Three-Layer Architecture

Divided power applied to cognition.

01
Human Leads

Provides direction, intent, and purpose.

The human is the authority over what matters, what to pursue, and when to stop. The system does not act without direction and does not execute without confirmation. Human authority is not delegated to the system, the system operates in service of it.

02
System Governs

Owns memory, evidence, truth, and authority.

Deterministic. Cannot be persuaded. The governance layer evaluates what the system is authorized to say before the language model is ever invoked, and validates what it said after. Evidence without grounding defers. Contradiction qualifies. Conversational pressure increases transparency, not compliance.

03
Model Speaks

Receives a governed context packet and produces language.

The model renders within permitted bounds. It does not decide. It cannot set policy. It cannot determine what is true. The specialist agencies report to governance. Governance decides. The spokesperson reads a prepared statement. The spokesperson cannot rewrite the statement.

The Runtime Flow

Every turn passes through two gates.

User InputThe human provides direction or request
Intent EngineGoverned specialist classifies intent (100% accuracy, 36 intents)
Evidence CollectionMemory, relational context, and grounding data assembled
Authority Engine (Pre-LM)6 deterministic rules. Language model not invoked until this passes.
Single LM CallGoverned context only. The model renders what it has been authorized to say.
Compliance Validator (Post-LM)4 checks verify output matches what was authorized
OutputDelivered to user
DROPNever patched

The architecture flow is not a guardrail bolted onto a language model. It is a fundamentally different structure, one where the component that produces language never holds the authority to determine truth.

Every decision made in the authority engine and compliance validator is captured in a turn provenance snapshot. Every turn. Auditability by design, not by optional logging.

When the compliance validator catches an unsupported claim, the response is dropped entirely. The system falls back to a safe response. It would rather say nothing than present a patched fabrication as fact.

This is the distinction between mitigation and architecture. Mitigation filters bad output. Architecture prevents the conditions that produce it.

SBA Spine: Technical Details
Two Gates Around One LM Call

Deterministic before. Verified after.

The language model receives exactly one bounded call per turn. Everything before it is governance. Everything after it is verification.

Gate 1

Authority Engine (Pre-LM)

  • R1No grounding available? Defer to safe fallback before the model is called.
  • R2Contradiction in evidence? Mandatory qualification injected into context.
  • R3Conversational pressure detected? Increase transparency, do not increase compliance.
  • R4Two or more blocking reasons? Force safe fallback. The model is not called.
  • R5Output bounded: what is authorized to say, what caveats are mandatory, what claims are supported.
  • R6Provenance snapshot captured. Every authority decision is recorded before output.

The language model receives a packet defining exactly what it may say. Nothing more is available to it.

Gate 2

Compliance Validator (Post-LM)

  • C1Unsupported factual claims present? DROP. The response is not delivered.
  • C2Missing mandatory caveats? Attempt reattachment. If reattachment fails: DROP.
  • C3Direct claims from a deferred response? DROP. The model overstepped its authority packet.
  • C4Compliance result appended to provenance snapshot. Full audit trail per turn.

The system attempts correction only for missing caveats. All other failures result in the response being dropped entirely. Never corrected and presented as truth.

The Core Claim

Not a better model. A different architecture.

We currently use frontier models (Claude, GPT, Gemini, and local models) as the expression layer, because that is what is available today. The architecture governs them. The relationship, the memory, and the governance persist regardless of which model is active.

A language model without state, memory, and governance is not a system. It is a component. What AI-nHancement has built is the rest of the system, the parts that make it complete.

The industry is selling components and calling them systems. We are building actual systems. The goal is a fully decomposed cognitive architecture where every function is governed, every specialist is purpose-built, and the frontier model is one swappable expression component in a larger structure we own.