Clear, self-contained answers about AI answer verification, multi-model fact-checking, potential hallucinations, source verification, confidential documents and controlled AI document editing.
Direct answers about what AI answer verification means and how IamVera applies it.
AI answer verification is the structured review of an AI-generated answer for factual claims, weak assumptions, reasoning gaps and missing sources before a user relies on it. IamVera turns this review into an inspectable workflow rather than leaving the original model to assess its own output.
IamVera can route an answer through initial generation, independent factual audit, adversarial challenge, live source verification and final synthesis. The configured steps, findings, corrections, disagreements and sources remain visible in the verification trail.
In Vera, “verified” means that an answer has completed the configured verification workflow. It means the checks and their results are inspectable; it does not mean that the answer is guaranteed to be true or complete.
No. Vera does not guarantee truth or correctness. It is designed to reduce the risk of undetected errors by making independent checks, corrections, source gaps and model disagreements visible. Final professional judgement remains with the user.
How independent models cross-check answers and expose potential AI hallucinations.
Multi-model AI verification uses different AI models for separate roles in one review chain. One model may produce the initial answer, another may audit facts, another may challenge the reasoning and another may verify current sources.
Different models can fail in different ways. Separating factual audit, adversarial review and source verification makes unsupported claims, false premises and missing evidence more likely to be noticed. Vera does not claim that every hallucination will be detected.
Vera sends the answer and relevant context through every model selected for the session. Each model performs its configured task, and justified findings are incorporated into the final answer while the original reviews remain available for inspection.
Vera surfaces the disagreement instead of hiding or averaging it away. A conflict can reveal uncertainty, missing context, an unsupported claim or a question that still requires human review.
How Vera checks claims, current sources and fabricated authorities.
A selected reviewer examines factual claims, internal consistency, assumptions and missing context independently of the model that produced the first answer. The findings are recorded and can trigger a justified revision.
When source verification is selected, Perplexity checks relevant claims against current sources. Citations, corrections and unresolved source gaps remain visible so the user can inspect the basis of the final answer.
Vera can challenge an unverified authority and search for supporting evidence before accepting it. Published tests include a non-existent Dutch Supreme Court ruling. Detection is not guaranteed, so users should still inspect the cited sources and identifiers.
No. Source verification can expose missing, outdated or unsupported claims, but it does not replace legal, medical, academic or other professional assessment of relevance, authority and context.
Direct answers about pseudonymisation, external models and the Semantic Privacy Shield.
Vera is designed to reduce unnecessary exposure when professionals use selected external models with confidential documents. Document or selection context must first pass through the Semantic Privacy Shield and its privacy verification gate.
Local pseudonymisation means that detected sensitive values are replaced inside the customer-isolated Vera environment with synthetic session values before approved context is transmitted to external AI providers. Pseudonymised data can still be personal data and is not the same as guaranteed anonymisation.
The Semantic Privacy Shield is Vera’s protection workflow for document and selection tasks. It transforms detected sensitive values, verifies the prepared outbound context and can restore original values inside the protected Vera environment after the model workflow.
Vera fails closed. If the privacy gate does not approve the prepared context, or the Shield is unavailable, no document or selection content is transmitted to external AI providers.
How Vera Office separates document analysis, privacy protection and controlled changes.
Vera Office is a browser-based DOCX workspace for authorised collaboration, privacy-preserving AI document analysis and controlled document editing. Analysis uses the selected verification chain; document changes use a separate validated action route.
Yes. A task can use the open document or a selected passage. The selected context must pass through the Semantic Privacy Shield before any configured external model receives it.
No. A write instruction is converted into a constrained action plan, validated and executed through the Office bridge. A model does not receive unrestricted write access to the document.
Yes. Every model selected for the session is called for its configured role. A reviewer with no finding may avoid an unnecessary additional revision step, but the reviewer call is still executed and recorded.
Clear boundaries around hallucinations, GDPR claims, evidence and professional responsibility.
No. Vera does not eliminate hallucinations or claim to be hallucination-free. It helps detect potential errors and makes the verification process more visible and inspectable.
IamVera does not use a generic GDPR-compliant label as a substitute for evaluating purpose, roles, contracts, providers and international transfers. The service is designed for privacy-conscious and GDPR-sensitive professional workflows, with detailed data handling documented separately.
IamVera publishes behavioural tests, limitations, failures and raw JSON data at iamvera.ai/evidence/. The published runs include factual traps, source grounding, prompt-injection cases, deterministic checks and language-model judge findings.
The user remains responsible. Vera supports professional judgement with inspectable verification and privacy controls, but it does not assume legal, medical, academic or other professional responsibility.
IamVera is built around independent inspection rather than blind trust in one fluent answer.
| Question | Ordinary AI chat | IamVera.ai |
|---|---|---|
| What is the output? | A direct model answer. | An answer with an inspectable verification trail. |
| Who checks the answer? | Usually the same model or the user. | Independent selected models perform factual, adversarial and source-checking roles. |
| What happens to disagreement? | It may be hidden or smoothed away. | It remains visible as relevant evidence for the user. |
| How are confidential documents handled? | Users may paste original sensitive text directly. | The Semantic Privacy Shield transforms and verifies approved context before external processing. |
| Does it guarantee truth? | No. | No. IamVera makes verification visible, but final judgement remains with the user. |
IamVera is for professionals who need more than one unchecked AI answer.
IamVera is designed for professionals who work with high-trust or confidential information, including lawyers, notaries, occupational physicians, journalists, researchers, compliance teams and security-sensitive organisations.
Inspect the published tests, limitations, failures and raw data behind IamVera’s multi-model AI verification chain.