ZillaAI

What it is

One small compound of buildings glowing alone in an enormous expanse of black night terrain, reached by a single lit road.

ZillaAI is being built for organisations whose material cannot leave, or should not, or simply does not need to.

The proposition is narrow and deliberate: AI that runs on infrastructure you own, governed the way any other production system is governed - so that it can act, and you can still answer for what it did, on whose authority, and why.

Who it is for

Organisations that hold something they are responsible for. Client files, case material, designs, source, records under a retention obligation. The common thread is not industry or size - it is that moving the material somewhere else is either forbidden, unwise, or simply unnecessary.

That includes a great many organisations too small to be sold a governance programme, and which are told, in effect, to wait. We think waiting is the wrong answer and that the process scales down further than the industry has bothered to find out.

What is deliberately not claimed

We would rather be believed later than impressive now, so a few things are worth stating plainly.

This is not a claim that nothing ever leaves a network. That is a binary guarantee, and binary guarantees invite someone to find the one exception and be right. What is being built is a system whose default is local and whose exceptions are deliberate, visible and recorded.

It is not a claim that a machine can hold responsibility. It cannot. The point of the record is that a person can pick it up and answer for what happened.

And it is not finished. What is described here is design intent, published in advance so it can be measured against the result rather than rewritten afterwards to match it.

How the thinking gets published

The writing is the substance - the arguments, the problems worth solving, and the principles we keep arriving back at, with dates on them. Where something can be shown rather than asserted, it gets shown.

There is an Atom feed if you would rather not check back. No account, no newsletter, no platform in between.

It stopped and asked

Authority is granted, bounded, and recorded.

An agent that can act is only useful if you can say afterwards what it did and on whose authority. The interesting moment is not the work it completed. It is the work it declined to complete alone.

Here, a task reached a step that fell outside what had been granted to it. It did not proceed and explain later, and it did not quietly narrow the job to the part it was allowed to do. It halted, recorded why, and waited for a person.

Not an illustration: It refused to half-upgrade — the incident this describes, written up in full.

Two things in that record are doing the work. Held, not attempted means the boundary was enforced before the action, not discovered after it. And approved by means the permission has a person attached to it, so the question "who allowed this" has an answer that outlives the session it happened in.

A system that asks for permission it does not need becomes noise, and noise gets clicked through. So the bar is not whether an action is risky in the abstract — it is whether it falls outside what was granted. That is a narrower question, and it is one a machine can answer honestly.