Research
Advancing trustworthy AI
in the open.
GammaLex develops open datasets, evaluations, frameworks, and technologies that study a single question: what an AI system needs to understand before it acts, and how anyone can check whether it had it.
What we study
Five questions nobody has settled.
Every one of them is still open. We publish the datasets and benchmarks so the answers can be argued with — not asserted.
Missing
How does a system know something was never there?
Everything looks answerable — until the one record that mattered was never filed.
Conflicting
When two records disagree, which one wins?
Neither, ideally. But showing both only helps if a person can see why they differ.
Changed
How do you tell that the ground has moved?
Knowing something changed is easy. Knowing it just invalidated your reasoning is not.
Uncertain
When is what you have simply not enough?
A system that always answers can never say “not enough.” Teaching it when to is unsolved.
Boundary
What if the most relevant record is off limits?
Sometimes it is. Relevance must never be allowed to argue its way past permission.
If you think one of these is already settled, we would genuinely like to be shown.
Published in the open
Everything is inspectable and evolving.
Datasets, benchmarks, methods, and specifications — open, so the claims can be checked rather than taken, and updated as the research moves.
- Dataset
Open Evaluation Corpora
Shared corpora and evaluation assets for studying whether the sources behind an answer hold up. Published as the DALI Citation Benchmark.
- Benchmark
Reproducibility Benchmarks
Measures of whether conclusions stay inspectable, reproducible, and reviewable. Published in the DALI repository.
- Framework
Verification Taxonomy
Methods for assessing verification, provenance, and review visibility.
- Specification
Evidence Package Specification
Reference structures for provenance, review, and durable decision records.
- Repository
Open Repository
Public tooling for transparency, verification, and reproducibility.