Releasing model weights has become a signal for openness in AI. It is a meaningful signal: access to weights enables independent evaluation, fine-tuning, and scrutiny that closed models do not allow. But weights alone do not make a model usable, understandable, or trustworthy. For health applications in particular, weights without context can be more misleading than no release at all.
A researcher receiving only weights cannot know what data the model was trained on, what populations are represented, what preprocessing was applied, what evaluation was done, how the model behaved on underrepresented groups, or what failure modes were observed. Without that context, the researcher cannot assess whether the model is appropriate for their use case, and any downstream application inherits unknown risks.
What genuine openness requires
Full openness for a health AI model includes, at minimum: a training data card describing the source populations, collection conditions, and known limitations; a model card describing intended use, known limitations, and evaluation results across subgroups; reproducible evaluation code and benchmark datasets so independent researchers can verify claims; and documentation of what was tried and did not work, so the field does not repeat the same paths.
This is more demanding than a weight release, and the additional demand reflects the higher stakes of health applications. A model that performs well on average but poorly for specific populations can cause harm if applied without understanding that heterogeneity.
Our commitment
Cytognosis treats full documentation as a release requirement, not a post-release improvement. Every model and dataset we publish will carry the provenance, evaluation, and limitation documentation that makes it genuinely usable by others. This page is part of our open notebook and will evolve as our release processes develop.