Limit 01
We don’t detect attacks
Prompt injection classifiers, content filters and anomaly scoring are not what this is. We remove capability. If you want detection as well, keep the layer you have. We assume it will miss things.
The Limits
What we don’t do.
Most companies publish what they can do. This page exists because being clear about the opposite protects both sides. An engagement that should never have started is worse for us than one we never won.
On The Product
What the control plane cannot do for you.
Every security control has a boundary. These are ours.
Limit 02
We can’t protect a compromised host
The threat model assumes the gateway host is trusted. If an attacker has root on the machine running the control plane, the control plane is not your problem.
Limit 03
We can’t fix a bad agent
If your agent produces wrong answers, enforcement does not make them right. It bounds what the wrong answer can cause. Those are different problems.
Limit 04
Cold start is real
A new agent with no observed behaviour has no learned policy. Shadow mode delivers the audit immediately, but enforcement waits for a confidence threshold. We will not enforce a guess.
On Services
Work we turn down.
Each of these has cost us revenue.
01
We don’t sell zero-human operation
Every system we build requires a named human owner with authority to intervene. Legal accountability requires a person. Anyone promising a fully autonomous department is describing something they cannot deliver safely.
02
We decline engagements below a sixty percent ceiling
If the honest autonomy ceiling for a department does not justify the cost of transforming it, we say so before you spend. A client who pays for a transformation that cannot deliver becomes a reference against us.
03
We don’t run multi-year programmes
Company-wide, multi-year change belongs to large systems integrators with the delivery scale for it. We do one department at a time, with a fixed scope and a fixed price.
04
We don’t provide legal advice
Our compliance work identifies which regulations apply and what obligations follow. It never produces the legal instrument that satisfies them. Where formal advice is required we coordinate qualified professionals rather than substituting for them.
05
We don’t take more than two engagements at a time
Delivery quality here depends on senior attention. If we are engaged, we will tell you when we can start rather than starting badly.
06
We don’t sell agent-building tooling
We govern agents and we build bounded systems. We are not a platform for other people to build agents on, and we do not compete with the frameworks you already use.
Not A Fit
Who we are not for.
Naming this saves everyone a wasted conversation.
| Not a fit | Why |
|---|---|
| Teams wanting a detection product | We remove capability rather than recognise attacks. If detection is the requirement, we are the wrong layer. |
| Organisations seeking zero-human operation | We do not sell it, and we will correct the expectation rather than take the engagement. |
| Companies without documented exception rules | Automation requires knowing what the exceptions are. Without them written down, the honest ceiling is very low. |
| Anyone needing enforcement on day one | Shadow mode runs first. A policy enforced before it is learned is a policy that breaks your agents. |
| Multi-year transformation programmes | We do one department at a time, with a fixed scope, rather than a programme nobody can scope honestly. |
On What We Claim
Things you won’t find on this website.
A company built on published error rates cannot decorate its own site with numbers it cannot evidence.
Absent 01
Accuracy percentages
No “94% accurate” claim. There is no independent benchmark in this category and a self-graded number means nothing.
Absent 02
“World’s first” anything
Unverifiable superlatives are the fastest way to lose a technical buyer. We describe what the system does instead.
If any of this rules us out,
that’s the page working.
And if it does not, you now know exactly what you are buying.