Correlation modelling

Independent uncertainty flatters the P80: with many activities sampling independently, overruns and underruns cancel and the tails thin out. In reality the same crew running late on one dig runs late on the next. Correlation is between activities' duration draws (not between register risks), and there are two tools for it.

1. Groups first — one ρ per shared driver. In Correlation groups on 4 Ranges, use the quick-add: scope by WBS branch, Activity code, UDF, Name contains or All activities, set ρ, click + Correlate these. Members share a correlated percentile; ρ = 0.8 is typical for one crew or one weather window. The demo ships a "Groundworks crew" group at 0.8 across the earthworks chain.
   Name the driver, not the maths: a group is "these activities share a crew", captured in one number.
2. Pairwise ρ when one group number genuinely can't express it. Pairwise ρ… opens the Iman–Conover editor: pick two activities, set ρ (−0.99 to 0.99 — negative correlation included). Ranks are reordered, so each activity keeps exactly the distribution you gave it; only the pairing of draws changes.
   The modal's status line checks your matrix as you type — groups remain the default and recommended practice.
3. Let the consistency check protect the run. An impossible combination (A↔B high, B↔C high, A↔C strongly negative) is flagged in plain English, naming the offending trio and the range that coefficient must sit in — and the run refuses until it's fixed. No silent "nearest valid matrix" repairs.
4. See what it did. Re-run with and without the group (snapshot both — same seed): the deterministic date barely moves, the P80 tail stretches. That gap is what independence was hiding.