Lawyers
Use AI as a second reader on case files, assessments and pleadings — chronology, inconsistencies, missing substantiation — without sending client identities, minors or criminal-law context to public AI.
Semantic Privacy Shield
AI review — without exposing the case file.
Vera's Semantic Privacy Shield replaces sensitive case data with synthetic, session-only equivalents locally, before any public AI model sees your document. The AI analyses a synthetic version of the file. The real names, dates, diagnoses and case details never leave your environment — and the final answer is restored locally with the original values. We call this utility-preserving dehydration: preserve context, shield identity.
The problem
Your documents are exactly the ones you can't paste into a chatbot
Lawyers, notaries, judges, physicians, occupational health physicians and psychiatrists work with material that is protected by professional secrecy and data protection law: medical diagnoses, criminal records, family disputes involving minors, wills, assessments, re-integration files. These are the documents where AI assistance would help most — and where sending the original file to a public AI service is not an option.
Classic redaction breaks the analysis
Replacing names with empty placeholders like [PERSON_1] strips the context AI needs. Is this person a client, an ex-partner, a minor child, an heir, a suspect? Without roles and relationships, the analysis is safe but useless.
Sensitivity is semantic, not just a pattern
The risk isn't only names and ID numbers. It's an ex-partner, a psychiatric diagnosis, a criminal suspicion, a disinherited son, a whistleblower, a minor. A pattern-matching scrubber does not understand what makes a fact sensitive in context.
How it works
Dehydrate. Verify. Rehydrate.
The Shield works in three phases. All three run locally — the only thing that ever reaches Vera's public AI verification chain is the synthetic version of your document.
Dehydrate — locally
Your document is analysed in your local environment. Sensitive values — persons, minors, dates, locations, case numbers, medical and criminal-law context, family relationships — are replaced with realistic synthetic equivalents and session-bound tokens. Roles and relationships stay intact, so the text remains fully understandable for AI.
Verify — fail closed
Before anything is sent, a local verification step checks whether any of the detected real values still appear in the prepared text. If the check is not clean, the document is not sent to the AI chain. No exceptions, no silent overrides.
Rehydrate — locally
The AI chain works entirely on the synthetic version. When the verified answer comes back, the synthetic values and tokens are replaced with the original values — locally, using a private vault that never leaves your environment. You get a normal, usable answer.
What the AI actually sees
Same context. Different reality.
A fictional example from a family-law file. The structure, roles and relationships survive — the real people don't travel with them.
Original — stays local
Mrs. Jansen states that her ex-partner Peter became aggressive on 12 March during the handover of her daughter Emma.
Dehydrated — sent to the AI chain
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 Ivy Peterse [[DOC001_MINOR_CHILD_001]].
The AI still understands who has which role, what the relationship is and why the facts matter — it just never sees the real values.
Final answer — restored locally
The statements of Mrs. Jansen should be assessed in relation to the documented conduct of Peter during handovers involving Emma…
The same applies to medical and occupational-health material: a diagnosis, medication or incident is replaced by a synthetic equivalent of the same kind, so structure and consistency can still be reviewed without the medical details being exposed. Legal and medical terminology that carries meaning — statutory references, professional roles, legal concepts — is deliberately preserved.
Why ordinary anonymisation isn't enough
Utility-preserving dehydration: preserve context, shield identity
A lawyer, physician, psychologist or claims handler gets nothing useful out of [PERSON_1] and [LOCATION_1]. Professional analysis depends on meaning: was this person a minor at the time, are they a patient or a suspect, do they live in a relevant region, do they fall within a medical risk group or under a legal threshold? Blind anonymisation destroys exactly that meaning.
Vera is not a simple anonymiser. Vera is a meaning-preserving privacy layer for professional AI analysis.
Vera does not blindly strip everything. Attributes the analysis genuinely needs — age, sex, role, region, medical context, legal thresholds, relationships between parties, timeline — can be preserved where relevant, while the exact identifying values are replaced locally. What is preserved and what is withheld is recorded per run, so the choice is inspectable rather than implicit.
Dehydrated — what the AI chain receives (fictional)
Patient A [[DOC001_PATIENT_001]] — context: female, age band 40–49, region retained, relevant medical context present. Exact name, date of birth and address are shielded locally.
The record of what was preserved versus withheld is part of the local audit trail of every run (illustrative, simplified — the vault holding the real values never leaves your environment):
{
"token": "[[DOC001_PATIENT_001]]",
"label": "patient",
"fakeValue": "Patient A",
"preservedAttributes": {
"ageBand": "40-49",
"sexContext": "female",
"role": "patient",
"relevantMedicalContext": "present (category only)",
"locationContext": "region, not exact address"
},
"withheldAttributes": {
"dateOfBirth": "shielded locally",
"fullName": "shielded locally",
"streetAddress": "shielded locally"
}
} The claim, stated precisely: Vera minimises exposure of identifying data while preserving the legal, medical and factual context the analysis requires. We do not claim that re-identification is impossible — unique combinations of facts can remain identifying, which is why the preserved-versus-withheld choice is always inspectable.
