The Triarcus Architecture
Cathedral is a layered constitutional governance architecture built on a simple idea: authority should be explicit, and a system should not be able to act outside — or quietly alter — the governance that constrains it.
A human authority sits at the apex. Governance policy is compiled and checked before deployment. Components operate only within their compiled authority. If the governance does not pass that check, the system does not deploy.
What is demonstrated today
Cornerstone is the working implementation — a pre-execution control layer that evaluates a proposed action before it executes and returns a determination with the reason and the authority that bound it.
Four determinations. Proceed — the action clears every authority. Block — an authority forbids it, attributed to the rule that stopped it. Escalate for discretion — the facts are complete, and the judgment is not the machine's to make. Escalate for data — a fact the rules require is missing; the action is held, and the record names what is needed. A determination may also carry flags: a reporting obligation annotates a verdict without being one, and never changes it.
Four concurrent authorities. Every action is evaluated independently by each authority in force, each holding a veto within its own domain. No authority outranks another. The determination is the most restrictive of the four, and the record names which one bound it.
Evaluation cost is constant in rule count. Rules are compiled before deployment, so runtime cost is a function of the configuration — authorities by dimensions — rather than of how many rules are loaded. Measured at roughly 22 microseconds from ten rules to ten thousand.
Determinations are reproducible. The same inputs against the same prior state produce a content-identical determination, verified across 500 runs under real multi-process concurrency.
Verified against a second implementation. The architecture was implemented a second time from its written specification, on a different internal rule-encoding scheme, and both were exercised against the same transactions — ten scenarios, nine governing rules, four concurrent authorities — compared on outcome, verbatim reason text, and binding authority. Ground truth was recomputed from source on every run. Agreement held on all three axes.
Both implementations are ours. The independence here is of internal model, not of authorship — the stronger property an auditor needs is not established, and we do not claim it. Four of the nine rules carry qualifications, and a fast evaluation path produces the correct verdict but not an equivalent record. All of it is written up in full.
What is designed and protected
The broader architecture — additional governance, reasoning, oversight, and interoperability layers — is documented, and its core elements are the subject of filed patent applications. These are design and roadmap, built as real demand and independent validation warrant them, not presented as finished capability. We would rather show what runs honestly than claim a system that does not yet exist.
THE
CONSTITUTIONAL
CEILING
Core Governance
Evaluates a proposed action against every rule in force, determines whether it is permitted, and monitors continuously whether the system is drifting outside what it was authorized to be.
Interpretation & Coherence
Resolves what a request actually asks for when it is ambiguous or conflicting, and checks that each decision stays consistent with what was decided before.
Perception, Memory & Infrastructure
Controls what enters and leaves the system, protects the integrity of what it remembers, and maintains an accurate picture of the environment it is acting in.
Reasoning & Strategy
Applies the governing rules to hard cases — where authorities conflict, where the facts are contested, or where the system must examine its own reasoning before acting on it.
Relational & Emotional
Governs how the system communicates with people and how it treats them, including whether comparable cases are handled consistently and fairly.
Defense, Oversight & Orchestration
Protects the governance layer itself from interference, monitors the system while it runs, and coordinates multiple governed components so they do not work against one another.
Output & Applied Governance
Applies the governance to specific domains and jurisdictions, and lets systems under different governance cooperate without either imposing its rules on the other.
Human authority → constitutional locks → compiled governance → every component
A human holds final authority, and no component of the system can override it.
Three constitutional locks cover the ways a system could change its own governance: what it may become, what it may erase, and what rules it may write for itself. The system refuses any change to anything they cover, and no single party can authorize it.
Governance is compiled and checked before deployment. If it does not pass, the system does not deploy.
INTERACTIVE ARCHITECTURE DIAGRAM
Govern the action, before it happens.
Cathedral evaluates what an AI agent is about to do — against every authority that governs it, each holding an independent veto — and returns a deterministic, auditable decision before the action reaches the world. It sits outside the agent, so its judgment does not depend on the agent's. It does not read the agent's reasoning or its explanation of itself; it evaluates the proposed action against the rules in force.
A capable agent is not the same as a dependable one
Autonomous agents plan, call tools, and act. When one drifts, is manipulated, or simply encounters a case no one anticipated, the safeguards built into its own reasoning are the least reliable thing to depend on — because the reasoning is exactly what failed. Governance has to be separate from the thing being governed, and it has to decide before the action executes.
Permitted under the rules that apply. It executes, with a complete record of why.
An interpretive or contested case is routed to the responsible person — never decided by the machine.
The action is held, and the record names what is needed and who must supply it. The system does not guess, and does not default.
Stopped before execution, attributed to the specific rule and authority that stopped it.
A determination is a verdict and a record. It is not a signed token or a cryptographic capability: the caller receives an answer and the reason for it, not a credential.
Five parts, one architecture
Cathedral is built from five governing functions. Each answers a distinct question about a proposed action, and together they form a single pre-execution decision.
Governance is compiled and checked before deployment
Rules are compiled before anything runs, and the compiled set is checked for contradictions and for completeness — a policy whose gates do not say what happens when they fire is refused at compile time, not discovered at runtime. A configuration that does not pass does not deploy.
Some constraints cannot be changed by anyone
A protected class of constraints cannot be modified by the client, the vendor, or the AI itself. Changing them requires multiple independent authorizations — no single actor, human or machine, can quietly move the boundaries.
Every applicable authority is satisfied at once
Real deployments answer to more than one rulebook — legal, regulatory, organizational, local. An action proceeds only when every applicable authority permits it, with no single authority's precedence hard-coded.
What the system does, and becomes, stays in bounds
Inputs are checked at the perimeter; state is protected from unauthorized change; and after an action, the system's own learning is inspected so it cannot quietly drift outside what it was authorized to be.
Independently governed systems can interact without imposing on one another
When two AI systems under different governance need to cooperate, they exchange information across the boundary without either being forced to adopt the other's rules — and any joint action is authorized by a human on each side.
What is built, and what is designed
We would rather show one working layer honestly than claim a system that does not yet exist.
Cornerstone — the working engine
A deterministic pre-execution control engine. It evaluates a proposed action and returns one of four determinations — proceed, escalate for discretion, escalate for data, or block — with the reason and the authority that bound it, and produces the same determination every time for the same inputs and the same prior state. It runs today.
The full architecture
The five functions above are specified in detail, and their core elements are the subject of filed patent applications. Layers beyond the working engine are designed and staged for build, not represented as operational.
Cathedral is developed by Triarcus Systems as a licensed governance engine for regulated, high-consequence environments. The engine is what we license; the system it powers is what our partners build on it.
See it decide
Add a governance rule and watch the same proposed action be handled differently — because of the rule you changed, attributed to the authority that changed the outcome.
The complete architectural engine is available as an interactive diagram. Click any part of the engine to see its description, role, and connections.