Live demo · AI governance
A trust layer between every person and every model.
For governments, universities and companies that want the best AI models without losing control of their data: one vendor-neutral layer that checks who is asking, classes the request, grounds it in documents the person may read, takes personal data out before anything leaves, routes it by tier, withholds what the sources don't support and holds risky actions for a named approver. The policy is a file with tests you can run, and every request leaves a sealed evidence record. Press play for a narrated run of about seven minutes, or use it by hand. It runs in your browser on an invented university.
What the run shows
One morning of requests, and the rules behind them.
Most organisations are stuck between sending sensitive data to commercial models and building everything locally at a fraction of the capability. The layer in between is the cheapest part of the stack to build, every request already passes through it, and it is the only part that survives a change of model vendor. The run follows one morning at an invented university through each of its controls.
01
Public work on the best model
An admissions FAQ is classed public from its text and its source, so it goes to a commercial frontier API. Nobody chose the model; the tier did, and every sentence comes back with its source.
02
A planted instruction taken out
A note attached to the board minutes tells the model to e-mail the grade sheet out. Grounded retrieval quarantines the passage before any model reads it, and a sentence no document supports is withheld.
03
Personal data leaves as tokens
A letter about a named student with her CNIC is Tier 2, so it runs on an open-weight model hosted in Pakistan, and even then her name and numbers leave the building as tokens and come back inside it.
04
Classified stays in the building
An exam paper before the exam is Tier 3: an air-gapped model on the exam cell’s own server, with nothing sent out at all.
05
Refused at the first control
A lecturer asking for a medical file he is not entitled to is stopped before anything is searched or any model is called, and the refusal is still on the record.
06
Agents wait for a named person
Tools are on an allowlist, and a bulk e-mail to 412 people is held until the Director Finance approves it. His name goes into a new evidence record.
07
Sovereign compute, sized right
Thirty days of traffic by tier show that only the regulated and classified share, about 15%, needs models inside the country or the building.
08
Policy as code, with tests
The rules are one policy file. Twenty-four conformance tests run against the live engine, still pass after the model vendor is changed, and fail the moment a control is switched off.
09
Evidence a regulator can check
Every request writes the prompt as it left, the sources, each decision, the model and version, the output, the approver and the time, sealed with SHA-256 and chained, so an edit made afterwards shows.
Simulated here
What the demo pretends.
- /The university, its people, documents and students, the traffic figures and the model vendors are invented. No real model is called: the replies are played by the demo, and typed requests are answered from the passages they retrieve.
- /The controls are not pretended. Identity, classification, retrieval, redaction, routing, attribution, the action rules, the 24 tests and the SHA-256 chain are real code running in your browser.
- /Nothing is sent anywhere: the approved e-mail is queued in the page only, and the evidence export is a file made in your browser.
In a build for you
What becomes real.
- /Your sign-in and roles, your document stores and their labels, your detectors for national ID, health, finance or classified material, and your own tiers.
- /The models you choose behind each tier: commercial APIs with zero retention, open-weight models in a data centre in your country, or a server with no network out.
- /The policy as a versioned file with its test suite in your release pipeline, and the evidence log on write-once storage for auditors and regulators.