The simulator is seeded: the "random" return paths are generated from a fixed starting point, so running the same plan twice gives exactly the same numbers.
That is deliberate, and it buys three things:
- Fair comparisons. When you change one input, or when an optimizer sweeps a whole grid of candidates, every variant is judged against the same set of market paths. A difference in the result is caused by your change, never by a luckier draw.
- Reproducibility. A success rate you saw yesterday is still there today. Results are testable, and a saved scenario re-opens to the numbers you remember.
- Hand-checking. With every volatility set to zero the fan collapses to a single deterministic path you can reconcile line by line in the year-by-year detail table. The Trace: zero volatility button on that table does it in one click. It zeroes Blended's Volatility %, the three per-asset vol fields, the taxable sleeve's own Volatility % (when you've set one), and the two uncertainty toggles, and Restore volatility puts your exact settings back. Nothing you trace is written down while the trace is on, so quitting mid-trace leaves your real settings where you left them.
The trade-off: a seeded run understates "simulation noise," the tiny wobble you'd see between two genuinely random batches. With 1,000 paths that wobble is small, and stable comparisons are worth far more to a planning tool than fresh dice each run.
When a run can't be believed
The Fat-tailed model's worst draw is floored at a total loss (a growth factor of zero) rather than allowed to go below one. A bucket in a floored year funds nothing: money dated inside that year comes from the next bucket in your draw order, and if nothing is left, the plan fails that year and the shortfall is reported. That is the honest answer, and it is rare: roughly one run in a hundred at the default 6% / 12% assumptions.
Some entries have no answer at all. A return or inflation assumption of −100%/yr or worse is not a distribution, and a plan with no simulated years or no paths has no population to measure. Rather than draw a chart from numbers that mean nothing, the projection says which assumption is out of range and shows no result. A blank chart or a success rate would be worse than nothing, because both read as an answer.