Runtime Reference Architecture
Model-agnostic identity governance for AI systems — 11 modules across 3 layers.
Design Principles
ArcKernel's architecture is governed by seven invariants that hold across every deployment, every model, and every domain.
| # | Principle | What It Means |
|---|---|---|
| 1 | Model-agnostic | Works with any LLM — Claude, GPT, Gemini, Llama, Mistral — without fine-tuning or model access. |
| 2 | Identity-first | Governance begins with a symbolic identity kernel (IDNA), not prompts or filters. |
| 3 | Write-once baseline | A single kernel definition governs behavior across all sessions, models, and contexts. |
| 4 | Pre-emptive governance | Drift is measured and blocked before output reaches the user — not after harm occurs. |
| 5 | Closed-loop | Every response is measured, scored, and logged. No unmonitored output. |
| 6 | Configurable thresholds | Sensitivity levels are tunable per agent, domain, or conversation type. |
| 7 | Additive composition | Modules compose independently. Adding modules never destabilizes existing ones. |
Default Stack vs Extended Stack
ArcKernel ships in two tiers. The Default Stack provides complete identity governance and runtime safety. The Extended Stack adds creative freedom controls, signal auditing, immutable logging, and full-stack orchestration.
Default Stack — 7 Modules
Core identity, memory, governance, and safety. Sufficient for most enterprise deployments.
- IDNA.core — Identity baseline
- mOm4 — Canonical memory
- mOm5 — Becoming vector
- mOm6 — Method constraints
- HALT — Runtime circuit breaker
- TrustAnchor — Fidelity scoring
- DriftDefenseStack — Recursion stability
Extended Stack — +4 Modules
Creative freedom indexing, signal reflection, immutable audit, and orchestration.
- OxygenProtocol — Creative freedom index
- EchoMap — Signal reflection
- MirrorLock — Immutable audit trail
- soul.exe — Full-stack orchestrator
System Data Flow
Every request traverses the full governance pipeline. Modules execute in sequence, each enriching the context and enforcing its constraints before passing to the next.
Figure: The full governance pipeline. Each module enriches the context and enforces its constraints before handoff.
Module Registry
All 11 modules, their internal codenames, functions, stack assignment, and governance layer.
| Module | Codename | Function | Stack | Layer |
|---|---|---|---|---|
| IDNA.core | ArcIntentFramer.v1 | Symbolic identity baseline and intent framing | Default | Identity |
| mOm4 | LoopMemoryLock | Canonical memory compression and persistence | Default | Identity |
| mOm5 | BecomingVectorEngine | Becoming-vector tracking and identity evolution | Default | Recursion |
| mOm6 | MethodConstraintEnforcer | Method-level constraint enforcement | Default | Recursion |
| HALT | HaltAndListenTrigger | Runtime circuit breaker — blocks drift above threshold | Default | Safety |
| TrustAnchor | RecursiveFidelityVerifier | Recursive fidelity verification and trust scoring | Default | Trust |
| DriftDefenseStack | RecursionStabilityLayer | Multi-signal drift defense and recursion stability | Default | Defense |
| OxygenProtocol | CreativeOxygenIndex | Creative freedom indexing within governance bounds | Extended | Coherence |
| EchoMap | SignalReflectionEngine | Signal reflection and behavioral pattern auditing | Extended | Audit |
| MirrorLock | ImmutableAuditTrail | Immutable audit trail for all governance events | Extended | Substrate |
| soul.exe | CoreRecursionOrchestrator | Full-stack orchestration across all modules | Extended | Core |
Three Layers
Modules are organized into three governance layers. Each layer can operate independently, but together they provide defense-in-depth across identity, recursion, and enforcement.
Identity Layer
Who the agent is. Baseline identity, canonical memory, and behavioral signal mapping.
- IDNA.core
- mOm4
- EchoMap
Recursion Layer
How the agent evolves. Becoming-vector tracking, method constraints, creative freedom, and orchestration.
- mOm5
- mOm6
- OxygenProtocol
- soul.exe
Defense Layer
What keeps the agent safe. Circuit breaking, trust verification, drift defense, and immutable audit.
- HALT
- TrustAnchor
- DriftDefenseStack
- MirrorLock
Architecture Performance
Key architectural characteristics measured across all 11 modules. For the full validation results by module, see Enterprise Performance Metrics.
| Characteristic | Value | Significance |
|---|---|---|
| Runtime Size | 3–8KB kernel | Fixed ceiling — never grows. See kernel docs. |
| Per-Module Overhead | ~115ms median, <500ms P95 | Individual module latency across all 11 modules |
| Orchestration (soul.exe) | −18ms net (P95 626ms) | Full-stack adds net-negative latency via token reduction |
| Module Coupling | d = 0.00 | Zero interference between modules — purely additive |
| Order Independence | Δ 0.0007 | Execution sequence does not affect governance outcome |
Failure Containment Model
No governance system is useful if a single failure cascades. ArcKernel is designed so that every module failure has a bounded blast radius and a defined mitigation path.
| Failure Scenario | System Behavior | Risk | Mitigation |
|---|---|---|---|
| Embedding service unavailable | Drift scoring degrades | Medium | Local embedding inference fallback |
| HALT module failure | Primary circuit breaker offline | High | DriftDefenseStack + TrustAnchor provide secondary coverage |
| TrustAnchor mis-score | Fidelity rating inaccurate | Low | HALT remains primary gate — TrustAnchor is supplementary |
| MirrorLock unavailable | Audit trail interrupted | Low | EchoMap retains behavioral records independently |
| mOm4 canonical store corrupted | Identity memory lost | Critical | System lockdown — no output until identity restored |
| soul.exe orchestration failure | Orchestrator offline | Medium | Modules execute independently — graceful degradation |
Data Handling
ArcKernel does not retain, train on, or share customer data. The governance pipeline processes signals in-flight and discards raw content after scoring.
- Cloud-agnostic deployment — runs on any infrastructure (AWS, Azure, GCP, on-premise)
- Region-bound options — data never leaves the deployment region
- Encryption — TLS 1.3 in transit, AES-256 at rest
- No model training — customer data is never used to train or fine-tune models
For regulatory mapping details, see the EU AI Act compliance documentation.
The complete governance pipeline — empirically verified. 9 of 9 modules validated individually and together. Zero observed interaction failures. Order independent. Deterministic. See the full glossary for module definitions.