Deterministic AI governance

Govern AI
before it acts.

VaultACE Hoku is the pre-execution control plane between intelligent systems and consequential action. Models may propose. Deterministic controls decide what is authorized.

Request a technical briefing See how Hoku works

Model-agnosticFail-closedReplay-resistantEvidence-linked

PACT · CONSEQUENCE TRANSACTIONTXN–HOKU–00418

VALIDATING

REQUESTExecute bounded actionAgent identity: GARDIT–027

✓Purpose lineageVERIFIED

✓Registered identityVERIFIED

✓Human authorityL3 ACCEPTED

✓Hard invariantsSATISFIED

✓Reality evidenceCURRENT

✓Replay protectionCLEAR

DETERMINISTIC DISPOSITIONAUTHORIZED

Illustrative trace · every real disposition depends on applicable policy, evidence, authority, and current state.

01 / THE CONTROL BOUNDARY

“Intelligence may propose. Deterministic systems must authorize consequence.”

01No proposal is permission.

02Scope can narrow, never silently expand.

03No identity or evidence means no action.

04The same inputs replay to the same decision.

Architecture

One governance plane.
Six accountable systems.

Hoku does not try to replace your models, optimizers, agents, or domain software. It governs the boundary where their outputs become real consequences.

Pre-execution gate

VaultACE Hoku

Separates intelligence from authority and issues a deterministic ALLOW, DENY, HOLD, or ESCALATE disposition before consequential execution.

Invariant enforcement

PHALANX

Applies hard rules for identity, authority, scope, containment, auditability, reversibility, purpose, and evidence.

Proof-carrying actions

PACT

Binds proposals to required proof, authorization state, execution receipts, postconditions, and final disposition.

Registered identity

G.A.R.D.I.T.

The Global Agent Registration Database Identification Tracker establishes which registered agent or service is making the request.

Human authority

HCCM

Maps consequential action classes to explicit human authority levels so software cannot silently grant itself more power.

Evidence lineage

OpenVault

Preserves tamper-evident decision lineage across request, policy state, authorization, execution, and post-action verification.

Consequence lifecycle

From intelligence to accountable execution.

Every consequential request moves through an explicit chain. Nothing jumps from recommendation to action.

Propose

A model, optimizer, agent, or operator submits a bounded request.

Validate

Hoku verifies purpose, identity, scope, freshness, evidence, and preconditions.

Authorize

Hard invariants and required human authority produce a deterministic disposition.

Execute

Only an authorized, allow-listed adapter may dispatch the approved action.

Settle

Postconditions, receipts, and the final evidence chain are checked and preserved.

Assurance by refusal

Trust is defined by what the system refuses to do.

Hoku is designed to expose uncertainty, missing authority, stale state, and broken evidence instead of hiding them behind an apparently successful action.

Request the evidence package

Representative failure conditions and expected dispositions
Condition	Disposition
Unregistered or unauthorized requester	DENY
Valid request with stale reality state	HOLD
Action exceeds granted scope	DENY
Previously used transaction is replayed	DENY
Required evidence is missing	NO EXECUTION
Materially conflicting state estimates	HOLD / ESCALATE

Applications

Built for teams accountable for consequences.

Start with one bounded workflow, measurable failure modes, and an integration test that shows whether governed consequence improves real assurance.

AI & agent platforms

Add hard authority boundaries around assistants, agentic workflows, multi-agent systems, and tool execution.

Aerospace & robotics

Govern the boundary between analytical output, mission scope, operator authority, and physical action.

Financial & regulated systems

Require explicit policy, evidence, limits, and accountable authorization before consequential transactions.

Critical infrastructure

Fail closed when identity, state, provenance, permissions, or post-action verification cannot be established.

Validation posture

Designed to be challenged—not merely believed.

Vaultagon supports reproducible review of deterministic decision logic, adversarial failure scenarios, evidence lineage, and integration boundaries. The practical question is not whether governance sounds convincing. It is whether it reliably catches invalid, stale, replayed, unauthorized, or under-evidenced requests.

✓ Release-candidate and adversarial test artifacts
✓ Deterministic replay and decision evidence
✓ Bounded technical-validation pathway
✓ Independent technical review welcomed

Hoku is not represented as a substitute for domain certification, a formal safety case, legal review, or regulatory approval. Applicable evidence should be evaluated for each deployment context.

VAULTAGON INDUSTRIES LLC · HAWAIʻI

Make autonomous capability easier to trust, test, and authorize.

Vaultagon develops deterministic governance infrastructure for AI agents, autonomous systems, and high-consequence automation. We work with technical teams through focused architecture review, evidence review, licensing discussions, and bounded validation projects.

Start a conversation

Govern AIbefore it acts.

One governance plane.Six accountable systems.