Facts
Sheffield’s Complex Systems Modelling group explicitly spans deterministic and stochastic models, principled approaches to uncertainty, and complex computational or physical simulation.
Official group page ↗Research practice
A practical framework for validating machine-learning evidence.
A model result becomes useful evidence only when the data-generating process, split strategy, uncertainty and intended decision are all examined together.
Published: 22 July 2026 · Last updated: 22 July 2026 · Facts are linked to sources; the framework is my interpretation.
Data provenance
Document who or what produced each sample, how labels were obtained, which preprocessing decisions were applied and whether repeated observations from the same subject can leak across splits.
Study design
Choose the unit of independence before training. Pre-register primary comparisons where feasible, separate exploratory tuning from final evaluation and retain a genuinely untouched test stage.
Metrics and uncertainty
Report class-level performance, uncertainty intervals, calibration and clinically or operationally meaningful error costs—not only an aggregate accuracy.
Transfer and deployment
Test temporal, geographic, demographic and device shift; monitor failure modes; define abstention or escalation; and avoid claiming decision support without workflow-level evidence.
My interpretation
That breadth creates a strong environment for asking whether a model is merely predictive, mechanistically informative, or suitable for a real decision—and for designing different validation evidence for each goal.