RL environments from owner-licensed engineering & operations work.
Engineering & CAD
Engineering work, made verifiable.
Drawings, native models, calculations and the review comments that changed them. We turn each finished job into a task an agent can run and a verifier can check.
Verifier
A design can look right and still be wrong.
The task: model a bracket from its drawing. A submission with the bore 0.4 mm oversize passes every whole-shape check. Measuring the bore catches it.
-
01
The drawing
environment/project/drawing.svg -
02
The submission

Bore Ø18.400; everything else matches the drawing.
-
03
Whole-shape checks
- Envelope
120 × 80 × 84 mm- Volume
−0.07%- Overlap
0.9993- Verdict
- Pass
A shape-only grader stops here.
-
04
Bore measurement
- Measured
Ø18.400- Drawing
Ø18.000–18.018- Reward
0- Verdict
- Fail
Sections at three depths, a circle fitted to each.
A generated demonstration task. The download has the drawing, reference solution, verifier and recorded run.
Environments
From a finished job to a reward you can train on.
01
Owner-licensed records
Drawings, revisions, calculations, email and chat from operating companies that license them to us for AI training.
02
A runnable environment
One Harbor-compatible folder per task: the request, the starting files and a container with the tools the work needs.
03
A verifier tested against reward hacks
The reward comes from the files the agent leaves behind. Before a task ships, its verifier must score reward hacks, plausible mistakes and correct alternatives as expected.
cad-bracket-001/ ├── instruction.md # the request, as an engineer sends it ├── task.toml # domain, time limit, split, rights ├── environment/ │ ├── Dockerfile # CAD kernel, network off │ └── project/drawing.svg ├── solution/solve.sh # the reference answer, reward 1 └── tests/ ├── test.sh # writes the reward ├── verify.py # 3 gates, 10 checks └── probes/ # 2 hacks, 5 mistakes, 3 alternatives
Two lines of work
Engineering drawings and operating records.
-
Engineering & CAD
Read drawings, build models and recompute the engineering, following the revisions and review comments that explain what changed.
-
Enterprise operations
Resolve quotes, reconcile records and follow amendments across email, chat and business systems, through to the recorded outcome.
Sixteen operating companies have signed agreements with us. How partnerships work
What we deliver
What a post-training team receives.
RL environments
An instruction, a container with the work’s tools, a reference solution and a verifier, in a Harbor-compatible folder.
Evaluation sets
Held-out tasks from the same work, kept out of every training split, to measure what training changed.
Licensed datasets
Drawings, native models and operating records, with a datasheet and the uses the owner permits.
Expert trajectories
Practitioners solving the same tasks step by step, as demonstrations for supervised fine-tuning.
Start with a sample pack.
Tell us the capability you are training. We send tasks you can run and verifiers you can read.
Own engineering or operating records? Become a data partner.