Deterministic Agent Control Plane

The model is
probabilistic.
The control plane is not.

Ark Sovereign evaluates action envelopes against explicit policy before tools execute — deterministically, with no LLM in the enforcement path. Execution control for AI agents that write code, move value, execute MCP tools, or touch production.

ark-sovereign — enforcement proof matrix
node demos/run_governor_proof_matrix.mjs
{
  "pass_count": 10,
  "fail_count": 0,
  "llm_calls_in_enforcement_path": 0
}
code_safe_patch ALLOW
code_secret_exfil DENY
code_prod_deploy REQUIRE_APPROVAL
mcp_tool_poisoning DENY
mcp_owasp_shadow_server DENY
mcp_owasp_unsigned_tool DENY
mcp_owasp_no_telemetry DENY
wallet_public_reply_transfer DENY
channel_github_issue_patch REQUIRE_APPROVAL
handoff_unknown_agent_deploy REQUIRE_APPROVAL

Seven layers.
One enforcement boundary.

Each governor targets a distinct attack surface in the AI-agent execution stack. No LLM in the permit/deny path.

GOV-01
Agent Governor CORE

Policy gateway for agent actions. Emits ALLOW, REQUIRE_APPROVAL, or DENY with deterministic reason codes before any tool executes.

GOV-02
ToolCall Firewall

MCP and API tool preflight: server allowlists, signed schemas, telemetry requirements, argument controls, and poisoning checks.

GOV-03
CodeGov

Controls file writes, shell execution, dependency installs, deploys, protected paths, and secret-exfiltration attempts.

GOV-04
WalletGov VALUE

Controls AI-wallet and payment-like actions with value thresholds, approval gates, destination policy, and audit evidence.

GOV-05
Audit Ledger

Replayable decision records with deterministic hashes that bind policy, action, verdict, and reason codes into a tamper-evident trail.

GOV-06
Rollback Planner

For blocked or paused actions, emits containment controls and a safe retry path for common failure modes.

GOV-07
Policy Packs

Reviewable policy packs for engineering, security, compliance, MCP, wallet, and agent-handoff workflows. Human-readable. Version-controlled.

67 replay cases.
3 buyer segments.
0 LLM calls.

The 10-case table below is the public sampler. The enterprise proof pack contains 67 deterministic replay cases across WalletGov, CodeGov, and MCPGov, with a clean legitimate allow rate of 1.0.

67
Replay cases
3
Segment packs
0
LLM calls in path
Tier 2 Enterprise Proof

Do not evaluate Ark Sovereign from the 10-case sampler alone.

The full replay pack shows realistic buyer workflows: agentic payments, software-engineering agents, and MCP runtime orchestration. Every case includes verdict counts, reason codes, deterministic hashes, and downloadable JSON.

25WalletGov fintech cases
24CodeGov engineering cases
18MCPGov runtime cases
1.0Clean legitimate allow rate
Action EnvelopeWhat It ProvesVerdictReason Codes
Code Safe Patch Safe coding-agent patch inside approved repo root. ALLOW POLICY_MATCH
Code Secret Exfil Secret exfiltration through shell command. DENY LOW_TRUST_SOURCE_FOR_ACTION
UNTRUSTED_CONTENT_BEFORE_ACTION
DANGEROUS_SHELL
SECRET_EXFIL_ATTEMPT
NETWORK_ACCESS_REQUIRES_APPROVAL
Code Prod Deploy Production deploy requires human approval. REQUIRE_APPROVAL HIGH_IMPACT_ACTION_REQUIRES_APPROVAL
Mcp Tool Poisoning Poisoned MCP tool description or result is blocked. DENY LOW_TRUST_SOURCE_FOR_ACTION
UNTRUSTED_CONTENT_BEFORE_ACTION
MCP_TOOL_POISONING_PATTERN
Mcp Owasp Shadow Server Unapproved MCP server is blocked. DENY MCP_SERVER_NOT_APPROVED
Mcp Owasp Unsigned Tool Unsigned tool schema is blocked. DENY UNSIGNED_TOOL_SCHEMA
Mcp Owasp No Telemetry Required telemetry missing for governed action. DENY TELEMETRY_REQUIRED
Wallet Public Reply Transfer Wallet-like value transfer exceeds autonomous value policy. DENY LOW_TRUST_SOURCE_FOR_ACTION
UNTRUSTED_CONTENT_BEFORE_ACTION
VALUE_LIMIT_EXCEEDED
Channel Github Issue Patch Untrusted channel-origin code patch requires approval. REQUIRE_APPROVAL CHANNEL_ORIGIN_REQUIRES_APPROVAL
Handoff Unknown Agent Deploy Unknown agent handoff before deploy requires approval. REQUIRE_APPROVAL UNKNOWN_AGENT_IN_HANDOFF_CHAIN
HANDOFF_REQUIRES_APPROVAL_FOR_EXECUTION
HIGH_IMPACT_ACTION_REQUIRES_APPROVAL

Input → verdict.
No model in the middle.

Ark Sovereign replays structured action envelopes through deterministic policy. The enforcement proof is reproducible from the same inputs every time.

Input
policy_json
event_json action envelope
reason predicates
tool call arguments
value, destination, file path, deploy target
Output
ALLOW / REQUIRE_APPROVAL / DENY
reason_codes + controls
deterministic audit hash
containment controls if blocked
safe retry path if applicable

Buyer-specific proof.
Not just a sampler.

Synthetic but realistic action logs for the three buyer surfaces enterprise teams ask about first: payments, coding agents, and MCP runtime orchestration. Every verdict is deterministic, replayable, and generated with zero LLM calls in the enforcement path.

WALLETGOV_FINTECH
WalletGov / Agentic Payments Replay Pack

fintech security, payments platform, wallet automation, treasury operations

Cases: 25
Verdicts: ALLOW 12 / APPROVAL 2 / DENY 11
Clean legitimate allow rate: 1.0

Download JSON →

CODEGOV_ENGINEERING
CodeGov / Agentic Software Engineering Replay Pack

VP Engineering, platform engineering, DevSecOps, regulated software teams

Cases: 24
Verdicts: ALLOW 12 / APPROVAL 6 / DENY 6
Clean legitimate allow rate: 1.0

Download JSON →

MCPGOV_RUNTIME
MCPGov / Multi-Agent Runtime Replay Pack

AI platform security, MCP operators, multi-agent orchestration teams

Cases: 18
Verdicts: ALLOW 4 / APPROVAL 4 / DENY 10
Block or approval gate rate: 78%
Clean legitimate allow rate: 1.0

Download JSON →

Send logs.
Get verdicts.

Send 5 to 20 sanitized AI-agent action logs. We map them into the action-envelope schema and return deterministic verdicts, policy gaps, reason codes, and a Phase 1 integration plan.

01
Send sanitized logs

5 to 20 agent action logs. No credentials, source code, production access, secrets, or internal prompts required.

02
Schema mapping

We map your logs into the action-envelope schema and identify relevant governor surfaces.

03
Deterministic verdicts

Every result has policy source, verdict, reason codes, controls, and audit hash.

04
Phase 1 integration plan

Policy gaps identified. Integration path scoped. Clear boundary: we govern execution, not conversation style.

No credentials required

No production access, source code, secrets, or internal prompts. Sanitized action logs only.

Replayable evidence

Every result has policy source, verdict, reason codes, controls, and audit hash. Deterministic from the same inputs.

Clear enforcement boundary

We do not ask a second model to decide whether the first model is safe. Policy decides whether the action executes.

0 LLM calls in path

The permit/deny decision is made entirely by deterministic policy evaluation.

Request Phase 0 Evaluation →