£48,500.00
Above the delegated authority the role carries. Recorded with the failing check by name.
Every action the platform governs is checked before it executes; an action that fails its checks does not execute. The console below shows what that looks like: a proposed action, the checks it faces, the verdict, and the record the verdict leaves.
1 An agent asks
Delete a record
Policy PL-88214
2 The kernel checks it8/9
3 The kernel answers
Refused
The action never ran.
Retention rule did not pass
rec_01J9X4F1Q5 · ed25519:7a20…8ef1Signed. Verifiable without us.Mode
Replay of curated scenarios, labelled on the console.
Determinism
Same input, same verdict, every time.
Failure mode
Cannot decide means refuse. Never wave through.
Scenarios
Each one is a moment where an agent proposed an action and a rule stood in the way — or nothing did, and the action went through with a record behind it.
£48,500.00
Above the delegated authority the role carries. Recorded with the failing check by name.
Not on the case
The addressee is not on the case record. Nothing is sent until a person releases it.
No rule allows it
No rule allows it, so the kernel refuses. Absence of permission is not permission.
In bounds
Within authority, within bounds, in the right order. Allow is a verdict too, and it is sealed like the others.
Certification lapses
The certification expires before the job ends. Refused below the application, on every write path.
None of them is about what a model said. All of them are about what a system was about to do.
The kernel's question is the same each time: is this action, from this actor, within these rules, right now?
Illustrative. Synthetic scenarios in the shape the kernel records them. No personal data, not live records.
A refused action is recorded with the reason it was refused.
An action held for approval resumes by replaying only what was approved — nothing else.
Every governed action is sealed in a signed, tamper-evident record.
Where the platform gates a dispatch, the refusal is enforced at the database itself, on every write path.
Honesty
Until the isolated sandbox is connected, the console on this page replays the five curated scenarios. It says so on the console itself, in the note beneath the output, every time it runs.
We do not present a replay as the kernel, here or anywhere else on this site.
The console then drives a real kernel instance: isolated from everything else, loaded with synthetic industry packs, rate-limited, accepting no personal data, and behind a kill switch on our side.
Determinism
Same input, same verdict, every time. That is the property that makes a verdict evidence rather than opinion. A sampled model gives you a distribution; deterministic checks give you a fact you can put in front of an auditor.
When the live sandbox connects, this console gains a replay control: one scenario, submitted a thousand times, and a running count of distinct verdicts.
Start here
Bring one where an AI action worries you. You will watch the verdict arrive, and see the record it leaves.