Trust needs evidence, oversight and clear limits.

People set the authority

The strategic goal is autonomy within agreed limits. People define the purpose, permitted actions, resource budget and decisions requiring approval. The system is designed to work within that scope and pause or escalate when the task requires a person’s judgement.

A proposed plan, permission to execute and evidence of an observed result have different roles. The architecture keeps them distinct so that a confident recommendation does not itself authorise an action.

Make the basis of a result visible

The project is developing methods for connecting outputs to their sources, recording relevant processing steps and communicating uncertainty. Explanations and provenance serve different purposes: an explanation helps someone interpret a result, while provenance records where its supporting information came from. Both need evaluation.

Improvement needs its own checks

Learning from outcomes produces candidate changes. Separate evaluation and release procedures determine where those changes may be used, with a recorded scope and a route to restore a previous version.

The aim is to preserve oversight as the system adapts, and to make the effects and limits of an update visible.

Keep people involved where decisions matter

The research programme includes human oversight, ways to question outputs and procedures for responding to unexpected behaviour. The form of oversight depends on the use case and its consequences.

Health and care pilots are research activities with defined protocols and safeguards. Their descriptions on this website are not clinical guidance, evidence of treatment effectiveness or an invitation to enrol.

Treat data according to its context

The project’s data management work addresses permissions, purpose, access, retention and the conditions for reuse. Pilot activities involving people or personal data require the relevant ethics and data protection arrangements before proceeding.

Open science does not mean that all data can be made public. Personal information, confidential material and restricted assets require appropriate protection.

Test fairness and robustness

Performance averages can hide differences between groups or operating conditions. CEREBRA-AI will investigate fairness, calibration, robustness and behaviour under changing inputs as part of its evaluation. Findings should make the tested conditions and remaining limitations clear.

Support regulatory readiness

The project includes work on risk management, technical documentation and alignment with applicable European AI requirements. The requirements for a particular system depend on its purpose, role and deployment context.

Documented evidence from development and evaluation will guide the assessment of each application. Findings will be reported with their scope and limitations.

Publish responsibly

Publications and releases will distinguish research objectives from demonstrated findings. Where assets can be shared, accompanying material should explain their intended use, evaluation conditions, licensing and known limitations.