Evidence-backed answers, or an honest abstention

Guardrails from the Lexora architecture run across every module. AI suggests. Humans approve.

Choose a deployment

Eight guardrails

How DuneSpan keeps AI honest

Grounded answers

Answers are built from your permitted sources, not from model memory.

Source verification

Each answer links to the page, drawing or record behind it.

Calibrated abstention

When evidence is weak, the system says so and routes to a person.

Observability

Every prompt, tool call and output is traced and reviewable.

Evaluation

Golden sets measure quality before and after every model or rule change.

Input and output guardrails

Filters for unsafe instructions, data leakage and out-of-scope requests.

Permissioned tools

Agents read, propose, or write with approval, through a credential proxy.

Human approval and audit

Nothing consequential happens without a person, and everything is logged.

Access follows the organisation spine

Scoped by entity, department, site and project

Access is default deny, then explicit role grants, then controlled inheritance down the organisation spine. A user scoped to one legal entity, one department and one facility retrieves nothing outside that scope, in dashboards, workflows or Lexora answers.

See the organisation spine

Platform controls

The same controls in every deployment mode

Tenant isolationNo cross-tenant data useLeast-privilege agentsFull audit trailVersioned models and knowledgeSSO, MFA and directoryRole-based access

What we never claim

That AI replaces design verification, testing or certification
That an estimation GA is a manufacturing drawing
Accuracy figures without a published test method
Security certifications we do not hold. Security documentation is shared on request.

Start here