Today, the Verify Quality tab opens to the public on PrismaX. For the first time, anyone can score the robot training data that models learn from, and earn for the work.

We call the people who do it validators. Over the next month, the community will score demonstrations, build a track record, and compete for a founding role. The 100 validators whose scores land closest to consensus become The First 100: the founding Data Quality Validators on PrismaX.

Physical AI is hardware-rich and data-poor. The robots are here and the models are ready. What’s missing is enough high-quality data to teach them, and a way to tell the data worth training on from the data that isn’t. The VLA Foundry is how we build that data layer, and the better it gets, the faster robots learn to do real work in the real world. Verify Quality is where it begins.

Better Models Start With Better Data

Better robotics models start with better training data. Better training data starts with people scoring it.

It's the same pattern behind the language models people use every day. Their fluency traces back to a layer of human judgment, people deciding what's good and what isn't. Robotics is no different.

Most robotics data isn't worth training on. Datasets are growing, but task selection remains repetitive, and much of what's sold as robotics data is too narrow, too synthetic, or too inconsistent to teach a model anything useful. Volume isn't the problem the field needs to solve. Quality is.

Software can measure the obvious things. A black frame, a camera out of sync, a hand that drifts out of the shot. What it can't do is judge quality. Whether the motion was smooth and deliberate. Whether the operator handled the robot the way you'd want a model to learn from. Whether the demonstration is one worth training on at all. That judgment is the work of a validator.

We already run a version of this scoring in-house. The data that clears it trains the robotics models our customers build, and we run it through our own training and evaluation as part of our research. With the Verify Quality tab, we're opening that work to the community for the first time.

How Scoring Works

Verify Quality turns data review into a clear, repeatable task.

You pick an environment, a task, and a dataset. Each dataset holds multiple episodes. You watch each one across its camera views, then you score it.

Every episode is checked against a set of pass-or-fail criteria first. Is the camera feed clear? Did the robot complete the task as instructed? Did the robot hand stay in frame? Were the cameras in sync? An episode that fails any of these isn't training-grade, no matter how good the rest looks.

From there, you score quality on a sliding scale. How well the operator controlled the robot. How smooth the motion was. Whether the pace made sense. Whether the task reached its intended final state. Whether the episode adds variation instead of repeating what the dataset already has.

Each episode is scored independently by many validators. Those scores combine into a consensus, and the consensus determines the final result. When validators disagree, the data goes to expert review. Once you submit a score, it's final. Locking submissions is what keeps the consensus accurate.

The closer your scores track consensus, the stronger your track record.

The First 100

The First 100 is how that track record turns into a founding role.

Scoring is open to anyone with an Amplifier or Innovator membership. Amplifiers get 10 submission chances in the first window. Innovators get 30. Every submission is a chance to score with consensus and earn Prisma Points, whether or not you make the cut.

When the first window closes, the 100 validators with scores closest to consensus become The First 100. Selection is open. There are no reserved seats. The role is earned by scoring well, and once you're in, you can score unlimited data each month.

The First 100 isn't a fixed list. Each month, the bottom 25 percent rotate out, and new validators rotate in. The standard holds because the people holding it have to keep earning the seat.

Built and Tested With Our Day 1 Launch Partners

Before today, our Day 1 Launch Partners and PrismaX Ambassadors tested Verify Quality, scored real data, and told us what worked and what didn't. Their feedback shaped the scoring criteria and the validator experience you're using now.

Our Day 1 Launch Partners are Blockchain Builders Fund, ChainGPT, Monad, OpenMind, Peaq, Perle Labs, Sentient, Stanford Blockchain Accelerator, Virtuals Protocol, XMAQUINA, and Zeno.

We're grateful for the time our partners and Ambassadors put in before launch, and for helping us hold the first version of this standard to a higher bar.

One Piece of the VLA Foundry

Verify Quality is the first public step into the VLA Foundry.

The Foundry is how PrismaX produces the kind of dataset robotics and AI companies actually need: diverse, large-scale, high-quality, and usable. It works because we run the whole loop ourselves. Operators upload data. Validators verify it. Customers use it. What we learn from training and evaluating models feeds back into how the next round of data gets collected.

We can hold the line on quality because we operate the system end to end. We've seen what separates training-grade data from a full hard drive, and we've built the standard around it. The validator layer is how that judgment scales past our own team and out to a community of 1,500,000+.

This is the service layer for physical AI, backed by real operation. The full VLA Foundry arrives in the coming months. The First 100 is where it starts.

What Comes Next

The first scoring window is open now. Score well, build your track record, and earn your place in The First 100.

If you want the deeper version of how we define quality, read Standards for Data. 

The standard starts with you. 

Begin Validating today: https://app.prismax.ai/