A reflex layer for autonomous machines.
A second detection channel that shares no model, no training pipeline, and no inference runtime with your perception stack.
In most autonomy stacks one network decides whether there is a hazard, and nothing downstream can tell when that decision is wrong. We run a second channel on the same lidar and camera input. It reaches its own answer by a different route and flags the frames where the two disagree.
A product line of Univault Technologies. Patent-pending.
Why this matters
A vehicle meets something your perception stack does not flag. Turning that one event into fleet-wide behavior means a labelling cycle, a retrain, a full revalidation of a network that changed underneath you, and a software-update file that has to be re-opened.
Autonomy is usually described as a perception problem. The bottleneck is the correction loop.
Label it, change the model, prove nothing else broke. Once the weights move, your existing evidence describes a system you no longer ship, so the safety case gets re-argued and, in type-approval markets, the update file gets re-opened.
Then the fleet finds four more.
Almost none of this is bounded by compute, and compute is the part you can buy more of. The bound is the people who write the safety argument and the independent assessors who sign it, and that capacity does not grow because you need it to.
The date moves because the loop that corrects a fault runs slower than the loop that finds one. You can tell where the industry thinks the constraint is by what it is paying for. Waymo is spending to compress the time it takes to open a new city. Mobileye built a map that updates itself off the fleet. Wayve is betting on not needing a map. None of that is a bet on better driving. All of it is a bet on making the next change cost less than the last one.
An independent monitor on your perception channel, built so that keeping it current costs less than rebuilding it.
The evaluation
The only quantity worth anything to you is how much of your own stack’s residual this channel recovers on your own fleet’s data. That number does not exist in a public benchmark and we are not going to manufacture a proxy for it. So we do not lead with numbers. We run the test you specify, on a bench, with your logs.
Pass and fail are defined by your team, in writing, before we see any of your data. Including the scoring convention and the false-alarm budget you actually have to live inside.
A replay on your recorded data against your own detector. No integration, no vehicle time, and nothing installed on anything that moves.
Your team scores it against the criteria your team wrote, and keeps the result either way. We see the same report you do.
How we release results. A result about a safety layer is only readable with its conditions attached: which detector it was measured against, the false-alarm rate both sides were held to, the distance band, and the definition of a miss. Those four travel with every number we release, under a mutual non-disclosure agreement.
Where it runs, and what it is for
None of these needs a benchmark to evaluate. Two of them you can check against hardware you already own.
The reflex layer has its own model, its own training pipeline, and its own inference runtime. Nothing it does depends on the perception network being correct. It is not sensor-independent: both channels read the same sweep, so a blinded sensor takes both down together. The independence is from the perception software, and that is the failure family it is there for.
Two kilobytes of state and a fixed-cost comparison against it. No floating-point network inference, no dynamic allocation, and control flow that does not branch on the data. That is a budget we are targeting at the lockstep microcontroller already in the vehicle rather than the accelerator inside the main system-on-chip. The port to that microcontroller is in build.
When the fleet finds a hazard pattern the car did not have, updating the channel for it is designed to be a smaller thing to argue under an update-approval regime than shipping a different network. No assessor has given an opinion on that yet.
What this is, and is not, today. At its current false-alarm rate the reflex layer is a monitoring, arming, and event-capture channel. It is not a brake authority. The false-alarm rate is the gate, and closing it is the program we are running.
Tier-1 supplier, OEM, robotics platform, eVTOL (electric vertical take-off and landing) program — if you have to argue independence for a perception channel, we should be talking.
