Viavitna's security model covers three things: what the AI is allowed to say, who is allowed to ask, and where your content lives. All three fail closed.
Every generated claim is checked for citation membership and number, date, and URL provenance, then reviewed by a separate semantic verifier for entailment and responsiveness. Unsupported material is rejected; if verification is unavailable or incomplete, Viavitna uses the extractive floor or abstains. Each decision carries a tamper-evident receipt. Full pipeline →
Your content is treated as untrusted reference text, never as instructions. If a model follows injected text, uncited output is rejected by the structural citation gate and semantically unsupported output is rejected by the verifier.
| Key | Where it lives | What it can do | Defenses |
|---|---|---|---|
pub_… publishable | Embedded in your website | Ask questions of your tenant only | Origin allowlist, per-IP rate limiting |
sec_… secret | Your servers only | Ask questions server-to-server | Never in a page; shown once at publish |
| admin | Your operators | Upload, test, publish, configure | Failed attempts throttled and logged; clickjacking-protected console |
The models are open-weight and run on infrastructure we or you operate. Your content and your customers' questions are never sent to an outside AI provider.
The cloud plan runs entirely on our servers — you add one script line to your website and that is all. The appliance runs entirely inside your own network; your content and your visitors’ questions never leave it. No inbound access, nothing installed on your systems.
On-premise deployments ship as a sealed virtual machine whose disk is AES-XTS encrypted. Without its license key, the appliance — models, content, engine — is ciphertext.
Temperature 0, fixed seed, schema-constrained output: the assistant's behavior is reproducible and auditable, not probabilistic theater.
Infrastructure failures degrade answers to verbatim passages — never to guesses. Only truth fails closed.
Before going live, every source file was reviewed line by line and the deployment was tested adversarially — authentication, injection, traversal, oversized inputs, transport security, clickjacking. Findings were fixed and re-verified against the live service, and the deployed code is hash-verified against its exact source revision on every release. Backups are encrypted and stored offsite, and restores are tested by actually restoring — not by trusting success logs.
Viavitna