The boundary
What stays local — and what leaves
A privacy layer only deserves trust if the boundary is stated plainly. Here it is.
Never leaves your environment
- The original document and its extracted text
- All real sensitive values: names, dates, locations, case numbers, medical details, criminal-law context
- The private vault mapping real values to synthetic ones
- The local verification results
Sent to the AI verification chain
- The dehydrated text: synthetic session values and tokens only
- Your question or instruction about the document
- Model calls are routed via OpenRouter to the configured AI providers — stated here, not buried in a policy
About the internals — a deliberate choice
We are radically transparent about the data flow: what stays local, what leaves, and in what form. We deliberately do not publish the internal detection and substitution techniques. The implementation is proprietary, it evolves continuously, and documenting its internals publicly would make it easier to craft inputs designed to evade it. Every run produces a local, inspectable audit trail — so you can always verify what was actually sent — without us handing out the blueprint.
Who it's for
Built for professions bound by confidentiality
If your files contain the kind of data that ends careers when it leaks, the Shield was designed for you.
Notaries
Check wills, deeds and powers of attorney for internal consistency and missing elements. Parties, dates and file details are shielded; legal concepts like usufruct, executor and statutory references stay intact.
Judges & court staff
Structure case material and reconstruct timelines across submissions without exposing parties, victims, suspects or minors to external AI services.
Physicians & occupational health
Have reports structured, summarised and checked for consistency while diagnoses, medication and incidents are replaced locally by synthetic equivalents. AI helps with structure and wording — not with your patient's identity.
Psychiatrists & psychologists
Work with psychological reports and treatment context — among the most sensitive data that exists — while the real person behind the file never reaches a public model.
Investigative journalists
Analyse leaked documents and source material, extract claims and build timelines while source identities and affected persons are shielded from every model in the chain.
Honest limits
What we claim — and what we don't
A verification company that oversells its own privacy layer would undermine everything it stands for. So here is the claim, precisely.
What we claim
Vera applies a local privacy layer that replaces detected sensitive values with synthetic session values before any public AI processing, verifies the prepared text against the detected real values, fails closed if that check is not clean, and restores the original values locally afterwards. Every run leaves a local, inspectable audit trail.
What we don't claim
We do not claim that every document is fully anonymised or that every sensitive detail is always found. Detection can miss subtle context, and unique combinations of facts can remain identifying. The Shield minimises exposure and enforces a protective process — it does not replace your professional judgement about what to share.
FAQ
Semantic Privacy Shield — common questions
The Semantic Privacy Shield is a local privacy layer for AI work on sensitive documents. Before a document is sent to Vera's public AI verification chain, sensitive values — names, dates, locations, case numbers, medical details, criminal-law context, minors, family relationships and more — are replaced locally with synthetic, session-only equivalents. A local verification step checks that no known real values remain; if it is not clean, nothing is sent. After the AI chain completes, the original values are restored locally before you see the final answer.
Classic redaction replaces sensitive values with empty placeholders, which strips the context AI needs to reason well. The Shield uses context-preserving synthetic substitution: real values are replaced by realistic synthetic equivalents that keep roles and relationships intact — client, ex-partner, minor child, heir, physician — so the AI can still analyse the document meaningfully without ever seeing the real data.
No. The public AI models in Vera's chain only receive the dehydrated version of the document — synthetic session values and tokens. The mapping between real and synthetic values is kept in a local vault that is never sent to any AI provider. If the local leak check finds real values remaining in the prepared text, the document is not sent at all.
No, and we do not claim that. Detection can miss subtle context, and unique combinations of facts can remain identifying even after substitution. What the Shield does guarantee is process: a local protection layer is applied before any public AI processing, a verification step checks the result against known real values, and the pipeline fails closed if that check is not clean. Professional judgement about what to share remains yours.
We are fully transparent about the data flow: what stays local, what leaves your environment, and in what form. We deliberately do not publish the internal detection and substitution implementation — it is proprietary, it evolves continuously, and publishing its internals would make it easier to construct inputs that evade it. Transparency about data handling, discretion about the mechanism.
Utility-preserving dehydration means Vera does not blindly anonymise. Attributes that professional analysis genuinely needs — age band, role, region, medical category, legal thresholds, relationships and timeline — can be preserved where relevant, while exact identifying values such as full names, dates of birth and addresses are replaced locally with synthetic session values. In short: preserve context, shield identity. Vera minimises exposure of identifying data while preserving the legal, medical and factual context the analysis requires.
Professionals who work with confidential case material and are often bound by professional secrecy: lawyers, notaries, judges and court staff, physicians, occupational health physicians, psychiatrists and psychologists, compliance officers and investigative journalists. Anyone who wants AI as a second reader without sending the original file to public AI services.
Use AI on your most sensitive files — without handing them over.
The Semantic Privacy Shield is part of Vera's early access programme. Verified answers, shielded data, every step inspectable.
Get early access