How it works
This page explains exactly how Xybern works, the architecture, the 5-stage pipeline, the verification engine, the identity layer, and the provenance vault. If you are evaluating Xybern for your organisation, this is where to start.
SHA-256 · HMAC-SHA256 · Merkle proofs · Cryptographic agent identity
Section 01 · The execution pathway
Every AI action in your organisation passes through all 5 stages before it can execute. This is not middleware. This is not a monitoring layer. This is the execution path itself.
An AI action that has not passed through all 5 stages has not been enforced. Xybern makes unenforced execution structurally impossible.
Section 02 · Each stage in detail
The pipeline overview names the 5 stages. This section explains what is passed in, what is checked, and what is returned at each one.
Stage 01
Your AI system calls Xybern instead of executing directly. One API endpoint replaces the direct call to your LLM, agent framework or automation tool. The action is held inside the pipeline, it cannot proceed to any subsequent system until all 5 stages complete. There is no way to route around the interception point.
Stage 02
The agent proposing the action is verified against its cryptographic identity before any evaluation begins. Xybern establishes who is acting, what their defined authority boundaries are, and whether their current behaviour is consistent with their historical baseline. An agent whose identity cannot be verified does not proceed.
Stage 03
The xybern-reasoning-v1 model receives the proposed action, the verified identity, the current trust score, the active policy set and the jurisdictional context for this action. It decomposes the action into atomic claims, evaluates each claim independently against evidence and policy, and produces a structured enforcement determination.
Stage 04
Binary. Authorise or block. The decision is deterministic, given the same inputs it produces the same output every time. There is no grey zone, no retry, no human-in-the-loop delay. The enforcement decision is made in under 50ms. The action either proceeds or it does not. There is no third option.
Stage 05
Every decision Xybern makes, authorised or blocked, is written immediately to the provenance vault. The entry contains the full action, the agent identity, the evaluation result, the enforcement decision, and a precise timestamp. It is anchored in a SHA-256 cryptographic hash chain with HMAC-SHA256 signatures. The entry cannot be deleted. The entry cannot be altered. The chain itself is evidence.
Section 03 · Stage 03 in depth
Most enforcement systems use rules. Rules are static. Rules do not understand context. xybern-reasoning-v1 is a proprietary 7B reasoning model trained specifically to evaluate AI actions against policies, not to generate outputs. It receives the proposed action, the agent's verified identity, the trust score, the active policy set and the jurisdictional context. It returns a structured enforcement determination.
Verification response object
Every enforcement decision returns a fully structured object. Not a score, a complete record of what was checked, what was found, and what was decided.
Claim decomposition
Actions are broken into atomic claims. Each claim is evaluated independently against evidence and policy. A single failed claim is sufficient to block the entire action, there is no partial authorisation.
Adaptive trust scoring
Trust scores update dynamically based on agent behaviour history across all sessions. An agent that has previously exceeded its authority boundary carries a lower baseline trust score on all future actions until rehabilitated.
Policy enforcement
Governance policies, regulatory frameworks and internal rules are all evaluated in a single pass by the same reasoning model. SM&CR accountability rules, FCA SYSC obligations and custom internal policies handled uniformly.
Section 04 · Stage 02 in depth
Not just a token.
API tokens tell you which application is calling. They do not tell you which agent within that application proposed this specific action, what reasoning chain it used to arrive at it, or whether its behaviour is consistent with its history. Xybern assigns every agent a persistent cryptographic identity, verified at every stage of every execution.
Section 05 · Runtime enforcement
This is Xybern running in production. Every row is a real enforcement decision, an AI agent proposed an action, Xybern evaluated it, and a decision was made. The decisions on the right are final. They cannot be reversed. They are already in the vault.
Section 06 · Stage 05 in depth
Nothing mutates silently.
Every enforcement decision Xybern makes is written to the provenance vault. The vault is an append-only ledger — entries can never be deleted. Each entry is anchored in a SHA-256 cryptographic hash chain with HMAC-SHA256 signatures and Merkle proof verification. The chain itself is the evidence.
Append-only ledger
Entries cannot be deleted
Once written, permanent. The vault has no delete operation.
SHA-256 hash chain
Entries cannot be altered
Any change to any entry breaks the hash chain, immediately detectable.
Merkle proof export
Chain integrity verifiable
Full Merkle proof export for any compliance audit, on demand.
Section 07 · The platform
The Xybern enforcement dashboard gives your security and compliance teams a live view of every AI enforcement decision across your entire organisation. Authorised actions, blocked attempts, trust scores, agent identities, policy flags, all visible, all permanent, all auditable. Nothing is hidden. Nothing expires.
Section 08 · Deployment patterns
One enforcement standard.
Xybern supports two deployment patterns depending on whether you are a platform provider embedding enforcement into your AI product, or an enterprise deploying enforcement across your internal AI estate.
Embedded deployment
Xybern integrates directly into your AI platform stack. It sits between your model outputs and your end users. Every output your platform delivers has been intercepted, verified, scored and recorded before it reaches the user. Your platform gains enforcement and provenance as a native capability.
Centralised deployment
Xybern deploys as an infrastructure layer that sits above every AI system in your organisation. Internal LLMs, employee copilots, autonomous agents, customer-facing AI, all route through a single enforcement point before any action reaches production.
Section 09 · Integration
Any provider. Any framework.
Integrating Xybern into your existing AI stack requires one change, you call Xybern instead of calling your LLM or tool directly. Nothing else changes. Your existing infrastructure, your existing models, your existing agent frameworks all remain in place.
Compatible with your stack
Deployment time
API integration
Point your first agent at the Xybern endpoint. First enforced action in under an hour.
Policy configuration
Define authority boundaries, trust thresholds and policy constraints for your AI systems.
Full deployment
All AI systems routing through Xybern. Full enforcement, full audit trail, full visibility.
Section 10 · Regulated environments
Xybern was designed from the ground up for regulated industries. The enforcement pipeline, the identity layer and the provenance vault were all built with financial services and legal services compliance obligations in mind.
Financial services
Xybern maps directly to SM&CR Senior Manager accountability requirements. Every AI decision is attributed to a specific agent with a specific authority scope and a specific enforcement record. FCA SYSC 8 outsourcing obligations satisfied by design. Full audit trail available for regulatory inspection at any time.
Legal services
Every AI action taken on behalf of a client passes through Xybern before it executes. Supervision obligations are met structurally, not by policy documentation but by architectural enforcement. Client data boundary rules enforced at the action level, not the application level.
Enterprise · UK & MENA
Jurisdiction-aware enforcement means the same policy engine applies different rules depending on where an action originates and where it has effect. UK, EU AI Act and MENA regulatory contexts are all handled in a single Xybern deployment, no separate configurations required.
Let's deploy it.
Start with a design partnership. We work directly with your team to deploy Xybern into one workflow in under two weeks. No lengthy procurement. No infrastructure rebuild. One endpoint.
