./contribute
We do the research together
Send us a task, data or an environment. We run the experiments with you and write them up together; contributors are co-authors of the work they contributed to.
What you can bring
Tasks
A factory cell, a drawing or a control problem you know is hard, with what a correct answer looks like.
Data
PLC programs, CAD drawings, I/O lists or operating records you have the rights to share.
Environments
A simulator, a test rig or a grader we can run models against.
We agree the scope and the rights first. Data stays private unless you decide otherwise; results are shared with you before anything is published.
Main experiments so far
- 01
Industry benchmark
CAD tasks GPT-6 Astra gets 0% on
When depth is measured from the back face, GPT builds the mirror-image part. Opus mostly gets it right.
- 02
Simulator
Give models a PLC simulator and their pass rate jumps
Agents could run their own PLC code in a simulator and revise it.
- 03
Industry benchmark
A process that finds defects automatically whenever a trace is produced, and updates the next version of the benchmark
Every task defect that changed a verdict was first found by reading model traces.
- 04
Industry benchmark
Claude models cost more
Against GPT models with a similar pass rate, running the whole benchmark costs several to over a hundred times more.
- 05
Trainingrunning
Teaching Qwen 0.8B to write PLC code
Fine-tuning a small open model on PLC tasks.
- 06
Trainingrunning
Judging PLC code with a jev-family model
PLC code behaves like a time series, so we fine-tuned a jev-family model on it, with some effect so far.
Each report has the numbers, from internal runs on the private task set. Running experiments have no report yet.