ROOTVIELROOTVIEL

Technology

How ROOTVIEL works

ROOTVIEL is a verifiable multi-agent system. Agents reason and coordinate; a deterministic layer decides what is authorized and what is proven. Below is the cognitive cycle at a glance, then the articles that go deep.

Not a scanner — the cognitive cycle

Read the full article →
  1. 1Observe the target and build context
  2. 2Maintain beliefs, goals and intentions
  3. 3Form and rank hypotheses — deterministically
  4. 4Select the next action by information gain
  5. 5Authorize through the tool gate (security first)
  6. 6Act, and observe the result as evidence
  7. 7Verify the evidence with a deterministic verifier
  8. 8Replan on the evidence
  9. 9Leave what is unproven as UNKNOWN
  10. 10A human reviews and validates the result

// budget exhausted → UNKNOWN, never a hopeful CONFIRMED

Not loops — guarantees

In the end, it is a guarantees problem

An LLM in a loop makes a good demo and a brittle system. What makes the system trustworthy is not autonomy — it is four guarantees, each enforced deterministically, outside the model.

Authorization

Authority lives in code, not in the model. Every tool call passes a single deny-by-default gate that reads capabilities, never the model’s confidence. The agent cannot self-authorize.

Evidence

A finding exists only when an independent, LLM-free verifier re-derives it from an observed, unforgeable state change. "Looks vulnerable" never becomes a reported fact.

Epistemic honesty

Claims are labelled PROVEN, NOT PROVEN or UNKNOWN. What cannot be demonstrated is marked UNKNOWN — never "safe" — and a run that finds nothing never means secure.

Traceability

Every decision and action emits a tamper-evident, hash-chained record, so any result can be reconstructed with the evidence that produced it.

The loop is the easy part. These guarantees are the work — and the reason a ROOTVIEL run can be trusted where an autonomous agent, on its own, cannot. The Academy develops each one from first principles.

Two planes, one signed bus

Authority is minted in one place, executed in another

The plane allowed to authorize an active test is not the plane that runs it. They speak only through a one-way, cryptographically signed event bus — so persuading the agent, or compromising the engine, cannot forge authority.

Authority across two planesDocumented architecture
CONTROL PLANEEXECUTION PLANEreads the signed message — never calls backsigned order · Ed25519trust boundary · one-way busauthorizedactsPrincipala human operatorAuthorize & sign orderdeny-by-default; two signing pointsAgent proposesreasoning — untrusted, propose-onlyTool authorization pointsingle deny-by-default gateExecution environmentisolated sandboxEvidencetraces + resulting stateDeterministic verifierno AI, no randomnessMembranepromotion frontier — does not reasonFindinga proven violation✕ Deniednot executed · counters stay 0UNKNOWNunpromotable — not “safe”EDGE KINDSauthority (grant / signed order)data / proposalevidencedeny / off-ramp (✕ = blocked)

The plane that is allowed to act is not the plane that does act. A signed order crosses the boundary one way; the engine reads the self-contained message and never calls back. The agent cannot skip the gate or self-authorize, a blocked action is never counted as work, and a claim only becomes a Finding once a deterministic verifier and the membrane promote it — otherwise it stays UNKNOWN.

Deep dives

The mechanisms, in detail

The cognitive cycle, the evidence model, the two-plane signed architecture and deterministic enforcement — each written to teach the mechanism.