A.3.3.PI:4.5 - Test the horizon, error and changing conditions

Work a case through to the requested future result. Include a pair merged by the simpler description when one motivated the repair. Identify which retained information now changes the calculation.

When readings are imperfect, propagate their stated uncertainty far enough to determine whether it changes the result. A reconstruction can amplify observation error even when the underlying equations have a unique solution. Parameter and input uncertainty require their own treatment; changing a sampling interval can change the update rule.

Check repeated use at its claimed horizon. A universally valid closed update can be iterated while its domain and input assumptions hold. A good fit over one observed step does not provide that result for an approximate predictor. Check the errors or bounds relevant to its use and shorten the horizon or change the account when needed.

Keep forecast performance and aggregate statistics tied to their questions. A model can reproduce a long-run average yet lose the history that changes the next-event probability. Conversely, a useful short-horizon predictor may not reproduce long-run behavior.