Authority before consequence

Your AI agent can call the tool.Should the action be allowed to run?

Ark Sovereign provides deterministic authority checks and replayable evidence for consequential AI-agent actions.

Supported inline workflows: ALLOW · DENY · REQUIRE APPROVAL

Independent record

The agent should not audit itself.

An agent can select tools, shape workflows, and generate a narrative around its own behavior. That narrative is not an authoritative record of whether the action was permitted.

Ark operates outside the agent. It evaluates declared authority before consequence in supported integrations, and preserves the applicable policy, evidence, reason codes, approval state, and observed side-effect status.

Two operating modes

One authority discipline. Clear operating boundaries.

The same packet discipline supports an inline authority boundary where integrated, and evidence-led replay where a workflow is not yet connected to a runtime path.

Runtime authority mode

For supported inline integrations

Ark can be placed before a consequential tool call, evaluates the declared authority envelope, returns a deterministic verdict, and records whether a governed handle resolved or remained blocked.

Supported inline paths use bounded wrappers or governed handles around selected tool calls; integration scope is confirmed during the workflow review.

  • ALLOW · DENY · REQUIRE APPROVAL
  • governed handle resolution
  • zero-operation receipt
  • boundary packet
Evaluation and replay mode

For offline or not-yet-integrated workflows

Ark evaluates whether proposed or recorded actions fit the authority an organization intended to delegate. It does not imply an action was intercepted or prevented.

  • sanitized traces or representative requests
  • reconstructed authority envelope
  • deterministic replay and policy-gap findings
  • integration recommendation
One integration point

Put authority before the consequential tool call.

For supported inline integrations, Ark governs whether selected actions cross into consequence. For evaluation workflows, it determines how those actions would be classified under the declared authority boundary.

01 · PROPOSE

Agent requests a consequential action.

02 · EVALUATE

Ark compares the request with the declared envelope.

03 · VERDICT

ALLOW, DENY, or REQUIRE APPROVAL.

04 · RECORD

Tool resolution or blocking status becomes a boundary packet.

{
  "agent": "customer-support-agent",
  "action": "issue_refund",
  "amount": 8500,
  "authority_envelope": "support-tier-1"
}
{
  "verdict": "REQUIRE_APPROVAL",
  "reason_code": "REFUND_LIMIT_EXCEEDED",
  "tool_invoked": false,
  "external_effect": false
}
Example scenario

The $8,500 refund.

Same agent. Same refund tool. Different outcome when the delegated limit is visible before the action crosses into consequence.

Customer support remains useful. The organization retains authority.

A customer-support agent proposes a refund above its delegated threshold. Instead of treating tool access or a trusted session as blanket permission, the policy evaluates the amount against the declared limit.

Tool availability is not authority.
A valid session is not blanket authority.
agent customer-support-agent
requested_action issue_refund
amount $8,500
delegated_limit $5,000
verdict REQUIRE_APPROVAL
reason_code REFUND_LIMIT_EXCEEDED
tool_invoked false
external_effect false
human_handoff finance-review-queue
Three product objects

What Ark produces.

01

Permit verdict

A deterministic policy result for a declared request: inside the envelope, prohibited, or requiring authorized human review.

02

Boundary packet

A replayable record of the request, governed target, policy and envelope version, evidence, reason codes, missing information, and observed side effects.

03

Zero-operation receipt

When a governed action remains blocked, the packet records that the handle did not resolve and no external side effect was observed within the evaluated scope.

Consequence examples

Start with one action that matters.

Refund or payment

Approval thresholds, credit actions, transfers, and exception paths.

Production change

Shell execution, dependency installation, deployment, or file modification.

Sensitive-data export

Scope, destination, retention, history, and declared data-handling authority.

External message

Customer communications, browser actions, outbound writes, and approval requirements.

Phase 0 authority boundary review

Review one agent workflow.

You do not need to redesign an entire agent platform. Start with one consequential action and representative evidence.

You provide

  • 5-20 sanitized traces or representative requests
  • intended authority limits and approval rules
  • actions that should never execute automatically

You receive

  • authority map and replayable boundary packets
  • ALLOW, DENY, and REQUIRE APPROVAL cases
  • policy-gap findings and one integration recommendation

ALLOW applies only when the declared action, target, context, and policy conditions match. High-impact actions can remain approval-gated by policy.

No production credentials or unrestricted infrastructure access required. Sanitized inputs accepted.

Proof before promise

Inspectable evidence, bounded claims.

10 / 10Boundary replay cases passed
0LLM calls in enforcement path
74,934Rows reviewed in offline RobotGov evaluation
falsepacket_authorizes_execution
Replay caseDeclared action boundaryVerdictPrimary reason code
Code safe patchApproved repository-root patchALLOWPOLICY_MATCH
Code secret exfilShell-based secret exportDENYSECRET_EXFIL_ATTEMPT
Code production deployProduction deploymentREQUIRE_APPROVALHIGH_IMPACT_ACTION_REQUIRES_APPROVAL
MCP tool poisoningPoisoned tool description or resultDENYMCP_TOOL_POISONING_PATTERN
MCP shadow serverUnapproved MCP serverDENYMCP_SERVER_NOT_APPROVED
MCP unsigned toolUnsigned tool schemaDENYUNSIGNED_TOOL_SCHEMA
MCP missing telemetryGoverned action without required telemetryDENYTELEMETRY_REQUIRED
Wallet public-reply transferTransfer exceeding value policyDENYVALUE_LIMIT_EXCEEDED
Channel-origin patchUntrusted external-channel code patchREQUIRE_APPROVALCHANNEL_ORIGIN_REQUIRES_APPROVAL
Unknown-agent deploy handoffUnknown agent in production handoff chainREQUIRE_APPROVALHANDOFF_REQUIRES_APPROVAL_FOR_EXECUTION

These results demonstrate deterministic replay and bounded packet behavior under published evaluation scopes. They are not security, legal, safety, compliance, or production-readiness certification.

Portability

Rent intelligence. Own authority.

Models may change. Enterprise authority should not change silently with them. Ark keeps policy versions, action permissions, tool boundaries, financial thresholds, external-write rules, approval requirements, and replayable evidence outside the model.

Primary surfaces

One authority layer. Multiple consequence surfaces.

The entry point is tool and action authority. These five buyer-facing surfaces apply the same bounded packet discipline to specific decision domains.

MCPGov

Governed tool calls, credentials, shell operations, network requests, and unmanaged MCP paths.

CodeGov

Repository scope, dependency installation, commands, deploys, and egress authority.

WalletGov

Refunds, payments, transfers, credit limits, and financial approval thresholds.

ContextGov

Governs sensitive context access, retention, quotation, transmission, and evidentiary use, including memory integrity, stale-evidence controls, and protection against poisoned or denied-event replay.

ClaimsGov

Evidence gaps, policy application, unsupported claims, and human claims authority.

Specialist cybersecurity capability

Ark Manifold Cyber

Deterministic DeFi and EVM security review using bytecode, control-flow, topology, and semantic analysis. Produces ranked, bounded vulnerability findings and sealed evidence packets without claiming exploit confirmation or a complete smart-contract audit.

Explore Ark Manifold Cyber
The principle

Put an independent refusal and evidence layer between an agent and consequence.

A capable agent is not automatically a trusted agent. A fluent explanation is not permission.