Your Spanish, or your language
On the chat page you choose how the mind speaks to you. Castilian as spoken in Spain, Andalusian, Canarian, Rioplatense or Mexican Spanish. Or Catalan, Valencian, Basque or Galician, which ALIA was trained on as first languages rather than as translations. Leave the choice open and it answers in whatever you wrote.
Regional varieties come from what the model has read, so Andalusian or Canarian will read as a register rather than a faithful transcription of speech. Catalan, Basque and Galician are proper languages of the model and read as such.
What is inside
Each part is a separate job. In the Spain edition each part is a separate ALIA conversation with its own instructions, so the parts still read your question independently and can still disagree. The check that measures whether they agree is the same Sheaf coherence engine.
Emotional reasoning
Reads how the people involved feel and what they need to hear.
Conscience
Pushes back. Says what is unsure, what is being avoided, and what would be a mistake.
Reason
Works out what is true, what follows, and what the trade-offs are.
Imagination
Finds the options and the reframes you had not thought of.
Memory
Keeps a few things it learns about you, stored encrypted.
The check
Before the voice speaks, Sheaf's coherence engine measures whether the parts agree. If they do not, the answer says so and names the tension.
The voice
One part speaks for all of them, in the language and variety you chose.
The model is ALIA-40b-instruct, released by the Barcelona Supercomputing Center under the Apache 2.0 licence. You can watch every part's notes as it thinks. Nothing is hidden.
It can run completely within Spain
Every part of this edition, the model, the coherence check, the memory and the record of decisions, can run on servers inside Spain: on an institution's own premises, in a Spanish data centre, or in a cloud region in Spain. ALIA's weights are open, so no call leaves the country and no AI lab, in Spain or elsewhere, sees the traffic. Sheaf Labs is a Spanish company, and this is the shape it proposes for schools, banks and public bodies.
The public demo on this site runs the same edition on Sheaf's own server in the United States, because that is where Sheaf's demo systems are. Even there, nothing is sent to the Barcelona Supercomputing Center or to any AI lab.
Data handling
The Spain edition keeps a small number of facts it learns about you, tied to this device and stored encrypted, so it can remember you. Nothing is used to train any model. Your question is processed by ALIA on a server Sheaf operates; no third-party AI lab receives it. See the privacy policy.
The mathematics behind it
The check is the same in every edition. Sheaf Mind uses sheaf cohomology: each part of the mind is one local piece, and before the voice speaks the mathematics asks whether all the parts fit into one whole. A tension that only shows up when the parts are put together is found and named in the answer. When the parts disagree, anything the mind was asked to do on the strength of that answer waits for a person.