− From evidence to governed claims

How it works

− Architecture walkthrough

Explore the six stages

Select a stage to see what goes in, what Validix checks and what moves forward.

Evidence

Choose an authoritative source

A policy, label or study provides the source text, version and context.

Current policy · §4.2

− Illustrative product test

Change one detail.
See what the evidence still supports.

Use the controls to edit a draft based on a component test.

 

Source record · Test report TR-07

In laboratory testing, Component X operated at 120 °C for 30 minutes.

Setting
Laboratory
Temperature
120 °C
Duration
30 minutes

Edit the draft

In laboratory testing, Component X operated at 120 °C for 30 minutes.

Supported by this source

The draft keeps the tested setting, temperature and duration.

− Structure

From validated evidence to
governed claims Infrastructure.

Validix connects evidence to claims and governs what can be generated, verified and released.

It combines structured evidence, governed constraints and AI generation into a controlled claim workflow.

The model generates. Validix governs.
How Validix works

01 — Evidence

Everything starts with validated evidence.

Scientific documents, studies, approved materials and other trusted sources are ingested into Validix and transformed into structured, traceable knowledge.

The goal is not simply to store documents.

It is to preserve the evidence, its context and its provenance so that it can be used reliably when constructing claims

Every claim starts with evidence.

02 — Atomization

Evidence is structured into Atoms

Validix breaks evidence down into small, reusable units called Atoms.
Each Atom captures a specific, evidence-backed fact together with the context needed to interpret and use it correctly.
For example:

  • Treatment X
  • reduces
  • systolic blood pressure
  • by 23%
  • in adults
  • after 12 weeks

Each element remains linked to the evidence it comes from.
Atoms can then be reused across claims while preserving their source, context and traceability. 

The evidence stays connected, wherever the claim is used.

03 — The Cage

Govern what can be claimed

This is where governed claim construction begins.
The Cage defines the boundaries within which claims can be constructed. It determines which combinations of governed Atoms can legitimately form a claim and prevents unsupported combinations from becoming releasable claims.

The Cage does not tell the model what to write. It defines what the evidence allows the system to say.

The Cage constrains what can be claimed.

04 — Construct

Build claims from governed evidence

Validix uses your preferred LLM to construct candidate claims from governed Atoms. The LLM handles language generation. Validix provides the evidence, context and constraints that govern what can be expressed. You can work with the models you already trust through a Bring Your Own Key (BYOK) approach.

Different formulations can be constructed for different audiences, channels, markets or content requirements. But construction is not approval. Every candidate claim moves to Reverse Verification before it can be released.

Your LLM generates the language. Validix governs the claim.

05 — Reverse Verification

Check what the claim actually says

Before a claim can be released, Validix works backwards.
The generated statement is decomposed into its individual assertions and each assertion is checked against the governed evidence and Atoms.

The question is simple:

Does the claim actually say what the evidence supports?

Illustrative scientific claims and figures.

Validix does not decide what is true.
It governs what can be claimed from validated evidence.

Supported Claim

Treatment X reduced systolic blood pressure by 23% in adults after 12 weeks.

✓ Population supported
✓ Endpoint supported
✓ Magnitude supported
✓ Timeframe supported

→ PASSES CLAIM CHECK. Any required approval still applies.

Unsupported Claim

Treatment X significantly improves cardiovascular outcomes in adults.

✕ The endpoint is not supported by the available evidence.

✕ Statistical significance is not established.

→ NR — Not Releasable

The claim is not released because its assertions cannot be supported by the governed evidence.

06 — Release

Only governed claims move forward

A claim is released only when its assertions have passed verification. The result is not simply an approved sentence.

It is a governed claim with a traceable relationship to its underlying evidence.

Each claim can remain connected to:

  • its supporting evidence
  • its source
  • its governing constraints
  • its verification status
  • its history

This creates a claim layer that can be reused, maintained and monitored over time.

From approved once to governed continuously.

One evidence base. Many uses.

Once a claim is governed, it does not need to be reconstructed from scratch every time it is needed.

The same governed claim can support different content, channels, markets and formats while remaining connected to the same evidence base.

A change in the underlying evidence can therefore trigger a change in the claims that depend on it.

Instead of reviewing documents one by one, teams can manage the claims themselves.

One governed claim. Multiple controlled activations.

due diligence

The questions serious teams ask.

01. Does Validix stop the models from hallucinating?

Although the model can still generate an unsupported statement, Validix governs what can be released. This means that every regulated assertion is checked against the governed Claims and their underlying evidence, and Validix can withhold unsupported assertions or route them for review, according to the configured workflow.

Validix does not make the model truthful. It makes the release process governed and auditable.

02. Is this just RAG?

No. Vector RAG retrieves similar text chunks. Validix selects governed semantic Atoms according to approval, support, jurisdiction, time, policy and scope. Similarity can help find material; it is not the release predicate.

03. Why use a cryptographic hash?

A semantic hash identifies a canonical assertion bound to its source span. The output check also evaluates the applicable graph and policy state. It can show that reviewed meaning and its source span have not changed. A hash does not establish truth; evidence, governance and graph relationships do that.

04. What happens when evidence is missing?

The system fails closed. When support cannot be verified for the request, Validix can withhold the assertion or route it for review. Leading to No Response or a governed human-review path rather than improvised language.

05. Can it work outside pharmaceuticals?
Yes. The architecture is domain-general. Any market with controlled evidence, policies, jurisdiction, temporal validity and auditable release decisions can define its own ontology and predicates.

− a new category

Generative AI created a new production engine.
Validix creates its control plane.

− Continue exploring

Our Edge → 

Inspect a decision and see why a claim is blocked.

Manifesto → 

Follow the effects of a source change.