AI security, policy and audit
Control every crossing.
Is this execution permitted, and under what conditions?
Protect sensitive information, enforce environment and route policy, and preserve a reviewable decision trail for every interaction between your systems and AI.
Current maturityDemonstrated in the browser · available in pilotsDeployment noteRequires deployment-specific hardening, scoped during the pilot.
Protect information · Control execution · Prove what happened
Modules
Everything the demonstration on the homepage does maps onto these modules; the same model is designed to govern Cortex, Forge and Axon.
Sensitive information is found and neutralised before it crosses.
Explicit, versioned rules decide what may cross and how.
Requests reach only approved destinations, by policy.
Each decision leaves a record an operator can review.
Enterprise policy management and production integrations around the boundary.
Run the synthetic demonstrationDeterministic, browser-local, fictional data.
A masking layer changes the payload. Boundary decides whether the crossing is allowed, what must change, where execution may occur, and how the decision is recorded. The demonstration above shows both outcomes: an authorised crossing and a refusal.
Boundary decides whether and where a task may run.
Forge decides which permitted model should perform it.
Together: controlled execution.See Kordane Forge ·the evidence each decision leaves.
How it works
Eight stages, in order, every time. The public demonstration runs this exact sequence on synthetic data. Above it sits the portfolio-level decision lifecycle:
Understand the taskClassify the informationApply environment policyEvaluate candidate modelsSelect the smallest sufficient approved modelExecuteEscalate if thresholds failRecord the decision
Decision lifecycle. Each step inherits the maturity of the product that implements it; end-to-end automation is configured per deployment.
Sensitive values are found in the request before anything moves: identifiers, codewords, references, contact channels.
Each value is assigned a class your policy can reason about: identity, document, contact, codeword, reference.
The active profile decides per class: pass, tokenise, redact, block, or contain. Decisions are explicit and versioned.
Values that may cross in shape but not in substance are tokenised; values that must never cross are redacted at source.
The request goes only to an approved destination: external model, private deployment, or a local model with no egress.
The downstream answer returns through the boundary, carrying tokens instead of your originals.
Where policy permits, tokenised values are restored locally, inside your environment, after the crossing.
Each decision lands in a detailed audit trail: what was found, what policy did, where the request went, what came back.
Boundary Audit
The decision record a single synthetic crossing leaves behind.
Operational briefing (synthetic fixture)
Every crossing produces a reviewable record: what was found, what policy did, where the request went.
Illustrative decision record · Synthetic data · Representative of current product capability
Boundary is the governance layer of the portfolio: Cortex, Forge and Axon are designed to route their crossings through it, so one policy model and one audit trail cover intelligence, models and conversations.