01 / Question
What if data could fade away? Could quantum computing make it easier to link scattered information back to a person? Why does an algorithm work the way it does, and what happens when it breaks?
We begin with childlike curiosity and a willingness to question what seems settled. Even a simple question can open a different way of thinking about the future.
02 / Testing
We use AI to help model problems, explore scenarios, and design experiments. We challenge assumptions, test alternatives, and look for failure cases that could affect people’s rights and freedoms.
The aim is to understand how a system behaves, where its protections might fail, and who could bear the consequences.
03 / Proof
Research experiments are how we test what we think we know. A result may support an idea, expose a failure, or leave us with a better question.
We make our methods, evidence, and limitations visible so others can examine our reasoning and challenge our conclusions. Where the evidence is incomplete, the question remains open.
Experiments provide evidence under the conditions tested. Formal proofs establish claims under stated assumptions.
Practical standards
Start with scope
A research claim needs a defined object: a particular algorithm, implementation, configuration, dataset, or threat model. The same result may support a narrow statement while leaving a broader claim unresolved.
For empirical work, useful context includes data provenance, selection criteria, baselines, controls, and what would count as a failure. For formal work, it includes the definitions, assumptions, and quantifiers on which the conclusion depends.
Separate types of evidence
- Conceptual illustration: explains an idea; it is not measured evidence.
- Synthetic experiment: examines behavior under constructed conditions.
- Implementation test: checks a specific system and configuration.
- Formal argument: establishes a claim under stated definitions and assumptions.
- Independent replication: requires someone outside the original execution to reproduce the result.
These forms of evidence complement one another, with distinct roles in evaluating a claim.
Make room for counterexamples
Negative results, failed replications, and counterexamples help define where an idea stops working. A clear limitation can be more useful than a broad claim that survives only because the test was too narrow.
Review should distinguish observed behavior from interpretation, and distinguish model-assisted analysis from independent human or machine verification.
Release deliberately
The publication standard is to pair findings with enough context to evaluate them: the research question, relevant methods, artifact versions, limitations, and review status.
Material that could disclose confidential information, proprietary implementation details, or a security-sensitive finding needs a separate release decision.