Governor AI
The management and safety system for AI agents that operate real infrastructure. Rules specific to each system, program and function are enforced outside the AI, at the point of action. The AI proposes; LawGuard decides whether the action may run, under which authority, and records the evidence.
THE PRODUCT, EXPLAINED
Put AI to work. Keep authority with you.
Start with the short film. Then see how it works in the full demonstration.
Useful autonomy, with LawGuard enforcing your rules outside the AI model.
Press Play. CC adds captions. Use full screen for the clearest view.
Illustrative scenarios. * Extended features in final R&D.
Why enforcement sits outside the model
An AI agent can be asked to follow rules, but a prompt is not an enforcement mechanism. Governor AI separates the two: the agent proposes, and an independent component, LawGuard, decides at the moment an action would run. The laws it applies are specific to the system, the program and the function in question, so a maintenance bot is governed by maintenance laws for its own fleet, not by a generic filter.
This is what makes governed AI operations safe and reliable without relying on the AI simply obeying its instructions. It is the mission of Governor Technologies: an AI management and safety system in which enforcement sits outside the model, so that you never have to rely on an AI simply following its prompt.
Function-specific laws
A maintenance bot is governed by maintenance laws for its own fleet, not by a generic filter. Laws are written for each system, program and function.
Authority you choose
Per-action approval, or explicit standing delegation within LawGuard limits. Selected per function and per client, and recorded as an exact grant.
Evidence for every decision
Each proposal, decision, authority and result is recorded so operators and independent reviewers can see exactly what happened and why.
The LawGuard decision cycle
Every governed action moves through the same deterministic cycle. Nothing in the cycle depends on the model's goodwill.
-
1
Observation
The governed program observes its system: an alert, a monitoring signal, a scheduled task or an operator request. Governor AI does not replace the program's own sensing; it governs what the program does about it.
Inputs Alerts, monitoring state, operator requests, scheduled maintenance triggers
-
2
Proposal
The AI agent analyses the observation and proposes a concrete action, with its intended target and scope. In the proposer role the agent can describe what it wants to do; it cannot execute it directly.
Proposal carries Function, action, target, scope, the observation it rests on
-
3
Law check
LawGuard looks up the laws for that function and checks the proposal against them: is this action allowed for this function, on this target, within this scope? The laws are written for the client's own systems and kept separate from the shared core.
Law pack Function-specific permitted actions, targets and scopes; client-specific, versioned, kept apart from the reusable core
-
4
Authority
If the proposal is within the law, LawGuard applies the authority chosen for that function: either an individual approval by a named operator for this action, or an explicit standing delegation that lets the action proceed within the delegated limits. Outside the law, the action does not run.
Authority modes Per-action approval · Standing delegation within LawGuard limits · Chosen per function and per client
-
5
Execution and evidence
A permitted action is dispatched to the program that carries it out, and the result is recorded with it. Every proposal, decision, authority and result is recorded as evidence for operators and independent reviewers.
Evidence record Proposal, law applied, decision, authority used, approver where applicable, dispatch, result
Three standalone, reusable programs
Governor AI is a platform of standalone, reusable programs. Each component keeps its own identity and responsibilities and can be deployed on its own or with the others.
LawGuard
The enforcement component. LawGuard holds the function-specific laws, decides whether a proposed action may run, applies the chosen authority and records the evidence. Its releases are reproducible, so the artifact qualified in the lab is the artifact that is deployed.
- Function-specific law packs per system, program and function
- Decision at the point of action, outside the model
- Per-action approval or standing delegation within limits
- Reproducible, qualified releases
- Lab results: 14 test cases (90 subcases) passed on disposable lab machines; 19 cases passed on a replica of the pilot fleet
Governor Memory
Persistent, model-independent project memory. Memory keeps an immutable journal, reviewed knowledge with provenance, exact-text sources and ranked search, so the knowledge a governed system relies on does not live inside any one AI model.
- Immutable journal of what happened
- Reviewed knowledge with provenance
- Exact-text sources and ranked search
- Projects isolated by default; shared knowledge never auto-accepted
- Available as a library, a command-line tool and an authenticated service, reused by other Governor programs
Governor Code
A project-centric workspace where several AI agents work on one project inside a single durable, governed runtime. Code brings the same discipline to software work that LawGuard brings to operations: one project, one runtime, governed agents.
- Several agents, one project, one durable runtime
- Governed by the same platform principles
- Built on the shared Governor foundation
Who decides is a choice, recorded as an exact grant
Governor AI does not impose one answer to the question "who decides?". The authority for each function is a product choice made with the client and recorded as an exact grant that LawGuard enforces. Both modes can coexist in one system: chosen per function and per client.
Individual action approval
Each proposed action waits for an explicit approval by an authorized operator before it is dispatched. The approval, the approver and the result are part of the evidence record.
- Scope: the single action in the proposal
- Approver: a named operator mapped to the function
- Suited to: high-impact or rarely exercised actions
- Evidence: proposal, approval and result recorded
Standing delegation within limits
For routine, well-understood work, the client grants an explicit standing delegation. LawGuard still checks every action against the function's laws; within the delegated limits the action proceeds without a per-action prompt.
- Scope: an exact grant: function, actions, targets, limits
- Enforcement: every action still passes the law check
- Suited to: routine maintenance that must not wait for a human
- Evidence: the grant and each permitted action recorded
| Question | Individual action approval | Standing delegation within limits |
|---|---|---|
| Who decides each action | A named operator mapped to the function | The client, in advance, through an exact grant |
| Law check | Every proposal | Every proposal |
| Per-action prompt | Yes, before dispatch | No, within the delegated limits |
| Outside the law | The action does not run | The action does not run |
| What is recorded | Proposal, approval, approver, result | The grant, each permitted action, result |
| Chosen | Per function and per client | Per function and per client |
One reusable foundation, a bespoke system per client
Governor designs and builds an exclusive system for each client around the shared foundation. Governor AI is not a public download or a self-service product.
Reusable foundation
Each component is a standalone, reusable program. The shared core is one authoritative source; client-specific function laws, configuration, interface profiles, credentials and live state are kept separate from it. Nothing is forked or rebuilt per client.
- Core: shared, standalone programs with one authoritative source
- Client layer: law packs, configuration, profiles, credentials and live state kept apart
- Per client: no fork, no rebuild of the core
Bespoke delivery
Governor designs and builds an exclusive system for each client around the shared foundation. The system runs on the client's own infrastructure, or Governor operates and manages it.
- Customer-hosted: runs on the client's own infrastructure
- Governor-managed: operated and managed by Governor
- Engagement: a bounded pilot with measurable success and failure gates, then expansion gates
Lab results to date
LawGuard is qualified before it governs anything live. Every number here is a lab result, labelled as such.
The lab qualification had two parts. The first exercised the reproducible LawGuard release on disposable lab machines, where 14 test cases (90 subcases) passed with a verdict for each. The second exercised it on a replica of the pilot operator's fleet, where 19 cases passed.
A decision-only shadow comparison was then replayed over the pilot operator's real maintenance-bot records. It showed what LawGuard would have decided for each record and made no changes to the operator's systems. Production cutover at that pilot is planned and gated by independent review and owner approval.
Platform characteristics
Core architectural characteristics of Governor AI as it is built and qualified today.
Request a technical briefing
The decision cycle, the law model, the authority modes, and how a bounded pilot is set up for your fleet.
Request a Technical Briefing