In every cell of every model measured.
The guarantee is measured against the record, not asserted: zero is the number of times a model got an action past the gate without a human's grant.
VEGA is a control plane that sits outside the model. Every consequential action passes a gate the model cannot talk its way past, so choosing a model stops being a safety decision. Work is reconciled against what actually happened rather than taken on the agent's word, and a hash-chained record shows what was done without trusting the agent's account. Open weights on your hardware, or a provider of your choice.
A benchmark world, not production. Every number below comes from benchmark runs, and each says which.
The guarantee is measured against the record, not asserted: zero is the number of times a model got an action past the gate without a human's grant.
Legitimate work refused in 1 of 240 runs across three models: 0 of 120 on qwen3:8b, 0 of 60 on nemotron-3-super, 1 of 60 on llama3.1:8b, where the right tool was called with a wrong parameter name. Quoted beside the guarantee rather than hidden from it.
In one synthetic benchmark scenario, an 8B model with the control plane finished the long-horizon task in 40 of 40 runs. Without it, 7 of 40 (17.5%), on the same model and the same task (p = 1.2e-15). One scenario, not a general rate.
Deployed, then plugged in, then trusted a little at a time.
The declaration is the one file the agent can never write: who answers for it, which hosts it may reach, which model it runs on. A node is not a folder on a PC.
It proposes a source, you click approve. Credentials sit in a vault under a key it can name but never read. No connector is shipped for any vendor; your deployment authors its own.
A mandate names actions, scope and a duration. Autonomy widens on two axes, never both at once by default. Every human step is one click on something the agent drafted.
Every proposed action is checked against typed preconditions, the mandate's scope and a host allowlist your organisation wrote. What persuaded the model to propose it is irrelevant to the check. A secret rides into a request by key; the model never holds the value. The record is hash-chained with no update and no delete.
Read the threat modelOpen weights on your GPUs, or the leading models from the high-end providers.
Locked to no vendor. Every model runs the same qualification battery before it is trusted with anything: seven cells, ten seeds each, two arms. The unaided arm shows what the model does on its own; the kernel arm shows what passes with VEGA in front of it.
A pass on the floor clears a model for use; it does not rank it above its neighbour. A person reads the record before deciding what the model may be trusted for. The full catalogue, every model, every cell → · See the fleets a node can run →
Your GPUs, your network, your keys. Nothing of yours touches a server we run.
We run the control plane; inference stays where you put it.
One node per customer, never shared.