Applied Labs.
We build synthetic plants, engineering tasks and deterministic verifiers. They test whether an AI system can work from drawings, scans and records that disagree, and whether it says when the evidence runs out.
The approach.
On a real plant the drawings often disagree with the scans, and some records are out of date or missing. Sometimes the correct answer is that the evidence is not enough to give one.
We recreate those conditions in synthetic plants, with ground truth our engineers author from the way they work on real sites. The synthetic plants are the classroom and real plants are the private exam. Before we claim a task family is fit for training, we calibrate it against held-out evidence from real sites.
Tasks.
- 01
Reconcile conflicting evidence
The drawing, the scan and the site observation disagree. Which is current, and which discrepancy matters for this decision?
- 02
Answer a bounded feasibility question
Can the named equipment go in this bay with the footprint, clearance, access and utilities it needs? The answer cites its evidence.
- 03
Abstain when it should
When the evidence is insufficient, the right answer is Not Demonstrated or Escalation Required, and the verifier scores it as an answer.
A drawing, a scan, a photo and a record of one machine, reconciled into one verified object.
Physical AI and autonomous manufacturing.
A robot that will work in a plant needs a plant to practise in first. We build that plant as a sim-ready stage, and our verifiers score every run before the robot goes near a real site.
- 01
A stage built for the task
A synthetic plant scoped to the task you want to test, delivered in an open format your simulator can load.
- 02
Your robot, run in it
Your team runs the robot or its policy on the stage, as often as it needs, with no access to a real site.
- 03
Every run scored
Deterministic verifiers grade each run against the authored ground truth, so the same run always gets the same score.
Deliverables.
Every deliverable is synthetic and authored by our engineers. Anything derived from customer data needs a separate grant of rights.
01
Industrial decision benchmark
Tasks on our synthetic reference plant. Each answer is scored on its verdict, the evidence it cites, whether it finds the discrepancies and whether it abstains when it should.
02
Environment packages
An environment for each task family, with the evidence, the ground truth and the evaluation interface. We grant evaluation use and training use separately.
03
Verifiers
Deterministic checks that grade a verdict, its citations or a bounded plant artefact against the authored ground truth. The same answer always gets the same score.
04
Expert-reviewed trajectories
Rubrics and reasoning traces on synthetic decision cases, written and reviewed by our engineers.
05
Calibration reports
A measurement of how far a synthetic task family is from the real decision pattern on the private exam. We issue it before we claim the family is fit for training.
06
Sim-ready plant stages
A synthetic plant stage, scoped to one task, for a robotics or physical-AI evaluation we have agreed with you. It comes in an open format.
We add physics task families, with ground truth from a solver, as each facility-physics domain comes in. Each proposal states which families are available.
Questions.
Can we test a robot before it goes on site?
Yes. We build a sim-ready stage of a synthetic plant for the task you want to test, in an open format. Your team runs the robot or its policy in it, and our verifiers score each run against the authored ground truth before any real site is involved.
How is synthetic performance checked?
We compare each task family against held-out evidence from real sites. We report a synthetic score as a synthetic score and state how far it is from the real decision pattern.
What comes next?
Next is a constraint-based plant generator, so we can produce synthetic plants in volume. Physics task families follow as each facility-physics domain comes in.
Start an evaluation.
If your team works on frontier or physical AI, we build tests from the way an engineer makes decisions on a plant.