Why your plan date is optimistic
Adding up most likely durations produces a date that is beaten less than half the time. Simulating the whole network shows you why, and by how much.
The merge bias
Take a project where two independent chains of work must both finish before the next step. Each has a 50% chance of hitting its planned date. The probability that both do is 25% — so the step after them starts late three times out of four, with nothing having gone wrong.
This is merge bias, and it is why deterministic schedules are optimistic in a way that feels unfair. The plan asked for several independent things to each go about as well as expected, all at once. The more parallel work a plan contains, the worse the effect gets.
No amount of care in estimating individual tasks fixes this, because it is not an estimating error. It is an arithmetic property of taking the maximum of several uncertain numbers.
What simulation actually does
Each run draws a duration for every activity from its own three-point estimate, then computes the critical path from scratch on those numbers. Do that several thousand times and you have a distribution of finish dates rather than a single one.
Crucially the critical path is recomputed every run rather than fixed in advance. That is the difference between simulation and the single-path arithmetic in PERT, and it is where the interesting result comes from.
The criticality index
The most useful output is not the percentiles. It is the proportion of runs in which each activity ended up on the critical path.
A deterministic schedule sorts work into critical and not-critical, and reality is not that tidy. An activity might be critical in 40% of runs — never critical on the plan, entirely capable of deciding the finish date. Those activities are invisible to ordinary scheduling and they are exactly where attention pays off.
A short path with wide uncertainty will often overtake a longer path with tight uncertainty. Path length alone does not tell you which chain governs your project; length and variability together do.
Independence is the weak assumption
Every simulation of this kind samples activities independently, and real delays are not independent. The specialist who is unavailable holds up four things at once. The supplier who slips affects everything downstream. A team that estimates optimistically does so consistently.
Correlated delays make the true distribution wider than the simulation shows. So a P80 from an independent model behaves more like a P70 in practice, and the right response is to treat the percentiles as a floor on your uncertainty rather than a measurement of it.
More runs do not mean more truth
Going from one thousand iterations to twenty thousand makes the percentiles more stable. It does not make them more accurate. Accuracy is limited by the estimates and by the independence assumption, and computation improves neither.
Twenty thousand runs on numbers guessed in a planning meeting produce a very smooth histogram of a guess. The output deserves exactly as much confidence as the three-point estimates behind it, and no more.
What this tool will not do
It calculates from the figures you enter and nothing else. It cannot tell you whether those figures are honest, and a precise-looking output built on rough inputs is still rough.