A model's benchmark score says what it can do. It says nothing about what it will do on a Tuesday afternoon, under a mandate, with a message in front of it that was written to be obeyed. We measured that second number, model by model, and it does not follow the first. VEGA exists for the distance between the two.
The industry ranks models by what they can do. We ran the same models under a standing mandate, inside a small world with real consequences planted in it, and counted what they did.
Across six models, the correlation between a public capability index and how often a model stepped outside its mandate was negative and not significant. Two models with the same index score differed by seventy-nine points on the same attack. Capability tells you what a model can reach for. It does not tell you whether it will.
A benchmark world, not production. Every number on this site says which.
A rule that lives inside the prompt is a rule the model can be talked out of, by a document, a message, a tool result, or a sponsor in a hurry. The only rule that holds is one the model never sees: a gate outside the prompt, in deterministic code, that every consequential action passes or does not.
Intelligence and authority are separate systems. VEGA is the second one.
Because the guarantee lives outside the model, it does not depend on which model. Open weights on hardware you own, or the leading models from the high-end providers on a route you declare: the same gate, the same record, the same battery run before any of them is trusted with anything.
A pass on the floor clears a model for use. It does not rank it.
Every company will run agents. The ones that keep running them will be the ones that can answer, for any action, who allowed it, on what evidence, and what happened next.
Autonomy is not the absence of a person. It is a person whose every act is one click on a draft the agent wrote: a mandate, a connection, a proposal. VEGA counts the clicks an outcome costs and works to make each one smaller, never to remove the person from the record.
Everything VEGA knows about a deployment is rebuilt from a journal nobody can edit, VEGA included. What stage a node has reached, what it may do, what it did, what it got wrong: all of it is read off the record, never asserted by the agent. That is what lets a company trust a thing it did not write.
Vega was the first star after the Sun to be photographed and the first to have its spectrum recorded. It is the star instruments are calibrated against: the fixed point other measurements are read from. A control plane should be that for an organisation's agents, a reference that does not move when the model does. Its constellation is Lyra, the lyre. The pages of this site are its stars.