A model should not only produce an answer. It should reveal what was measured, what was assumed, and where the claim stops.
A correlation can describe motion without explaining what produced it. So the hepatocyte model does not fit a curve to glucose output; it fires individual reactions — exact Gillespie events where copy numbers are low, a Langevin diffusion term where they are high — and lets the observed behaviour fall out of the chemistry.
In the Cell engine this is a mechanism, not an intention. Every parameter carries a checksummed source reference and a verification date; rates that fail the SBML, compartment, and per-cell-unit gates are quarantined as exploratory instead of quietly entering the model. The engine currently reports 0 of 36 reaction parameterizations as source-backed — and says so on its own front page.
Sampling decides who can appear in a dataset. Voxel geometry decides where a molecule can move, and which compartment it is forbidden to enter. Interfaces decide what a user is able to ask. Constraints are not footnotes to a result; they are the boundary of the space in which it means anything.
I began with people, moved through cognition and biology, and arrived at machine learning and scientific simulation. Mechatronics taught me to ask how parts become a system. Research software gives that question a form I can inspect, test, and revise.