When an answer matters, people need more than “the AI said so.” Qriton KB links answers to their sources, refuses unsupported claims, and gives teams a record they can check. It supports accountable human work; it does not make decisions or certify compliance.
The areas the AI Act treats as high-risk, and the reason each one ends up in front of someone asking how the answer was reached. Qriton supplies governed evidence to the accountable human; it is not an autonomous decision-maker.
Software that is a device, or a safety component of one, already carries a conformity assessment. The AI sitting inside it inherits that scrutiny.
Safety components in the management of traffic, water, gas, heating and electricity, where a wrong answer has physical consequences.
Recruitment, task allocation, promotion and termination. Decisions a candidate or an employee is entitled to have explained.
Creditworthiness scoring, and access to public benefits and emergency services. Refusals here are contested, and get contested in writing.
Assessment of evidence, risk of offending, and profiling in the course of investigation, where the record of how a conclusion was reached matters.
Visa and asylum examination, risk assessment and identity verification, under duties of documentation and review.
This comparison is about operational controls, not an unsupported blanket accuracy claim. “Baseline RAG” means the reference retrieval path in this demo, not every system built with RAG. Measure answer quality on your own corpus and evaluation set.
Each answered or refused chat can carry a receipt linking it to the cited evidence. Authorized users can check that record later.
This supports review and traceability; it is not a guarantee that an answer is correct.
Check a receipt yourself ›Supported answers carry inspectable cited passages. If the evidence bar is not met, it refuses and shows the nearest sections instead of inventing one. Extractive mode returns exact source text.
Answers carry a record of their cited evidence that authorized users can check later.
Approved changes follow release controls and can be removed from active knowledge without retraining.
Run it with a local open-weight model, extractive answers, or an explicitly configured hosted provider.
No model training is required. Point it at your documents, choose the deployment model boundary, and start reviewing cited answers.
It prepares and indexes documents by section.No model-training step is required.
Supported answers link to cited passages. Below the evidence bar it refuses and shows what it considered.Refusing is the feature, not the failure.
Answered and refused chats on an audit-enabled deployment are written to a tamper-evident chain that authorized users can verify against the deployment's audit log.The receipt shows provenance; it does not prove the answer is correct.
Use a non-confidential EU-regulation question to compare baseline RAG with the governed Qriton KB path, then inspect its receipt.