Academic peer review
Follow-up publications are being prepared for peer-reviewed venues. While that work is underway, preprints are publicly available on arXiv and Zenodo, so that the contributions remain visible independently of any review status.
Methodological foundations
โWhich standards make research results verifiable beyond the lab that produced them?"
A piece of research is only as robust as the methodology that produced it. This page documents the operational standards of our work โ from the pre-registration of individual research decisions to the external validation pathways against which we measure ourselves.
This page is written for research institutions, funding bodies and public authorities who want to assess the methodological rigour we apply. It extends the methodology overview on the research page with the operational detail.
Methodological standards do not emerge from a single manifesto. They emerge from concrete decisions made in day-to-day research. We document here those decisions that make our work verifiable by third parties.
The research process
Every research output passes through the same process โ iterative, not linear. Ablation and negative results feed back into hypothesis revision.
Pre-registration
Central research documents are time-anchored via OpenTimestamps. The cryptographic hash of a document is committed to a public blockchain, which creates a priority record independent of any later publication date and verifiable without involvement of the author.
OpenTimestamps is open infrastructure, not an institutional intermediary. Third parties can verify the timestamps later without any cooperation from the researcher.
What we register before the work starts
Reproducibility
Research findings are only robust if third parties can reproduce them under clearly stated conditions. For us, replication material belongs with the publication, not in a later appendix.
Software is released under open-source licences (Apache 2.0 for ZenBrain, a mix of Apache 2.0 and MIT for evaluation scripts). Configurations, hyperparameters and seeds live in the code repository โ so that a forgotten hyperparameter cannot turn a study into something that cannot be reproduced.
Where our replication materials live in public
External validation
Research that certifies itself has not been examined. We build the architecture so that external validation is possible at several points โ not only after the work is finished, but along its phases.
Which validation pathway we use depends on the subject matter. For algorithmic contributions, peer review is the academic standard. For safety-relevant components, notified-body pre-assessments and legal opinions on fundamental-rights compatibility come into play.
Follow-up publications are being prepared for peer-reviewed venues. While that work is underway, preprints are publicly available on arXiv and Zenodo, so that the contributions remain visible independently of any review status.
For the civil-liberties architecture of the Public-Safety track we have held preparatory talks with notified bodies. TรV SรD, TรV Rheinland and Bureau Veritas have experience with conformity assessments under the EU AI Act. These exchanges are non-binding; once the system is used in production, they would be transferred into a formal conformity assessment.
Architectural decisions with fundamental-rights implications are reviewed by lawyers. The assessment focuses in particular on compatibility with GDPR Art. 89, Regulation (EU) 2024/1689 Art. 5, and the Brokdorf doctrine of the Federal Constitutional Court (BVerfGE 69, 315).
The arXiv endorsement procedure carried out by established researchers is a form of upstream validation: an endorser confirms that a contribution meets the academic standard of the relevant category. ORCID, Semantic Scholar and Google Scholar link the publication with verifiable researcher profiles.
Negative results and ablation
Research that only publishes successes is methodologically incomplete. Ablation studies show which components contribute what โ and whether a finding is the result of a single architectural decision or of their interplay.
Our ablation register documents hypotheses that were refuted and architectural variants we discarded. This makes it traceable why the final architecture takes the form it does, rather than another.
In the internal development repository, backend/src/algorithms/ablation.ts is maintained as a feature-flag registry. It allows individual published algorithms to be disabled for comparative studies โ for instance to isolate the contribution of a specific memory layer.
Architectural variants we examined and discarded โ certain reranker configurations or rejected memory topologies, for instance โ are documented with their rationale. Later revisiting of a discarded path is therefore a conscious decision, not an oversight.
Follow-up publications will explicitly flag negative and counterintuitive findings instead of editing them out in favour of a smoother narrative. Replications that show divergent results will be linked from the publications page.
Data minimisation
Data minimisation is not only a data-protection obligation under GDPR Art. 5(1)(c); it is a methodological decision. Annotation schemes that capture more than the research question requires conflate objects of study and undermine the explanatory value of the findings.
In the Public-Safety track, this means concretely: we work with skeleton and motion data, not with facial features. Trajectories are aggregated within bounded analytical windows, not chained across spaces or time. No biometric templates are created.
Annotations capture the construct addressed by the hypothesis โ not everything an annotator might observe. Inter-rater reliability with a Cohen's ฮบ โฅ 0.61 target safeguards the quality of the operationalisation.
Research processing under GDPR Art. 89 is pseudonymised wherever technically possible. Re-identifiability is limited to what the research question actually requires.
Selection decisions are repeatedly validated against outliers โ three independent validation steps before data is used. This prevents biases in sample selection from quietly entering the architecture.
External standards
These standards are not ours โ they are established in their respective communities. We orient our methodology to them.
Findable, Accessible, Interoperable, Reusable (Wilkinson et al., 2016). Research data and software are filed in such a way that they are discoverable via persistent identifiers (DOI, ORCID, GitHub) and remain reusable under Apache 2.0 or CC BY 4.0.
The reproducibility checklists established in the ML discipline โ maintained by JMLR, ICML and comparable venues โ are the reference point for reporting in publications: data splits, seeds, hyperparameters, compute budget, confidence intervals.
Outputs of AI-supported systems are labelled in accordance with Art. 50 of Regulation (EU) 2024/1689 once they are presented to natural persons. Labelling is not a downstream patch but part of the output pipeline.
Research processing is carried out under the safeguards of GDPR Art. 89: pseudonymisation, data minimisation, purpose limitation, technical and organisational measures.
Publications appear as preprints on arXiv (CS.AI) and are persistently anchored via DOI on Zenodo. Software is released under Apache 2.0, replication material under CC BY 4.0.
Open methodological questions
Methodological maturity is a process. We name openly where our standards are still evolving.
The ZenBrain algorithms are public, but independent replications by third parties are still outstanding. As soon as such replications become available we will link them from the publications page โ including those that report divergent findings.
Annotation schemes for the Public-Safety track are currently being developed internally. External annotation by independent annotators โ for instance within the framework of a consortial project โ would be methodologically desirable.
The retention curves modelled in our memory algorithms are validated in the short- to medium-term range. Multi-year follow-ups require research infrastructure we cannot provide on our own โ collaborations with universities or non-university research institutions would be a methodologically appropriate route here.
For Bayesian confidence propagation we report 95% confidence intervals. For more complex pipeline stages โ such as GraphRAG retrieval accuracies under realistic data distribution โ CI reporting is still being expanded.
Methodological discussion
We are happy to answer concrete methodological questions โ for instance about replication material, ablation studies or validation designs. For interest in joint methodological work we welcome exploratory conversations.
Send a methodology inquiryMore from this research
Three tracks, one core, a shared ethics foundation.