Architecture

The SovereignClaw architecture is built for AI agent runtime security: deterministic execution gating, 33 Rust crates, 1,345 workspace tests passed in the October 7, 2026 release run, and 9 formal security properties.

SovereignClaw's architecture treats every LLM-proposed action as untrusted input. The seven-stage execution pipeline turns model output into a deterministic, policy-evaluated decision before any side effect can occur. For teams evaluating enterprise AI governance, this page shows how runtime security, evidence generation, and operational control fit together in one system.

This architecture is designed for organizations that need more than best-effort AI guardrails. Instead of filtering or post-hoc review, SovereignClaw canonicalizes intent, verifies facts, assigns risk, and binds any permitted execution to signed artifacts and receipts. Pair this architecture view with the security properties, compliance coverage, and deployment tiers to understand the full operating model. Published on SSRN (ID 6290760). The DetGate execution-gating method behind SovereignClaw is patent pending with the USPTO.

Execution Pipeline

Every agent action passes through a deterministic 7-stage pipeline before any execution can occur. No shortcuts. No bypasses.

Stage-by-stage execution flow

Intent
→
Canonical IR
→
Fact Infer
→
Iron Gate
→
Adapter
→
Receipt
→
Merkle Ledger

Risk Tier Classification

Every action is classified into one of four risk tiers. The gate behavior is determined by the tier, not by the LLM's self-assessment.

Policy behavior by tier

TierNameExampleGate Behavior
T0Observels -laAllowed by default policy
T1StandardWrite app configAllow with signed receipt
T2ElevatedRestart production serviceRequires threshold signatures (default quorum 1 of 2)
T3SovereignDelete IAM / DB / infraMechanical refusal: refused by default policy

Crate Architecture (33 Crates)

The workspace is grouped by role: the kernel, platform services, skills and supply chain, agent and LLM integration, self-improvement, CLI tooling, binaries, and test crates. The October 7, 2026 release run passed 1,345 workspace tests with zero failures. That separation is what lets SovereignClaw map execution controls cleanly into audit and compliance evidence.

Core, control-plane, and runtime crates

CrateRole
libs/coreImmutable SovereignIR schema, canonical hashing, execution graph
libs/gatePolicy engine: fact recalculation, tier classification, threshold signatures, WASM plugin sandbox, revocation check
libs/adapterAttested runtime: 14-step verification, nonce replay protection, AEAD credential store, SQLite revocation registry
libs/receiptAppend-only, hash-chained ledger of dual-signed authority receipts with Merkle anchors
libs/tenantMulti-tenant isolation: context hashing, quotas, cross-tenant denial
libs/vaultPer-tenant, per-environment credential vaults with lease-gated access
libs/cryptoEd25519 and SLH-DSA-SHA2-256f (FIPS 205) receipt signatures, threshold signature collection, and key rotation
libs/identitySSO/SCIM identity bridge: IdP group → approval quorum mapping
libs/approvalPersistent async approval workflows with quorum, TTL, escalation
libs/telemetryObservability and metrics export
libs/enterprise16 governed enterprise adapters (Terraform/Pulumi, kubectl, Vault/AWS Secrets Manager, psql, AWS CLI, gh, PagerDuty, ServiceNow/Jira) with tier classification (T0 to T3)
libs/skill_registryVerified manifests, publication state, quarantine, intake freeze, tenant allowlisting
libs/skill_scannerWASM bytecode behavioral analysis - flags network, file-write, env-read imports against declared scope
libs/wasm_runtime, libs/toolsSandboxed WASM skill runtime and governed tool definitions
libs/meshEd25519-signed inter-agent envelopes, delegation chains with scope attenuation, propagation depth limits
libs/agentLLM orchestration loop, provenance-tagged memory, context injection boundary, swarm execution runtime
libs/provider, libs/memoryLLM provider integration and agent memory
libs/browser-adapterGoverned Playwright browser adapter, driving a separate stdio-only worker with no network listener
libs/strategy, libs/evaluator, libs/learningZero-authority Stage 1 learning: captures and drafts lessons into a tenant-local shadow store; cannot change memory, policy, Gate, or adapters
libs/cli-core, libs/cli-generatorCLI tooling shared by the binaries
libs/gatewayHTTP control plane with deployment attestation, compliance metrics (CEF/SIEM), production mode profile, signed POST /api/v1/evaluate, and governed A2A v1
bin/sovclaw, bin/sovclaw-gateway, bin/sovclaw-tuiCLI, gateway server, and terminal UI
bin/sovclaw-improve, bin/sovclaw-browser-workerSelf-improvement runner and governed browser worker
tests/adversarial, tests/proptestAdversarial and property-based test suites
Request Early Access

Frequently Asked Questions

What is the SovereignClaw execution pipeline?
Intent → Canonical IR → Fact Recalculation → Iron Gate → Adapter → Receipt → Merkle Ledger. Every agent action passes through this deterministic pipeline before execution.
How many crates make up SovereignClaw?
SovereignClaw is a workspace of 33 Rust crates: the kernel (core, gate, adapter, receipt, crypto); platform services (tenant, identity, approval, vault, telemetry, enterprise, mesh, gateway); skills and supply chain (skill_registry, skill_scanner, wasm_runtime, tools); agent and LLM (agent, provider, memory, browser-adapter); self-improvement (strategy, evaluator, learning); CLI tooling (cli-core, cli-generator); five binaries (sovclaw, sovclaw-gateway, sovclaw-tui, sovclaw-improve, sovclaw-browser-worker); and two test crates (adversarial, proptest).
What are the SovereignClaw risk tiers?
T0 (Observe): reads, allowed by default policy. T1 (Standard): ordinary writes, deletes, and execution, allowed with a signed receipt. T2 (Elevated): production execution, infrastructure writes, confidential data, or crossing a trust boundary; requires threshold signatures, with a default quorum of 1 of 2 operators. T3 (Sovereign): identity roots, unbounded impact, or infrastructure deletion; mechanically refused by default policy, and signatures cannot override a deny.