Standardise governance across every AI system you run.
One evidence model, inside your own perimeter, for the whole estate. What one surface reports, another can prove, and an auditor can check without us in the room.
In brief
A one-minute read for the person deciding whether the rest of this page is worth their time.
- What it is
- A governance runtime that sits beside the AI systems an organisation already runs. It records what each operation was allowed to do and what it actually did, and turns that into evidence a third party can check. It installs into existing tooling rather than replacing it, and it runs entirely inside the organisation’s own perimeter.
- Why organisations standardise
- Because governance assembled from separate tools produces separate records, and reconciling them is the work. One evidence model across every AI system removes that work permanently. Executive, audit, engineering and compliance all read the same record rather than keeping their own.
- Why it is different
- It refuses to answer questions it cannot support. There is no composite trust score, no severity ranking and no trend line, because nothing in the platform computes them. What it reports instead is what it examined, what it could not reach, and why. That is the part that survives an audit.
- Deployment
- Your infrastructure. Four postures.
- Data egress
- None required to operate.
- Starting scope
- One agent, one repository.
- Operator requires
- No specialist role.
Three things become true.
One evidence model, not one per tool.
Every AI system writes into the same records, so governance does not fragment as the estate grows.
Every claim carries the file that proves it.
A finding is not a statement about the past. It is a record with a digest, and the digest can be recomputed.
Verification does not require us.
Records export in open formats and verify offline. Removing VerifAIer does not invalidate what it already produced.
Nothing to flag.
●Ready
16 capabilities were checked. 14 answered. Two could not be reached, and are reported as unknown rather than as failures.
14 / 16 answering
flight and governance declare no operation an assessment can run, so they answer nothing. A ratio, not a grade. No composite score is computed anywhere in this platform.
Inside your perimeter. All of it.
No outbound call is required to answer.
- Server-rendered HTML
- No client-side framework, and no build step in the path of a page.
- Works without JavaScript
- Navigation is links. Mutation is form posts. The product is usable with scripting disabled.
- No external asset
- No CDN, no webfont and no third-party request from any product surface.
- Synchronous execution
- An assessment completes before the page renders. There is no queue and no worker to operate.
- Fixed surface area
- 21 destinations in 9 groups, one route table, asserted by the test suite.
Four postures, stated plainly.
Self-hosted
Runs on infrastructure you already own and already govern.
Disconnected
No outbound dependency exists to remove, so an isolated network changes nothing.
Container image
A packaged container path is not yet declared. This page will not claim one until it is.
Hosted service
There is no hosted service in the path of an assessment, and no plan to put one there.
Two vocabularies sit in that row and answer different questions. AVAILABLE and PENDING are maturity: whether a capability exists yet. BY CONSTRUCTION and NOT OFFERED are deployment posture: how a capability is delivered, or that it is deliberately not. A posture is never a maturity, and NOT OFFERED in particular is a decision rather than an open question. Deployment specifics sit on the Trust Center.
From operation to something an auditor can check.
An agent is about to act. The runtime sees it first.
The operation is sealed into a record with a digest over its bytes.
A receipt is issued against the envelope and bound to its hash.
Posture is derived from the records, never from an opinion about them.
Records leave in open formats and verify without the platform present.
Flight Recorder
Keeps enough of the run to re-examine a decision later.
Agent Passport
Says which agent acted, under which identity and version.
- Bound envelope
ev_bf6442d03d1d4…89a3b708bbd19d28- Receipt
rcpt_711e51de6f0…b3c82256fdeca7c3- Produced by
evidence_engine- Verification
- MATCH
Policy packs, adopted deliberately.
eu_ai_actEU AI Act
Ships today, as metadata your organisation adopts. Nothing is enforced without you adopting it.
nist_ai_rmfNIST AI RMF
The same evidence, read against a different framework. No second ledger is created.
iso_42001ISO/IEC 42001
Management-system alignment, drawn from records that already exist.
Governance Simulator
Ask what a policy would have done, before it governs anything real. Exploratory by design, and never a verdict.
Three compliance packs ship today, and the same evidence is read against each of them. Custom packs are declared in the plan model and are not yet implemented. SSO is the other one, and both are named in entitlements/models.py as unimplemented rather than advertised as available.
Access is stated, not implied.
Roles and permissions
Actions a role cannot take render as unavailable with the reason on screen, rather than disappearing.
⊘Not available to you
Selective Disclosure
Prove a property of a record to an outside party without handing over the record itself. Architectural selective disclosure, not a zero-knowledge proof system, and it says so where it is implemented.
Entitlements
The plan model distinguishes a feature that is not in your plan from one that is in your plan and not yet built. Both are shown.
Not in plan In plan · not implemented
Data movement
None is required. Evidence is produced, stored and read inside the perimeter it was produced in.
Four modes, and what none of them claim.
Quick
A first read of the estate, on demand.
Standard
Broader capability selection, same evidence model.
Deep
Widest selection. Still a selection, and it says which.
Continuous
Planning semantics for repeated invocation. There is no scheduler, worker or queue behind it, and selecting it does not switch monitoring on.
- A mode name is not a verdict.
- COMPLETE means the selected mode succeeded, not that the platform was assessed.
