Verifiable AI (Pillar 8)
Last reviewed: 2026-04-20 · v2.2.1-D
ARIA's eighth pillar is the one most vendors skip: a hard guarantee that generated claims are backed by retrievable evidence. Below is a live demo plus a static example.
Live demo
Static example — grounded vs ungrounded
Grounded output
Per our 2026 compliance posture, GDPR Art 5(1)(b) [1] requires purpose limitation. Retention windows default to 90 days [2] unless the operator overrides per tenant.
[1] gdpr-regulation.pdf p12 · [2] internal DPA §4.2
Ungrounded output (blocked)
GDPR fines can reach $500M per tenant. All EU companies must delete data within 7 days.
No supporting citation found in knowledge base. ARIA refused to ship this response.
Attribution sample
FAQ
- What is Pillar 8?
- Pillar 8 is ARIA's Verifiable-AI pillar: every generated response is grounded against an authoritative knowledge source and citations are emitted inline. Ungrounded output is blocked, not shipped.
- How does grounding work?
- The grounding engine retrieves candidate passages from the tenant knowledge base, scores each claim against them via a peer-review model, and attaches citation markers. Claims without supporting evidence trigger an ethics-gate refusal.
- Can I export the grounding evidence?
- Yes. Every response produces a downloadable attribution PDF containing the exact source passages, confidence per claim, and the peer-review verdict.
- Is Pillar 8 exposed over MCP?
- Yes. MCP tool #20 'verify_claim' and tool #21 'attribution_report' let external agents verify ARIA output programmatically.