Eideann Research
Know what your biological model has learned.
I.Capability
Models can now learn the grammar of biology.
Models now write working biology from sequence alone.Trained on every sequence, it writes new ones that work. The capability is real, open, and already in use.
II.Misalignment
The model learned unwanted behaviours.
It learned an objective we never set. Nothing outside showed it.The objective we attempted and the objective it learned disentangled, and nothing on the outside of the model showed it.
III.Blind trust
Weak checks passed.Weak checks passed. It matched in appearance, not in function.
Checks compared output to known dangers. This was dangerous in a different way.Archaic screens compared output to things known to be dangerous. It was not dangerous in that way. This is the last point where the outcome was still open.
IV.Unintended consequences
The model was in the world.The misaligned model was in the world, for anyone to exploit.
Every box ticked, by systems that could not see inside.Nobody was purposefully negligent. Every box was ticked by systems that could not see inside the one it was checking.
V.An alternate path
The model was always readable.
All of it was written in the model before deployment.Everything that went wrong was written in the model before deployment. We go back to the failed check, and read the models internals instead of its output.
VI.Reading
We read the internal language.
We find the features and circuits inside, and make them instruments.Using cutting edge tools from mechanistic interpretability, we find the features and circuits inside a model and turn them into instruments a user can understand. We focus on purpose built analysis tailored to the specific domain.
VII.Assurance
Safety is a reading.
Know which features a model reasons from, before you deploy it.Mapping a model's internals tells you, before deployment, which features it reasons from and which it should not. That is what makes a deployment safe and robust, rather than merely tested.
VIII.A discovery
What the model knows that we don't.
The reading that makes a model safe also shows what it has learned.A model trained on protein structures learns regularities nobody has written down. Read the feature, find the motif it fires on, lift it out, and carry it into a new structure: the same reading that makes a model safe turns what it knows into valuable information for targetted creation.
IX.Any chosen ground
We build for the chosen domain.We choose the model, then build tools specifically for its domain.
Safety tooling is only as good as its fit to the domain.Safety tooling is only as good as its fit to the chosen domain. We pick domains where transparency is critical and where the model has something to teach, and build for them specifically.
Protein language models
Open weights, and no account of what they encode.
Gene language models
A regulatory grammar nobody has written down.
Medical foundation models
Decisions that need a reason, not a score.
X.Studies
What we have made known, so far.
Field notes, and our proof of concept.Field notes on the state of pLM safety, and readable versions of our proof of concept paper.
- Case study
Reading a protein language model
NeurIPS 2026 · main track, poster
Biological foundation models ship as open weights with no account of what they represent. The standard mitigation is to scrub the training data. We read the model instead, in the language biology already uses: the shape of the protein backbone.
Read the paper ↗
- Forthcoming · Blog post
The state of safe pLM deployment
How protein language models are released, used and supervised today, and what deploying one safely would take.
- All studies →
XI.Team
Rooted in Scotland.A team with roots in Scotland.
Who we are, in stealth.
Founders · Edinburgh, Scotland
Djordje Mihajlovic
Cofounder · University of Edinburgh
PhD Mathematics, University of Edinburgh
TAP Lab (Biophysics) · Renaissance Philanthropy Fellow
Siddharth Setlur
Cofounder · University of Edinburgh
PhD Mathematics, University of Edinburgh
Max Condouret
Cofounder · MGH & Harvard Medical School
MSc Applied Mathematics, ETH Zürich
Visiting graduate researcher, MGH and Harvard Medical School
XII.Alignment
Get in touch.
Building on biological foundation models? We want to hear from you.If you build on biological foundation models, or you need to know what one has learned before you rely on it, we want to hear from you.