Scientific & Medical Disclaimer

Corollary is a non-commercial research demonstrator. It is free, unsupported, and offered to show what open scientific models can do when they are put behind one interface. It is not a product, it carries no service level, and it may change or disappear without notice.

It must not be used for commercial purposes, and it must not be relied upon for any decision that matters.

Nothing on this platform is a medical device, an in-vitro diagnostic, a clinical decision support tool or a laboratory-developed test. Nothing here has been assessed under Regulation (EU) 2017/745 on medical devices, Regulation (EU) 2017/746 on in-vitro diagnostics, or by the FDA or any other regulator.

Output must never be used to diagnose, treat, prevent or mitigate disease in any person or animal, or to make or support a clinical decision. If you have a health concern, consult a qualified clinician. Nothing here is medical, veterinary, legal, regulatory or financial advice.

Every number this platform produces is a model output, and models are wrong in ways that are not always visible in the output:

• A structure prediction is a hypothesis about a fold. A mean pLDDT of 62 is low confidence; a high pLDDT is confidence, not correctness, and says nothing about the conformation a protein adopts in a particular condition.
• Binding affinities and docking scores are estimates, frequently off by orders of magnitude, and rank-order is more meaningful than any single value.
• Generated sequences and designed binders are candidates with an expected success rate well below one. They have not been synthesised or tested.
• Guide RNA efficacy predictions describe on-target activity only, and say nothing about off-target effects.
• Property, toxicity and ADMET predictions are trained on narrow chemical space and degrade sharply outside it.
• Literature summaries can misattribute, and a citation that looks plausible should be checked against the source before you rely on it.

Confidence scores are surfaced precisely so you can judge this. Use them.

Answers are generated by a large language model. It is instructed to ground every claim in a tool result and to say when a tool failed rather than guessing, and it is generally good at this. It can still be wrong, including fluently and confidently, and it can misread a result it correctly retrieved.

You are interacting with an AI system, not a person and not an expert. Verify anything you intend to act on, publish or cite, against the primary source.

Corollary orchestrates models and databases built and maintained by other people. We did not train them, we do not validate them, and we make no representation about their accuracy, currency, availability or fitness for anything.

Public scientific databases are updated, corrected and occasionally retracted by their maintainers. A record retrieved today may not match the record retrieved next month, and the platform reflects upstream errors faithfully.

Model and dataset licences belong to their owners and are yours to honour. See the licences and attribution page.

Runs are kept with their inputs, model version and parameters so a result can be traced. That is a reproducibility aid, not an archival guarantee — this is a demonstration, and stored data may be deleted when it ends. Keep your own copy of anything you need.

The service is provided “as is”, without warranty of any kind. Our liability is limited as set out in the Terms of Service, which you should read alongside this page.

If you use output from this platform to inform an experiment, a publication, a submission or a decision, that is your judgement and your responsibility.