Agents that leave a receipt.
Every agentic action — traceable, measurable, accountable. Five layers of governance hold this up: Security, Policy, Guardrails, Compliance, and an append-only Audit ledger.
The summary below is a starting point your vendor-review team can use today; full artefacts are available on request.
What governance looks like in production.
Every request passes through four gates before it acts.
Agent actions follow a four-gate path — authenticated, policy-checked, executed, logged. A high-level view of the request path in production.
What we actually ship.
Your model keys, your control.
Bring your own Anthropic, OpenAI, Google, or xAI key at the org level. Swap or revoke any time — Kaiva uses your key only while you tell us to.
- BYOK for Anthropic · OpenAI · Google · xAI
- AES-256 at rest · TLS 1.3 in flight
- Keys scoped to tenant, cached briefly
Every action, reviewable.
Agent actions write into an append-only impact ledger — UPDATE and DELETE are blocked at the database. Operators post a reversal entry referencing the original, so the full history survives.
- Append-only · Postgres-trigger-enforced
- Reversal entries reference originals
- Actor · timestamp · outcome on every row
EU-resident by default.
Our production infrastructure is located in the EU. Self-hosted and multi-region deployments are on the roadmap — scoped per Order Form when customers need them.
- EU-hosted production
- No cross-region processing today
- Sub-processor list on request
Responsible disclosure.
Found something concerning? Email us directly and we'll coordinate a fix. One inbox, human triage, fixes called out in release notes.
- Coordinated disclosure · 24h ack
- Fixes called out in release notes
- Scoped test environments on request
No training on your data.
Kaiva does not train foundation models. Where enterprise options exist, we prefer no-training configurations with our model providers. Your inputs are subject to the provider's own terms — see the Privacy Notice.
- No Kaiva-side model training
- Prefer provider no-training settings
- Retention purged per policy
A human reads every escalation.
Disclosure reports, vendor-review questions, and incident queries go straight to engineers — not an L1 queue. First reply within 24 hours, weekdays, London time.
- One inbox · human triage
- 24h first response
- Fix status visible in release notes
A starting point your vendor-review team can use today.
// controls matrix with regulatory references · full artefacts on request
Vendor documents, on request.
Write to [email protected] for the current sub-processor register, DPA, and vendor-review artefacts. Formal audit reports (SOC 2, ISO 27001) are on the roadmap.
Write to us directly.
Vendor reviews, disclosures, coordinated incident response — one inbox, human read, 24-hour first response.
[email protected]