TELEMETRY
Full state at frame resolution. Sensors, actuators, compute load, comms.
The Record
This is where a league becomes a technology company.
Every kata, every round and every arena match is captured. Telemetry, engagement decisions, failure modes, human overrides, and the exact conditions that produced them.
Adversarial data is the hard kind to get.
Most autonomous systems are trained against the world, not against something actively trying to defeat them. Adversarial data at scale is expensive, dangerous and slow to generate, which is why almost nobody has it.
We generate it as a sport. Continuously, from a field of builders who are motivated to try things we would never have thought to test, because there is money on the line and their name is on the machine.
Six things, from every round.
DECISIONS
What the machine chose, what it considered, and the confidence attached.
FAILURE MODES
Every fault, classified. What broke, when, and what preceded it.
OVERRIDES
Every human intervention, with the state that triggered it. This is the highest value signal we hold, and it is the one nobody else is collecting.
CONDITIONS
The exact scenario seed, so any round is reproducible by anyone.
OUTCOMES
Labelled by result. Adversarial data with ground truth attached is the whole point.
Three things the record does.
The simulator.
Real behaviour makes simulated opposition harder. The dojo gets more honest every season, which means qualifying gets harder every season.
The grading.
Belts stop being a judgment call. Thresholds come from the distribution of what machines actually do, not from what we think they should do.
The gap.
We measure the difference between dojo performance and arena performance across the whole field, and we publish it. Closing that number is our core technical problem and our most honest scorecard.
Who gets it, and on what terms.
Builders own their code and always have full access to their own machine's record. That is not a privilege we grant, it is yours by default.
Aggregate, anonymised data is published openly after each season, because a field that can check our claims is a field that will take them seriously. If we ever stop publishing it, assume the numbers stopped being flattering.
Commercial and institutional access is licensed, not sold outright, and it is governed by the data policy. Nothing that identifies a builder or exposes their stack leaves this company without their written consent.
If you need capable autonomous systems, we should talk.
Institutional, commercial and defense enquiries go to the partnerships team, and they get a straight answer about what we have actually measured.