VEGA

Any model. The last agent you'll ever need.

VEGA turns a language model into a complete working agent. Every consequential action passes a gate outside the model, so a mistake never becomes a consequence. And the longer it works inside your organisation, the better it gets. Open weights on your hardware, or the leading models from the high-end providers.

Scroll to dive in
Sovereign enterprise autonomy Your hardware or the leading providers

01 · The guarantee, in numbers

Each number holds the screen while you scroll
unauthorised actions

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.

docs/findings · 2026-08-16 · the price re-measured on a fixed grader
legitimate actions refused

The price of the gate.

Legitimate work refused in 0 of 180 runs across two models, and quoted beside the guarantee rather than hidden from it. A gate that blocked real work would be a brake; this one is a record.

docs/findings · over-blocking · two models · 180 runs
long-horizon work completed

The gate is not a brake.

With the control plane, 100% of long-horizon tasks completed on an 8B model. Without it, 17.5%, on the same model and the same tasks (n = 40 per cell, p = 1.2e-15).

docs/findings · 2026-08-19 · the reply verb inverts sign on long-horizon · qwen3:8b · kernel arm vs baseline

02 · How a node comes to life

Tiles assemble from the star field as you scroll

Deployed, then plugged in, then trusted a little at a time.

01 · Deployed

An identity, an accountable human, a model route.

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.

02 · Plugged in

It asks which services carry your organisation's life.

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.

03 · Under mandate

It drafts the smallest authority for the next thing. You grant it.

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.

03 · Outside the prompt

The journal is the object

The kernel imports no model.

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 model
journal · liveverify_chain ok
  • #4166 · 09:40:51 · source.observed · registered:poll_board · 3 items
  • #4167 · 09:40:52 · signal.raised · unreviewed_change · invoice 2026-03-118
  • #4168 · 09:41:07 · action.proposed · http_request · host:billing.example · secret:BILLING_KEY
  • #4169 · 09:41:07 · action.refused · host not in mandate scope · value never shown
  • #4170 · 09:41:09 · mandate.drafted · reconcile March invoices · host:billing.example · 24h
  • #4171 · 09:52:31 · mandate.granted · by the accountable human · one click
  • #4172 · 09:52:33 · action.proposed · http_request · host:billing.example · secret:BILLING_KEY
  • #4173 · 09:52:33 · action.receipt · executed · postcondition met
  • #4174 · 09:52:34 · work.reconciled · satisfied
  • #4175 · 09:52:34 · agent.outreach · "Eleven invoices reconciled; two need a human eye."
  • #4176 · 10:03:10 · operator.answered · "Leave the two for Monday."
  • #4177 · 10:03:11 · domain.fact.confirmed · invoice 2026-03-118 · disputed
  • #4178 · 10:03:12 · chain.verified · 4178 entries · sha256 · no gap
A realistic sample, not a customer's record

04 · What holds

Verified against a file and a test on disk
The gate
  • The gate sits outside the model. Prompt injection changes what the model proposes, never what passes.
  • Every consequential action is typed: preconditions, postconditions, idempotency keys, reconciliation and recovery.
  • A secret rides into a request by key. The model never holds the value, so it cannot leak one.
The record
  • Loopback by default. Every route needs a bearer token or your identity provider. Rotation locks out a running server.
  • The journal has no update and no delete. Verifying it recomputes every digest and raises on any gap.
  • Backups never contain the vault. Personal data is erased by destroying its key, and the chain still verifies.

The full threat model →

05 · Your model, measured before it is trusted

Rebuilt from the journal · a floor, not a ranking

Open 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.

ModelFloorUnaided, failedWith VEGA, passed
qwen3.5 · 4Bpassed6 of 7068 of 70
qwen3.5 · 9Bpassed0 of 6969 of 69
qwen3.8 · 27Bpassed0 of 7070 of 70
gemma4 · 31Bpassed0 of 7070 of 70

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 →

06 · Three ways to run it

Pilot pricing on request · the endpoint is always yours
Pick № 1

Self-hosted licence

Your GPUs, your network, your keys. Nothing of yours touches a server we run.

Pick № 2

Hosted kernel, your endpoint

We run the control plane; inference stays where you put it.

Pick № 3

Fully hosted, yearly

One node per customer, never shared, metered on what it does.

You have seen the numbers

Give it a mandate now.