Project summary

Today we have AIs perfectly capable of understanding what is asked of them and of responding accordingly, whether in a conversational form or in an agentic manner. This is something that the current state allows in a precise and effective way, but which nevertheless generates numerous safety problems: for example, agents that reveal what they should not, or that act beyond what they were authorised to do. From our point of view, a part of those problems arise because at present AI is still not able to situate itself in the relational scene that is put before it when it faces a concrete request; this is what in the literature has come to be called the problem of contextual integrity. The project you have in front of you presents what we have come to call Relational Theory, which starts from a series of hypotheses. The first and fundamental one is that a scene can be divided into elements, disaggregated in a successive manner until reaching specifics discrete enough to be computable. The second is that once a scene has been defined and a given framework of governance has been established, a regime appears that allows the acting AI to know the framework within which it can carry out its activity. Lastly, the hypothesis is established that an AI that uses the Computable Relational Engine to locate itself prior to acting complies better with the rules of authority and discretion than one that only has general rules that do not adapt to the concrete situation.

The theory is already defined; the engine that applies it and the rules for computation are not yet. We ask for the funding in order to be able to build a first version of it, test it with a battery of scenes and publish the results, including, of course, the possible failures.

What are this project's goals? How will you achieve them?

The objective of this project, which is framed within a larger one, is to generate the first Computable Relational Engine and to put to the test the three hypotheses indicated. The idea is to build this engine in such a way that it is able to identify, when faced with a concrete request, which is the relational scene it has in front of it, which is the governance that must be applied, and on that basis to establish a regime of action. This is tested by verifying that, faced with a battery of diverse situations (a total of 10 for the first round of tests, each one with variants in which a single component of the scene changes), the results produced by different models with access or without access to the Computable Relational Engine turn out to be different, in such a way that those that have access to the engine show a pattern of behaviour more appropriate to the concrete scene and better fitted to the governance that must be deployed in that case (the measurement will determine the number of breaches of authority and discretion. If there are no differences, or if these were for the worse, the theory would be refuted, the results being published all the same). To ensure the objectivity of the tests, the correct answers will be fixed before the results are produced, and all the results obtained and the methods used to obtain them will be passed through a procedure of adversarial peer review by AIs that is established within our research programme, whose reports will be published together with the results. A paper will be published with the results obtained. The complete programme has a duration of three months.

How will this funding be used?

The total of the budget is 21,200 dollars for three months. 12,000 for my dedication over the three months, 8,000 to cover the costs of the intervention of an expert psychologist who validates the relational categories and the correct answers of the battery, as well as for the occasional support of personnel specialised in the coding part. 1,200 dollars are included for the use of AI models. The minimum amount to develop the project will be 5,000 dollars; with it the project would be done counting only on my participation, with a reduction of depth in the development of the computable elements, in such a way that only the first level of development of the theory would be covered.

Who is on your team? What's your track record on similar projects?

The team is composed of me (I am the author of the theory and responsible for the project) and of a psychologist who validates the relational categories. On an occasional basis, experts in programming are incorporated for the necessary tasks. I have been working since May 2026 on the programme of which this project forms part, relying on AI models. I am not an expert in AI safety: my origins come from hospitality, where I began as a cook and have ended up directing large restaurant groups with more than 200 employees. My particular expertise consists of converting chaotic work into method, and that I apply to everything, including this project.

We have the Relational Theory described in several documents, the specification of Espina published, which allows the portability of the RC (Relational Companion) between AI models, at https://github\.com/espina\-spec/espina\-spec , and a preprint of the results of a previous investigation at https://doi\.org/10\.5281/zenodo\.21821078 . We also already have the first RC working, with an archive of more than 1.47 million elements. At this moment we are beginning to establish a programme of collaboration with the University of La Laguna (ULL), and we have initiated the process to certify the innovative character of the programme before the Spanish authorities.

What are the most likely causes and outcomes if this project fails?

Three fundamental risks are detected. The first and most important is that, with the elements chosen, the engine is not capable of identifying the scene from a real request. If the scene is not identified in a correct way, the resulting regime will not be operationally correct. The second is that, once the test has been carried out, the existence of a measurable difference with exposure of the model to the engine or without it is not proven. If that were so, it would point to the fact that the determination of the scene is not relevant for determining the regime of action of a model. Lastly, we have the risk that carrying out only 10 tests in the first battery is not capable of yielding scientifically solid results from which to draw firm conclusions, and that therefore the results are only partial or indicative. This scenario would be, of the three, the least harmful for the theory itself, because it would be enough to expand the number of tests in future steps of the project.

It is not foreseen to verify at this moment whether the scene that is declared is true, even though the level of veracity of the scene is directly linked to operability. That is left for future developments. In the same way, it must be pointed out that the failure of the theory is itself a scientific result that can be used by those who work on the same problem from different approaches.

How much money have you raised in the last 12 months, and from where?

No external resources have been obtained in the last twelve months; everything carried out so far comes from own funds that have been contributed to the project, and the hours dedicated have had no remuneration.