Software tier · in build
Attested Semantic Compare checks each output inside an attested private enclave, against a versioned rule library, on hardware you control. Your rules are never exposed in plain text to the AI side, and no data leaves your boundary for the check. Each verdict is bound to the library version that produced it.
Separation of knowledge
Your rules are never exposed in plain text to the AI side.
A swapped input, library or verdict fails verification.
No data leaves your boundary for the check.
Three outcomes
Inside the envelope. The action proceeds.
Governed convergence pulls the output back toward compliance rather than only refusing it.
Outside the envelope and refused. Nothing proceeds.
What is proven, and what is being tested
| Check | What it asks | Status |
|---|---|---|
| Range check | Is this state inside the approved distribution at all? Measurable without the experiment. | First, from mid-2027 |
| Distance check | How far is the output from the authored reference? This is the meaning check. | After the experiment |
| Governed-reasoning engine and ledger | The lifecycle behind the check: semantic modelling, constraint reasoning and continuous change detection. | Demonstrated in pinpole |
The distance check depends on a pre-registered experiment of about ten weeks, with its decision rule declared before any measurement. If it fails, the meaning check does not load; the rules tier and the range check are unaffected. pinpole, a live commercial product, demonstrates the engine on a structurally different problem.
Who it is for
Regulated entities
A bank or insurer running its own model on customer communications or claims notes, checking each output against a policy library it cannot share with the model's operator.
Operators
A defence or critical-infrastructure operator checking recommendations against a confidential rule set on its own premises, with no data crossing the boundary.
What it does not do
It runs in software. Physical non-bypassability is the hardware appliance.
Whether distance tracks compliance is exactly what the experiment tests. Until then, it is labelled a hypothesis.
A certified miss rate holds for the distribution it was calibrated on, not against an adversary outside a stated budget.
Evidence rung it fills
That an output met a declared policy, and what was checked against which library version.
Where the policy is about meaning rather than form, this is the producer for the attestation and audit rungs, with each verdict bound to the library version.
Related
The same check, on inputs
Everything bound for a model passes one governed path.
Semantic Gate →How we grade claims
Demonstrated, built, designed or hypothesis.
Status legend →The hardware rung
The meaning check on a hardware interlock, after the experiment.
The appliance →Get in touch
Thirty minutes on where your AI acts, what evidence it would need to leave, and whether this is the right rung.
Book a 30-minute conversation →