TrustFabric demo
Stop secrets and restricted data reaching an unapproved model
An AI coding agent asks the gateway before sending code to a model provider; the policy and the sovereignty profile decide, and the decision is recorded as evidence.
The challenge
Coding agents send source code and context to model providers on their own, and a team cannot show which data went where or under which rule.
The scenario
A platform team lets developers use an agent runner they operate. A developer's task would send a file containing an API key, from a repository classified as restricted, to a model provider that is not on the EU profile's approved list.
What goes in
- Approved policy bundle
- The rules for model providers, repositories and secrets, approved by a second person.
- Sovereignty profile
- Which providers, endpoints and regions each data classification may reach.
What you can ask
Illustrative prompts. Results depend on your own data and are not deterministic.
- “Can this agent send this repository's code to this model provider?”
- “Why was the agent's last tool call denied?”
How Clearception approaches it
How the work is structured. Each step shows what its output is grounded in.
The agent asks first
ObservedBefore the tool call runs, the runner's guard sends the action, resource and context to the gateway.
Secrets are detected
ObservedCredential detection runs on the request before any policy is evaluated.
Policy and sovereignty decide
ObservedThe rules and the profile return deny, with the reason code and the rule that matched.
Evidence is written
SourceThe decision is appended to the organization's hash chain; if that fails, the answer is deny.
What you get
The shape of what comes back — not a promised result.
Decision with its reason
ObservedDeny, the reason code and the matching rule, returned to the agent before it acts.
Evidence record
SourceA chained entry that an auditor can verify and export.
Why it matters
- Prevented, not just logged
- In a runner you operate the call does not happen unless the gateway allows it.
- Explainable
- Every denial names the rule, so a developer knows what to change or whom to ask.
Related demos
TrustFabric
Gate a production release on policy and a human approval
A pipeline asks TrustFabric before deploying; missing checks fail the stage, and a production deploy waits for an approver who did not request it.
3 minute walkthrough
- Deterministic Policy Engine
- Human Approval Gates
- Evidence Trail
Lawgorithm
Take in and organise a discovery set
Deduplicate, classify and index discovery material on arrival, with provenance recorded.
4 minute walkthrough
- Document Intelligence
- Chain of Custody
- Evidence Trail
Lawgorithm
Find the material that matters
Search and group a large document set by entity, topic and date without reading everything.
4 minute walkthrough
- Document Intelligence
- Entity Resolution
- Evidence Trail
Want to try this with your own work?
Open TrustFabric and bring your own data. No signup is needed to read these demos.