Trust through mechanics — not borrowed badges.

Principles

• Tenant / organization isolation in the product data model
• Source permissions and least-privilege retrieval intent
• Role-aware access to workspace surfaces
• Source-level traceability for material claims
• Expert-note attribution and review state
• Human approval for consequential actions
• Audit history for sensitive lifecycle events
• Data deletion and retention treated as customer policy, not marketing slogans
• Secrets managed via environment configuration — never embedded in the public site
• Model-provider boundaries: private corpora are not training data
• Public analytics avoid free-text PII

Current safeguards

• Session cookies with HMAC integrity checks in middleware
• Membership-scoped workspaces in the application core
• Role checks on action approve/execute paths with deny audit events
• Cross-tenant isolation covered by automated tests in the monorepo
• Public demo uses synthetic fixtures only — no customer uploads on the marketing form

Permission and tenant model

Workspaces are organization-scoped. Source permissions are inherited from connected systems and enforced at retrieval time — restricted records never enter an answer and are not merely filtered after the fact.

Source permission inheritance

When a document, record, or knowledge entry is connected, its access group travels with it. If a user cannot access the underlying source, the derived claim, citation, or answer does not reveal it.

Auditability

Questions, evidence selected, review decisions, and action previews are recorded. Operators who secure the workspace can replay the retrieval and decision lifecycle.

Retention and deletion

Retention is treated as a customer policy in production. Marketing site analytics are retained only as configured by the analytics host; no free-text questions or PII are sent.

Model-provider boundaries

Deterministic product paths do not require leaking private corpora into model training. Where models are used, tenant data is not retained for model improvement.

Under development

• Production multi-tenant hosting hardening for app.veldarium.com
• Broader connector permission mapping
• Customer-managed retention policies in production
• Expanded evaluation and monitoring for retrieval quality

Planned

• Enterprise identity integrations beyond the OIDC skeleton
• Formal third-party security review when the product stage justifies it
• Documented subprocessors list for production SaaS

What we do not claim

• SOC 2, ISO 27001, HIPAA, FedRAMP, or equivalent certifications
• Completed formal penetration testing or external audits
• Regulatory compliance for any specific industry regime
• Production security maturity equal to large horizontal platforms

Security contact

Questions: chris@veldarium.com