AI Disclosure

The EU AI Act Article 50 transparency obligations are in force since August 2026. People must be told when they are interacting with an AI system. This page explains how we meet that on this website and in the agents we build.

1. The sandbox on this website

• The Brewkraft demo agent is an AI system, labelled as such in its header before you type.
• It answers only from a fixed product knowledge base (brewkraft-store.json) and cites the entry it used.
• Refunds, payments, damage claims and order changes are blocked before any LLM call and routed to a human.
• Prompt-injection and discount-generation attempts are blocked on input; codes, offers and external links are filtered on output.
• Three test buttons run the same gate logic in your browser and use no model tokens; the order-status button runs on the server against fictional demo orders.
• Conversation content is not stored.

2. The 5 gates

• Gate 1 — Pre-screen: high-risk intents never reach the answering model. Clear cases are caught by multilingual patterns; the rest are labelled by a small classifier model, and the route for every label is fixed in code.
• Gate 2 — KB grounding: the model sees only the knowledge-base entries relevant to the question.
• Gate 3 — Source citation: every factual sentence must cite a KB entry, and every number or claim word (price, days, warranty, free) must appear in the cited entry — otherwise the answer is not sent.
• Gate 4 — Scope: questions outside the knowledge base go to a human instead of a guess.
• Gate 5 — Injection & output shield: instruction overrides, discount codes and foreign links are blocked.

3. Agents we build for clients

• Designed for EU AI Act Article 50 transparency requirements, including an AI disclosure badge in the widget.
• High-risk intents (refunds, payments, order changes) never reach the answering model and are handed to a human.
• Each agent ships with an eval benchmark on the client's own tickets before go-live.

4. Limits

Gates reduce risk; they do not make an AI system infallible. If an answer looks wrong, contact the store directly. Questions about this disclosure: hello@quietstack-team.com.