Tests for factual traps and potential hallucinations
False-premise traps, source grounding, prompt-injection resistance and adversarial pressure to guess — each run through the configured chain and scored by deterministic checks.
One AI can be confidently wrong. IamVera is a multi-model AI verification layer for professionals. Claude answers, GPT fact-checks, Grok challenges and Perplexity verifies live sources — helping surface potential hallucinations and unsupported claims before they enter professional work.
Powered by
Built in the EU · Designed for professional workflows
Generative AI can be fluent, confident and wrong. For professionals who act on AI output, undetected hallucinations, unsupported claims and missing sources create avoidable risk.
"You wouldn't publish a single unverified source. Why would you publish a single unverified model?"
Verification only earns trust if it's honest about its own limits.
An AI answer verification layer for professional work. Vera cross-checks one answer through independent models that audit facts, challenge reasoning and verify live sources, then shows every correction, disagreement and source behind the final answer.
Not a single language model. Not a guarantee of truth. Not a replacement for professional judgement. Final responsibility for acting on an answer stays with you.
Every Vera answer passes through a configurable, auditable AI verification chain. No single model has the final word without independent factual review, adversarial challenge and source verification.
Claude formulates the initial response
GPT verifies every factual claim
Grok critically reviews and pushes back
Perplexity checks live sources, Claude synthesises
Claude (Anthropic) formulates the initial answer using its full reasoning capability. This becomes the baseline that every subsequent model evaluates — not a final answer, but a starting point for structured verification.
GPT (OpenAI) receives Claude's answer and independently verifies each factual claim. Corrections, confirmations, and confidence levels are returned — not as a replacement answer, but as a structured audit of the first response.
Grok (xAI) plays devil's advocate. It looks for missing context, ambiguous framing, and important distinctions that the first two models may have glossed over. Disagreement is a feature, not a bug.
Perplexity performs a live source check, grounding the answer in citable references. Claude then synthesises all model feedback into a final, consolidated answer — with every correction and source visible.
Every verification step is inspectable. You see the factual audit, adversarial challenge, AI source verification and final synthesis — including corrections, disagreements and supporting references.
We publish behavioural tests of the multi-model AI verification chain, including factual traps, source-grounding checks, prompt-injection resistance, raw run data and the cases where the chain fell short.
False-premise traps, source grounding, prompt-injection resistance and adversarial pressure to guess — each run through the configured chain and scored by deterministic checks.
Where a check fails or the models disagree, we show it. A benchmark that only ever passes proves little; the disagreements are where the chain earns its keep.
Every result links to the underlying JSON: model calls, corrections, costs and judge output. Nothing summarised away.
Configure the chain per session. See tokens and cost per model call. Verification depth is your choice — and so is the bill.
Get early access — pricing announced at launch →Vera supports privacy-preserving AI document analysis through the Semantic Privacy Shield. Before a document reaches any external AI model, detected sensitive values are replaced locally with synthetic, session-only equivalents. The approved synthetic context can then be analysed while the original values remain inside the protected Vera environment. This utility-preserving transformation retains relevant roles and relationships while reducing exposure of identifying data.
Names, dates, minors, diagnoses, criminal-law context, case numbers — replaced locally with realistic synthetic values. Roles and relationships stay intact, so the AI still understands the file.
A local check confirms no detected real values remain in the prepared text. If it is not clean, nothing is sent. Fail closed — no exceptions.
The verified answer is restored locally with the original values. You get a normal, usable answer — the public AI never saw the real data.
// Sent to the AI chain — fictional example
Mrs. Dylan Knoers [[DOC001_ADULT_PERSON_001]] states that her ex-partner Mr. Tobias Evers [[DOC001_ADULT_PERSON_002]] became aggressive on 06-01-2026 [[DOC001_DATE_001]] during the handover of her minor daughter [[DOC001_MINOR_CHILD_001]].
Built for professionals bound by confidentiality — lawyers, notaries, judges, physicians, occupational health physicians, psychiatrists, investigative journalists.
Vera Office combines privacy-preserving AI document analysis, multi-model verification and controlled AI document editing inside a browser-based DOCX workspace. Open a document and ask, for example:
"Summarise this document."
"Rewrite this paragraph more formally."
"Replace this phrasing throughout the document."
Vera processes the document content through the Semantic Privacy Shield. Sensitive values are replaced locally before the selected AI models analyse the document. Afterwards, the necessary values are restored locally.
For edits, Vera first builds a controlled action plan. Only validated changes are executed and saved — a model never gets unrestricted access to your file.
Document analysis, multi-model verification and controlled document editing in one secured working environment.
A verification tool that obscures its own data handling would undermine the very thing it promises. So we don't.
Documents are processed temporarily in memory for the active chat run. Only metadata is retained for session history. Files are not written to disk.
Model calls — including document content in the context — are sent to the configured AI providers via OpenRouter. This is stated clearly, not buried in a privacy policy.
Vera is built and operated within the EU. We are working toward full GDPR processor documentation. We will only claim compliance when the documentation is complete.
Vera is designed for work where AI output is acted upon — not just read.
A compliance officer uses Vera to verify regulatory summaries before they become internal policy. Every claim is traceable to a source — or flagged as unverified.
Full case study — coming soonA journalist uses Vera to fact-check background research before publication. Grok surfaces the contested claim; Perplexity finds the primary source.
Full case study — coming soonA researcher uses Vera to cross-check AI-generated literature summaries. The chain catches outdated statistics before they enter a published report.
Full case study — coming soonDirect answers about multi-model AI verification, potential hallucinations, source checking and confidential documents.
Use independent checks for factual claims, reasoning and sources rather than relying on the original model alone. Vera automates this by routing the answer through a configurable multi-model AI verification chain and exposing each correction, disagreement and source. The result remains subject to professional judgement.
Vera does not claim to identify every hallucination. It reduces the risk of undetected errors by having independent models audit factual claims, challenge assumptions and verify live sources. Conflicts are surfaced rather than hidden, so unsupported claims and missing evidence are easier to inspect.
A single model produces and judges its own answer. Vera separates roles across independent models: one formulates, another audits facts, another challenges the reasoning and another verifies sources. No single model has the final word without review.
Yes. Vera runs every model selected for the session in its configured role and records the resulting findings, corrections, sources, token usage and cost. The exact models can be configured per session.
When source verification is selected, Perplexity checks current public sources and returns references for the claims under review. Claude then incorporates justified findings into the final answer. The sources and model steps remain visible for inspection.
No. It means the answer passed through independent factual audit, adversarial challenge and live source verification, with corrections, disagreements and sources shown. Vera does not guarantee truth; it makes the verification behind each answer inspectable.
Vera's Semantic Privacy Shield transforms detected sensitive values inside the customer-isolated Vera environment before approved context is released to external models. A local privacy gate checks the outbound text and blocks transmission if verification fails. This reduces exposure but does not guarantee that every indirect identifier can be detected.
Vera surfaces disagreement rather than collapsing it into a false consensus. The final synthesis can incorporate justified corrections, preserve unresolved uncertainty and show which model raised each issue. You can inspect the evidence and remain responsible for the decision.
PDF files are not permanently stored by Vera. They are processed temporarily in memory for the active chat run; only metadata is retained for session history. Model calls, including document content in the context, are sent to the configured AI providers via OpenRouter.
Yes. You configure the verification chain per session. You can choose which models participate, see the tokens and cost per model call, and decide how much verification depth each question deserves.
Vera replaces blind trust with structured verification. Every step visible. Every source traceable.