Trusted AI
An LLM in the loop,
without breaking ALCOA++
It is the question we are asked most: how do you put a model inside a process that has to pass an audit? The answer fits in one sentence — and it carries everything else.
“The model creates nothing
and signs nothing.”
The action blocks and the rules are written and validated in advance, by humans. The model assembles; it does not manufacture.
First principle
The model creates nothing
The actions the system may perform are written and validated in advance, before any model comes into play. The model builds patterns out of those pre-prepared blocks: it chooses, it orders, it proposes. It does not write code on the fly, it does not invent an action, and it has no hand on the data.
- The blocks exist before it does. Every possible operation is written, reviewed and approved by humans. The catalogue of actions is closed.
- It assembles, it does not manufacture. A model that cannot create an action cannot create a non-compliant one.
- What it produces is a proposal, never a record.
Second principle
Nothing is written without human validation
No value enters a regulated record without a named person validating it — and that validation is itself an entry in the log, with their name, the timestamp and the machine.
To the machine
Read, measure, compare, propose, order.
To the human
Decide, validate, sign — and carry the responsibility for what they validate.
To the log
Keep both, in order: the model's proposal and the person's decision.
In practice
What it changes for the operator
Take document reading. The model looks at a scan and proposes the values of the fields. The operator sees the proposal, compares it with the document, and validates or corrects. Nothing the model proposed enters the record without that gesture — and the log keeps both: “the model proposed this”, “the person validated that”.
The direct consequence: a model that is wrong costs a re-read, not an invalidated record. And because the model does not hold the pen, you can swap it, update it or remove it without invalidating a single validated record. That is exactly what a quality department wants to hear.
Third principle
Where the models run
Where your rules demand it: on your hardware, inside your network, or with a provider you choose. We have no dependency on a service you do not control, and no public cloud by default. A slower local model beats a value that leaves your building without your decision.
What we do not do
- Let a model decide on a regulated value on its own.
- Let a model sign anything.
- Write into a record what no person has validated.