PRISM · THE ROBOTICS OPPORTUNITY
Know what made
the robot better.
The Light Company helps physical AI labs get reliable model evaluations. Prism projects each setup onto the work surface and uses cameras to verify it.
10 paid units installed
3 customers
Enact · New Theory · Trossen Robotics
01 / THE MISSING CONTROL
Same code. Different physical world.A few centimeters
can change the answer.
A team trains a new robot model. The result improves. But did the model get better—or did someone place the cup in an easier spot?
When operators and starting positions change between runs, teams keep repeating trials just to find a useful signal.
MAKE THE IMPROVEMENT VISIBLE
The model change stays the same.
Move the setup control.
A fixed model improvement, with changing setup variation. These dots explain the principle; they are not customer measurements.
The results overlap. The improvement is hard to see.
02 / SOFTWARE REACHES THE SURFACE
See the physical loopThe test starts
before the robot moves.
Test code can specify a position. It cannot make an operator put the cup there. Prism closes that gap with projected guidance and camera verification, connected to the lab’s code through our SDK.
- 01
Define the case.
Upload placements through the SDK, or generate a set for the task.
- 02
Project. Place. Verify.
The instructions appear on the table. Cameras check the actual physical setup.
- 03
Run the robot.
Clear the guidance before the evaluation. Compare models against the same starting conditions.
- 04
Connect the result.
Link the outcome from the lab’s evaluator or Prism’s scoring workflow to the task, model version, and verified setup.
Perception, calibration, projected guidance, camera verification, placement generation, and the SDK. Each layer has to agree about the same object on the same surface.
Compare models. Recreate the same conditions to measure a change.
Explore capability. Vary conditions deliberately to find weaknesses.
03 / WHAT CHANGED IN THE LAB
Field results · Updated September 2026Fewer trials.
A clearer answer.
ENACT · ROBOTICS LAB
Engineering feedback
in about one-tenth
the evaluations.
Prism guides and verifies the starting conditions. Engineers can compare model changes with far less repetition.
evaluations for useful engineering feedback
using the 300-to-30 comparison
NEW THEORY · 30-DAY MEASUREMENT
Evaluation variability
fell from 40% to 3%.
Consistent physical starting conditions make it easier to see whether the next model actually improved.
in evaluation variability
Two customer workflows, two measures of improvement. Enact’s earlier process took roughly 300–400 evaluations; the 10× comparison uses 300. New Theory’s 40% and 3% describe evaluation variability, not model accuracy.
04 / THE DATA OPPORTUNITY
From “did it work?” to “where does it fail?”Find a failure once.
Make it useful again.
A single success rate hides the useful detail. A failure tied to a verified physical setup becomes a case a team can run again—across stations, after a model update, and when deciding what to investigate next.
FOLLOW ONE PHYSICAL TEST CASE
A better average can hide
a new weakness.
Choose a position. Arrow keys move across the table. Compare the same case in both models.
FROM A RESULT TO A REUSABLE CASE
Keep the conditions.
Test the next model.
A failure becomes a regression test.
Recreate this setup on a compatible, calibrated station. Run the next model against it and check whether the weakness returns.
WHAT TRAVELS WITH THE CASE
- Task and success criterion
- Object requirements, placement, and tolerance
- Model versions, results, and verification context
Establish locally: matching objects, a calibrated work surface, a feasible task, and the same scoring rule.
This illustrates the product direction. It does not connect to a robot or use customer records.
The next test
starts with what
we already learned.
We have tested initial and final image capture, task-completion scoring, and spatial success maps. We are rolling that fuller loop into customer deployments, building on the placement and verification system already in the field.
Our next data product is a library of recreatable failure cases. Start with the customer’s own history, then add cases contributed by comparable labs. A team pays for useful challenges it has not discovered itself—and evidence about what to test next.
05 / CUSTOMERS INSTALL THE INFRASTRUCTURE
Hardware opens the door. Usage expands the business.One station.
Then the rest of the lab.
installed and running
Enact came through another lab, tried a unit, and expanded to five. We built and installed those units by hand. The next step is standardized production and setup that customers can complete without us onsite.
Hardware economics before ongoing software, support, payroll, and other operating expenses.
Every model release
creates another decision.
Keep a change. Catch a regression. Choose the next test. Customers build a library of useful cases and return to it as their models evolve. That is the recurring software value we are building.
Explore the software economics Illustrative scenarios
WHAT USAGE COULD BECOME
The more a lab evaluates,
the more useful Prism becomes.
We are developing recurring software pricing. Explore a monthly model or usage tied to evaluations.
Proposed pricing · illustrative scenario1,000 stations × 100 evaluations × 250 active days × $1
Scenario, not current revenue or a forecast. Assumes the displayed paid adoption, activity, and pricing. Usage counts completed, verified evaluations; rejected setup checks are not extra billable events. Hardware sales are excluded.
Stations
Paid hardware that fixes a daily problem in the lab.
Evaluation software
Recurring usage for test orchestration, results, coverage, and the next test.
Failure-case library
Contributed cases that help another lab discover weaknesses sooner, with benchmarks and data products built from that evidence.
THE NEXT PROOF
Targets · not current tractionScale the installations.
Prove the software.
The next stage is about repeatability: units customers can install without us, software they return to, and failure cases useful enough to pay for.
Show that a contributed case exposes a previously untested weakness in another compatible lab—and is useful enough to run again.
06 / THE LIGHT COMPANY
Robotics is where we start.The room becomes
the interface.
Our original vision is AR without glasses: intelligent light that understands a workspace, guides the next action, and checks the result.
Robotics gives us a concrete place to start. The same perception, calibration, task guidance, and verification tools can help people learn, assemble, repair, and work in the physical world. Each new workflow adds its own task knowledge.
Explore the wider visionJON SOLOMON · FOUNDER
Built beside
the people using it.
I helped ship Apple Vision Pro, leading two stability labs with about 150 robots and fixtures. I built Prism with the labs that needed it, working beside their operators. Now we’re turning those first deployments into infrastructure that can grow across entire labs.
THE NEXT CHAPTER
Physical AI should iterate
at the speed of code.
Change the model. Recreate the conditions.
Measure the improvement. Know what to train next.