A core design decision in OncoTwin is that calibration never produces a bare number: every estimated parameter carries an uncertainty range forward into simulation, and that range is visible to the clinician reviewing the output.
The problem with point estimates
A single "most likely" growth rate is easy to display and easy to misread as certainty. For a system meant to support, not replace, clinical judgement, that's the wrong default.
Our approach
We treat calibration as a distribution-fitting problem rather than a single-value optimisation, and propagate that distribution through simulation so the final projection (survival, side-effect risk, cost) is itself a range.
This is a draft working paper, shared for early feedback ahead of formal submission.