet or RL environments), after a final acceptance test.
Guarantees
What every engagement includes.
- 10 to 30 percentSolve rate we tune each task to
Your model solves it 1–3 times in 10.
- 3 layersMulti-layer QA
Automated checks, two reviews, an acceptance test.
- 100 percentOriginal and copyright-clean
Written fresh for your project, and never resold.
- 48 hoursA free sample task
On us, sent under NDA within 48 hours.
Library
See the work for yourself.
Why it matters
Models learn what humans write down.
A frontier model has already seen most public text, and training it on its own output repeats its mistakes. New tasks, written by senior engineers and domain specialists, give it material it hasn’t seen.