When to Trust a Model
Trust in a model is earned per question, per regime, and per required accuracy, never granted to the model as a whole.
Trust Is Scoped, Not Global
A model is never simply trustworthy or untrustworthy. It is adequate or inadequate for a specific question, in a specific regime, at a specific required accuracy. A weather model trusted for tomorrow's temperature is not thereby trusted for next month's. The correct question is never is this model good but is this model good enough for this decision.
A Checklist
- Is the code verified, with a demonstrated order of accuracy and bounded discretization error?
- Has it been validated against experiment in a regime close to the one being predicted?
- Is the prediction an interpolation within the validation domain, or an extrapolation beyond it?
- Are the input uncertainties propagated to output uncertainties?
- Do independent methods or codes agree on the same quantity?
The Extrapolation Trap
Models are most dangerous exactly where they are most useful: predicting regimes no experiment has reached. Confidence built by matching existing data does not transfer automatically outside that data's envelope. An extrapolation should always be flagged as such, and its distance from the validation domain stated, so a reader can weigh it honestly.
Convergence of Evidence
Trust grows when independent lines agree: a verified solver, validation against experiment, an analytic limit, and a second code built on different assumptions. No single line is decisive. This is why Kronos design work keeps verification evidence, validation comparisons, and stated assumptions as separate ledgers, and treats simulation-only results as conditional until hardware exists to test them.
The honest posture is calibrated: state the regime where the model has earned trust, and refuse to launder that trust into regimes it has not reached.