Context for the task
Connect the case material, instructions and relevant criteria. Give each agent only the sources and tools authorised for its task.
Loading operating surface
ContextClone / Safety & security harness
Frontier capability. Your rules.
Use OpenAI frontier models to build agent capabilities and workflows under your control. ContextClone brings the task, approved sources, permitted tools and human review into one governed process.
The work, in context
Every draft needs sources. Every decision needs an owner.
A permit check, a procurement review or a case summary follows rules. Design the agent’s task around those rules and the people who apply them.
Checks before decisions
Check claims against their sources. Flag missing information and questions that need professional judgement before a draft reaches its reviewer.
Control for the operator
Set the model, budget and approval steps. Define when an agent should continue, ask for help or stop.
Build at the frontier / Run within your boundary
Build with frontier models. Govern the run.
Develop on approved material with OpenAI frontier models. Review the release candidate before evaluating casework on a selected NVIDIA runtime.
01 / Build & evaluate
Build at the frontier.
ContextClone frames OpenAI frontier-model workflows around approved context, tools and review rules. Simulate tasks and evaluate the evidence.
- Tools & workflows
- Simulation & evaluation
- Policies & release review
Reviewed release candidate→Workflow · Models · Policies
Boundary review required
Boundary review required
02 / Run & govern
Operate under your authority.
Designed for NVIDIA GPUs with selected NVIDIA agentic AI and Physical AI world-model components.
- Operational data & context
- Local permissions & controls
- Evidence & human oversight
The ContextClone assurance loop
Prove it before it handles the work.
A useful agent needs more than a successful answer. Test the task, its sources, its limits and the people who can stop it.
- Build
Define the task, tools and permitted context with approved sample material.
- Test
Check expected outputs, source use, refusal rules and escalation paths.
- Simulate
Rehearse normal cases, edge cases and failure conditions before live use.
- Prove
Record what passed, what failed, the model version and the conditions of the test.
When evidence fails
Back down. Learn. Rebuild.
Stop or reduce the workflow’s scope, examine the failure and return to the build with a revised test.
When evidence holds
Validate. Approve. Secure.
Review applicable constitutional and legal requirements, policy and ethics. An authorised person approves the release; the team secures and tests the chosen runtime.
The task & its evidence
Keep every decision connected to its source.
A proposed answer is only one step. Context, checks and a responsible reviewer give it meaning.
Deployment review covers package contents, model provenance, logs, telemetry, support access and network policy. A signed package alone does not establish a data boundary.
ContextClone / Architecture direction
Task & policyApproved context
Agent workspacePropose · Check · Review
EvidenceHuman decision
Three deployment profiles / Architecture options
Choose where the work runs.
Three architecture profiles frame the deployment discussion. The installed configuration must demonstrate the controls your organisation requires.
Connected
- Environment
- Selected sovereign SE/EU cloud infrastructure.
- Network design
- Explicitly permitted, logged connections.
- Release path
- Reviewed network delivery.
Protected
- Environment
- Your controlled data centre and NVIDIA runtime.
- Network design
- Internal services; outbound restrictions defined and tested.
- Release path
- Approved internal distribution.
Sealed
- Environment
- Isolated infrastructure designed for air-gapped operation.
- Network design
- No external network path in the target design.
- Release path
- Inspected, signed offline media.
NVIDIA and Systown model choices, performance and available tools are evaluated for each profile. An air-gap or no-egress statement requires evidence from the installed system.
Start with a bounded task
Less time assembling. More time judging.
Start a demo with one workflow, approved material and the criteria your team uses to review it.
Municipalities & authorities
Check application completeness, assemble a records-request response or compare procurement material with published criteria.
Healthcare administration
Support administrative completeness checks and guideline retrieval, with patient-data access and professional review defined locally.
Government & regulated operators
Compare consultation responses, organise evidence bundles and prepare sourced drafts for a responsible reviewer.
Questions, answered
What to know before you start.
What is ContextClone?
ContextClone is a safety and security harness for agentic AI: the layer that defines a workflow’s task, approved sources, permitted tools, checks and human review. It uses OpenAI frontier models to help build agent capabilities under the operator’s control. The harness does not replace deployment testing or accountable human decisions.
How does ContextClone use OpenAI and NVIDIA?
OpenAI frontier models support development and evaluation using approved material. A release candidate is simulated, reviewed and evaluated for a selected NVIDIA-based runtime. Changing the runtime model requires evaluation of the actual workflow; it does not automatically preserve the frontier model’s performance.
Can ContextClone run air-gapped?
Sealed is the proposed air-gapped deployment profile. Activation requires verification of the installed models, tools, network isolation, logging, support and offline update process. A selected profile or a signed package alone is not evidence of an air-gap.
Does building with OpenAI expose sensitive operational data?
The architecture separates development on approved non-sensitive material from operational casework on the selected runtime. The project must verify what enters prompts, release packages, model weights, telemetry and support systems before making a no-egress claim.
One workflow. A clear evaluation.
Bring one workflow. Define a useful result.
Choose a recurring task and approved sample material. We’ll map its sources, review steps and deployment requirements, then agree what the evaluation must prove.