Before intelligence, context.
Why the foundation of a clinical AI application deserves as much thought as the model itself.
An answer starts before the prompt
Every AI application makes a choice about what a model gets to see. In clinical workflows, that choice might include a laboratory result, an active medication, or a note from a previous encounter. It also includes everything left out. Preparing context is therefore part of designing the application—not simply a retrieval step.
From a pile of records to a deliberate input
More data is not always a better input. Duplicate observations consume space. Old notes can compete with recent information. A token limit can silently turn a rich record into an incomplete one. A compiler makes these decisions explicit: normalize the inputs, remove repetition, rank by the task, and fit the result to a budget.
Requirements belong in the build
A ContextSpec describes the information a workflow needs and the boundaries it should respect. Guards make those expectations testable. Provenance records the transformations behind the output. Together, they give teams a way to inspect the foundation beneath an answer.
A more thoughtful starting point
Calyra is built around this idea: the context layer should be something a team can define, test, and understand. Models will continue to change. The need for deliberate inputs will remain.