Architecture · the third generation of AI
A mind, not an agent
Language models gave AI a tongue. Agents gave it limbs. GAIBRA AI gives it a mind. In every other system, whatever produces the language also produces the verdict. In GAIBRA we took verdict authority away from the language model and gave it to engineered minds.
Three generations of AI
The question that separates them is not "what can it do?" but "what does the thinking?"
Generation one · the tongue
AI that answers
The language model understands, thinks, judges and phrases in a single step. The fluency is remarkable — but fluency is not correctness: a verdict can change between two conversations, and a missing number can be filled with a confident-sounding estimate.
Generation two · the limbs
AI that acts
Agents book, send, pay and execute. Real progress — but the thinking still belongs to the language model, and the agent sits beneath it, carrying out what it decides. Here an error does not stay words in a chat: it becomes an action in the world.
Generation three · the mind
AI that thinks, then judges
Specialized minds carrying deterministic thinking engines that compute and weigh by written rules, with the language model as their tongue only. The verdict arrives with its reasons and with what did not enter it, and the decision stays human. This is GAIBRA AI.
Who holds verdict authority?
A mind and an agent both work with a language model. The difference is direction: which one leads.
The mind in GAIBRA
Sits above the model and leads it
The mind decides what is right, how we know it, and what we do not know. Then it gives the language model one job: phrasing. The model's authority here is to express the verdict, not to issue it.
The agent
Sits below the model and executes its decisions
An agent's strength is action: tools, systems, execution. But what it does is decided by the language model that drives it, with everything a tongue brings — drift and estimation included. That is why every agent needs a mind that thinks before it moves.
The path of a verdict inside GAIBRA
Seven stations, each with a single holder of authority. The language model appears only in the first and the sixth.
- The tongue understandsThe language model reads your question and attachments and turns them into structured signals. It issues no opinion.
- Routing to a mindThe signals go to the mind that owns the right tools for your matter — or to the mind you chose yourself.
- The engines thinkDeterministic engines compute, compare and apply written rules to what you supplied. The same inputs give the same engine result.
- Every claim carries a passportEach material claim carries an Information Passport: its source, its kind (measured, estimated or inferred) and what is unknown about it — per the Epistemic Custody Standard (ECS).
- The verdict with its reasonsThe verdict is issued with its reasons, what did not enter it, what would change it, and a source card when there is one. What is missing is declared and requested, never filled with an estimate.
- The tongue phrasesThe language model returns to write the verdict in plain language. Its authority is phrasing, not judgment.
- The decision is yoursYou adopt the verdict or contest it. And when your matter needs a licensed doctor, lawyer or financial advisor, GAIBRA tells you so, when, and why.
Engineered minds, not role-play
Telling a language model to "act as a CFO" makes it talk like one, not think like one. This is what we call Mind Engineering: from prompts to engineered personas.
Law
What it may judge
Each mind has a written scope and limits: what falls within its competence, and what it refers to another mind or to a licensed professional.
Engines
What it computes and weighs with
Arithmetic, comparison and written rules for its field, running on your inputs — not on a linguistic estimate.
Knowledge and services
What it rests on and what it delivers
Approved knowledge sources for its field, and defined services with their inputs and outputs — not an open-ended chat without limits.
GAIBRA AI among the world's architectures
The industry is now adding deterministic engines that guard what agents do. Determinism is not new — what is new is where it sits: others made it a guard on the limbs; we made it the mind itself.
| Engineering question | General chat assistants | Enterprise agent platforms | Deterministic process automation | GAIBRA AI |
|---|---|---|---|---|
| Who holds verdict authority? | The language model | The language model, with a deterministic engine policing actions | The model plans, the executor executes | The mind and its engines — the model is the tongue only |
| What does governance govern? | Nothing, architecturally | What the agent may do | The steps of the process | What the model may decide |
| Same question, new conversation | The verdict may change | Follows the model | Execution is fixed | Same inputs, same engine result |
| A missing number | May be filled with a confident estimate | Follows the model | The flow stops | Declared and requested |
| Information passing between components | Free text | Shared context | Process state | An Information Passport with source, kind and unknowns (ECS) |
| What you receive | An answer | An executed action | A completed process | A verdict with its reasons and what did not enter it |
Epistemic custody: nothing crosses without lineage
When several minds or agents work together, the most dangerous failure is not a wrong action — it is an estimate turning into a fact along the way.
The problem
An estimate at the start, a decision at the end
The first says: "the company is expected to grow 12%". The second passes on: "the company is growing 12%". The third recommends: "steady 12% growth — invest". Nobody lied and nobody exceeded their permissions, yet the estimate lost its lineage and became a decision.
The answer · ANSANA's open ECS standard
An Information Passport for every claim
Every material claim crosses with its source, its kind, what is unknown about it, a record of who received and handed it on, and the limits of its use. No mind and no agent can promote an estimate to a fact.
Published evidence
What this page says is published with permanent identifiers, and its measurements are published with their limits — whether they passed or not.
Three theses
From tongue to epistemic custody
From Tongue to Mind · Mind Governance · Chain of Epistemic Custody
Custody self-check No. 002
111 / 111
material claims in production turns, all carrying a valid Information Passport — published with the limits of the check. Report
Judgment stability and fairness
8 · 1 · 2
Across 105 runs on 5 minds: 8 pairs passed, 1 failed and we published why, 2 were inconclusive. Report (Arabic)
For agent builders: the mind inside your agent
You have the tongue from a language model and the limbs from an agent. What is missing is the mind that thinks before it moves. We don't compete with your agent — we give it a mind.
Coming soon · developer platform
GAIBRA's minds as your agents' mind
Every mind and every deterministic engine — not a reduced edition — with a key and metering per integration. Announced when it is complete and tested, not before.
Available now
Try the mind yourself
Ask a question you already know the answer to, then look: where each number came from, what did not enter the verdict, and what would change it. 15 Thinking Units free at sign-up, once — no card.