Represent dependencies without hiding judgment

Independence is an assumption, not a neutral default. Shared quantities, market conditions, productivity, design maturity, weather, and common causes can make uncertain inputs move together. Record a correlation only between stable typed stochastic-variable references such as a cost item, schedule activity, risk occurrence, or event impact. Labels are not relationship keys.

This workbench interprets coefficients as latent dependence in a Gaussian copula. The observed Pearson correlation of sampled nonlinear or Bernoulli outcomes may differ. Every pair needs a coefficient, basis, evidence links, and explicit assumption status. Do not infer coefficients automatically from a short or non-comparable history, and do not translate verbal terms like “strong” into numbers without a declared elicitation scale.

Matrix validity

For an enabled scenario, the complete matrix must be symmetric and positive semidefinite. Self-pairs, duplicate reversed pairs, fixed or zero-variance variables, missing references, and invalid coefficients block correlated simulation. The engine does not silently shrink, clip, or repair an invalid matrix. A scenario may deliberately disable correlations and run independently, but the unresolved dependency issue remains a visible readiness warning.

Use correlation to represent dependence, not causal propagation or mutual exclusivity. Conditional events and causal risk networks are outside version 1. Inspect whether base uncertainty and risk events are being linked to the same underlying cause and whether that also creates double counting. Document why zero correlation is plausible where material. Then configure simulation and convergence. ## Interpretation discipline

Keep inputs, derived results, and human decisions visibly separate. Every material number needs linked evidence or an explicit assumption flag and basis. Unknown is not zero. A percentile is conditional on this model, scenario, seed, engine version, and source version; it is neither a guarantee nor an approval. Any material model edit makes an earlier result stale.