Iterative rather than linear

One pass shows how a result moves through the four stages and can also go back at evaluation. A click on a stage shows what this page says about it.

Four stages and one feedback loop
Ablation & negative results → hypothesis revision 01 Pre-registration OpenTimestamps anchor 02 Reproducibility Code · data · configs 03 Evaluation Comparison runs · ablation 04 Publication Open access · DOI Ablation & negative results → hypothesis revision 01 Pre-registration OpenTimestamps anchor 02 Reproducibility Code · data · configs 03 Evaluation Comparison runs · ablation 04 Publication Open access · DOI
Illustration · a symbol, not a real project history

Process diagram of the research methodology in four stages: first pre-registration via OpenTimestamps, second reproducibility with code and data, third evaluation with comparison runs against other systems and ablation of our own components, fourth open-access publication. A dashed feedback loop carries ablation and negative results back into hypothesis revision.

Four stages, one feedback loop

A click on a stage or the feedback loop, in the picture or in the list, shows here what this page says about it. A second click shows all again.

Pre-registration

Anchoring research decisions in time before the work begins

Central research documents are time-anchored via OpenTimestamps.

Third parties can verify the timestamps later without any cooperation from the researcher.

Reproducibility

Replication as default, not an add-on

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.

Evaluation

Comparison runs · ablation

Evaluation puts comparison runs against other systems next to ablation of our own components; whatever does not hold up there goes back into hypothesis revision.

Publication

Open access · DOI

Publications appear as preprints on arXiv (CS.AI) and are persistently anchored via DOI on Zenodo.

Feedback loop

What did not work is part of the result

Research that only publishes successes is methodologically incomplete.

Our ablation register documents hypotheses that were refuted and architectural variants we discarded.