Agentic infrastructure · sovereign by construction
Every consequential action stops and waits for a person. Every inference is routed by policy before a model ever sees the data. Not because it is written in a policy document — because there is no code path around it.
Mission
To put agentic AI inside institutions that cannot outsource their accountability — owned outright, routed by policy, and stopped by a person before it does anything that matters.
An open, declarative platform
A new service is six declarations over one contract — agents, knowledge, connectors, prompts, experience, policy. Runtime, retrieval, routing, gates, placement and audit are inherited, and a service cannot loosen them.
Partners build packs on the same contract we do. Which is why the tenth service costs domain work rather than platform work.
The architecture
Seven planes, owned once and operated everywhere. A vertical service declares six things and inherits everything else — which is why the ninth service costs domain work rather than platform work.
One platform, many agentic services
the Platform Service Contract v1.1 is the only interface between them
each pack declares six things and nothing more
TerraSigna agentic platform — seven planes, owned once, operated everywhere
The pack carries domain meaning. It never carries runtime, retrieval, routing, gates, placement or audit.
Each is owned by the platform, versioned by the platform, and reachable by a service only through the contract.
The platform
Ten agent archetypes, one kernel, and a single path from an intent to an action — the same eleven steps for a 5G design, a meditation session and a factory work order.
A service designer picks an archetype and specialises it. The archetype carries proven defaults for memory, tooling, safety and model policy that a specialisation may tighten and may not loosen. Adding an eleventh is a platform decision, never a service workaround.
| Archetype | Behavioural contract |
|---|
Two of these carry the weight. The Actuator is the only archetype permitted a side-effecting external call — so the set of things that can change the world is enumerable by reading declarations rather than by reading prompts. The Guardian is mandatory in every service and red-teamed before release.
The same path every time. Only the declarations differ.
The class decides the tier and the site. A request carrying regulated personal data cannot reach a frontier API, whatever the prompt asks for, because the router evaluates the data rather than the request.
The part most companies leave out
Both sit inside the execution path rather than in front of it. There is no configuration that removes them and no prompt that talks past them.
Mechanism one · architecturally unbypassable
Three stacked rulebooks — privacy, sovereignty, regulation — evaluated on every single inference. It sits inside the reasoning gateway rather than in front of it, so there is no path to a model that avoids it.
A regulated context cannot reach a frontier API whatever the prompt asks for, because the router evaluates the data and not the request. The decision is made before a model sees anything, and it is recorded.
Mechanism two · declared per service, enforced by the platform
Each service declares which of its actions are consequential and who may confirm them. A network configuration push. A grade written to a permanent record. A maintenance command on a production line. Identical mechanics everywhere, and the declaration is data the platform enforces — not a behaviour the service is trusted to implement.
Execution pauses for a named human. The decision is recorded immutably whether it is approved or refused, so the audit trail contains the refusals too.
Only the Actuator archetype may make a side-effecting external call, and every consequential one raises the gate. That is what makes the claim checkable: the set of things this system can do to the world is enumerable by reading declarations, rather than inferred by trusting prompts.
The catalog
Each is a certified, versioned bundle that installs onto the platform the way an application installs onto an operating system. Each declares six things and inherits the rest.
What a service pack actually is
six declarations — everything the vertical owns, and nothing else
TerraSigna agentic platform — inherited unchanged · zero pack engineering
What the pack adds is domain meaning. What it never touches is runtime, retrieval, routing, gates, placement or audit.
What is different here
Each one is a property of the architecture rather than a policy we promise to follow. That distinction is the entire proposition.
01
The hardware is yours from the first invoice. The control plane is yours. The data never leaves ground you control, and an air-gapped site never connects at all — it takes updates on physical media. No hyperscaler sovereign region can say that, because the control plane always stays with the vendor.
02
Policy routing, the human-confirmation gate and a mandatory Guardian arrive with the platform. A service cannot opt out of them and a partner cannot loosen them. The posture may be tightened for a market. It may never be relaxed.
03
Every delivered site contributes its design, its acceptance results and its operational telemetry back into the knowledge domain, so sizing accuracy and proposal speed improve with each engagement. A competitor starting today starts that accumulation at zero.
A demonstration proves something worked once, for the person who built it. An evidence pack proves it works for you, and can be audited after everyone involved has moved on.
Engagements begin with the customer's own requirements specification, not with our product sheet. The design that comes back is sized against it and validated against the rules that apply in your jurisdiction.
Every criterion is tested and its result recorded with the evidence attached, so a procurement office that has changed staff twice can still read what was accepted and why.
A service release must pass adversarial red-teaming of its Guardian, functional tests of both gates, quality thresholds on its own evaluation sets, and a conformance run in a clean environment its authors did not set up. Including our own.
Cost models and facility validators must return identical verdicts for identical inputs — enforced as a service objective, not hoped for.
Contact
The most useful first conversation is about your requirements specification and the rules you have to satisfy — not about our product.
TerraSigna AI, Inc. · Boca Raton, Florida, United States
Non-negotiables
Everything this pack owns. Everything else is inherited from the platform.