Kytos Observatory — why we publish every prediction, and every time biology says we're wrong

Imagine a student takes a test and gets every question “correctly formatted” — right structure, no blank answers, nothing flagged by the grading software.

Imagine the answers are all wrong.

That’s what most AI model evals miss: they check if the output looks right, not if it is right. It’s like spell-check passing a grammatically perfect essay that says nothing true.

We’re building tools that predict how cells react to drugs — before you test them in a real lab. So “looks right but is wrong” isn’t a typo, it’s a drug that doesn’t work, discovered a year too late.