- No mode computes a score of any kind.
- Findings carry no severity, because nothing in the platform ranks them.
Runs are local and synchronous
By the time the page renders, the run is over.
Reads are bounded
Evidence tables stop at 25 rows by design, and say so.
Results are not persisted
Which is why the platform refuses to report a trend.
One agent first. Never the estate.
- Run an assessment
- No account, nothing installed, nothing to configure. It finishes before the page renders.
- Install the runtime beside one agent
- One team, one editor, one codebase. The blast radius of a mistake is one repository.
- Read what it produced
- Receipts, envelopes, provenance. Verify one offline before trusting the rest.
- Adopt a policy pack
- Read the same evidence against the framework your auditors already use.
- Give executives the read
- A situation, a confidence and one recommendation. Nothing they have to interpret.
- There is no packaged Sentinel installer. Three surfaces install; the rest do not.
- No composite trust score exists, and the product refuses to compute one.
- No baseline is stored, so improvement over a period cannot be reported.
- Marketplace distributes metadata. It never executes or downloads anything.
- Billing is not configured. Nothing on the pricing page can be purchased yet.
Governance nobody uses is not governance.
Most governance platforms are adopted by a programme and operated by specialists. This one is adopted by a team and operated by whoever is already there.
- A cluster to provision. The runtime installs beside an agent, not in front of one.
- A data migration. Evidence is produced where the operation happens.
- A specialist operator. An assessment is one action with nothing to configure.
- A training programme. Every screen states its own limits in plain language.
- A consulting engagement to begin. The first assessment needs no account at all.
- One editor with the runtime installed. Two install today.
- A decision about which policy packs apply to you.
- Somebody willing to read one screen. Once.
The reason this matters commercially: a governance layer that only a specialist can operate is a governance layer that covers whatever the specialist has time for. Coverage is a usability problem before it is an architecture problem.
- Under 1s
- An assessment completes before the page renders. No spinner exists.
- No account
- The first assessment runs without signing up for anything.
- 1 command
- Installing the runtime beside an agent is a single install script.
- 0 bytes out
- Nothing leaves the perimeter. There is no hosted service in the path.
Every door opens onto the same runtime.
An editor, a browser, an API and a terminal are four ways into one architecture, not four products to license, integrate and reconcile. Each door carries its own maturity, stated plainly.
The editor
An MCP server beside the coding agent. Claude Code and Cursor install today.
The workspace
21 destinations in a browser. Server-rendered, no client state.
The API
12 open operations, named in one allowlist. Everything else needs a credential, and the allowlist is what makes an endpoint open.
The SDK
11 typed services. It composes the runtime; it never reimplements it.
The terminal
The vai console script and python -m vailidator.sdk. Both ship in the wheel.
The desktop
A native application. Planned for VerifAIer Me, after v1. No date.
The phone
App Store and Google Play. Planned for VerifAIer Me, after the desktop build.
The mark appears at every door because the same architecture is behind each one. It is an identity, not an interface, and nothing is verified by looking at it.
Six steps. No architectural change at any of them.
Each step is a decision about scope, not a deployment project. Nothing is rebuilt, re-platformed or re-integrated when an organisation moves from one to the next.
- One agent
One editor, one install command, one person.
no approval needed - One team
The same install, repeated. A shared workspace appears.
+ shared evidence - One department
A policy pack is adopted deliberately. Rules become explicit.
+ policy - Business units
Scopes separate. Access is stated per unit rather than inherited.
+ scope boundaries - Organisation-wide
Executive surfaces become useful because coverage is broad enough to read.
+ executive view - Sovereign scale
Disconnected posture, national or sector policy packs, no change to the model.
+ deployment posture
Step six runs the same runtime as step one.
A receipt produced on day one is readable on day one thousand.
Packs are added. The way rules are stated does not change.
Inside your perimeter at every step, including the first.
There is no step at which an organisation must stop and re-platform. That is the property being described here, not the speed of any particular rollout, which depends on your organisation and not on ours. It is not a service commitment, and no date on this page is one.
The cost is in the seams.
Most of what an AI governance programme costs is not the tools. It is the work of holding disconnected tools together: reconciling records, maintaining integrations, training people to operate each one, and paying somebody to explain the result.
One architecture removes that work rather than automating it. There is nothing to reconcile when there is one record.
No ROI figure, no payback period, no cost-saving percentage and no benchmark against other tooling. We have not measured your organisation, and a number we cannot show the evidence for would contradict everything else on this page.
- Reconciliation between tools
- One record. Nothing to align, and nothing to explain when two systems disagree.
- A governance headcount
- The budget line for a dedicated governance role does not open.
- Consulting to begin
- No statement of work stands between the decision and the first result.
- A migration project
- Nothing moves. It installs beside what already runs.
- Retraining per product
- Adding a capability adds a surface, so the training cost is paid once.
- A separate audit exercise
- Evidence is produced as operations run, not assembled afterwards.
Start with something you can verify yourself.
An assessment needs no account and touches nothing. Read what it produced, then decide whether the rest is worth a conversation.
Local-first and advisory. Findings are evidence about what happened, not certification and not legal advice.