Larchmere Library and Community Hub (Demo Data) Monte Carlo simulation results · Basis: working days

The schedule's forecast finish of 3 September 2026 is not reached in any of the 5,000 simulated runs: the even-odds (P50) finish is 25 September 2026, 15 working days later, and the 80% confidence (P80) finish is 1 October 2026, 19 working days later.

Holding 19 working days of schedule contingency behind 3 September 2026 covers 80% of the simulated runs; the decision asked of the reader is whether to carry that contingency or to act on the work that drives it (section 2). Read 1 October 2026 as a planning scenario, not a commitment: no convergence diagnostics were run and the duration range is one illustrative range applied to every activity, not calibrated to this project.

Every duration and contingency in this report is in working days on the project calendar. The update was filed at data date 5 August 2025 but records actual progress up to 14 October 2025, so it is read and rescheduled at 15 October 2025: the file shows a finish of 24 June 2026, and its logic at the corrected date gives 3 September 2026, the forecast finish used throughout (section 4).

The finish moves most with the four curtain wall and windows activities (A2380, A2260, A2340 and A2300), all on the critical path: in the runs where that work takes longer, the finish comes later.

Correlation measures how closely the finish follows an activity across the 5,000 runs: 0 means no link, and 1 means the finish moves in step with it. After them, the next strongest are A4410 Owner furniture and fit-out (by others) at 0.22 and A4380 Deficiency correction at 0.20. All ten of the strongest drivers sit on the critical path. This is the first place to look for a firmer estimate or a recovery measure.

Of the 252 activities modelled, 56 have a correlation strong enough to tell apart from chance at 5,000 runs (above 0.028); the rest are noise. The tornado chart (exhibit A-05) and the full ranking (exhibit A-06) list all 252.

No risk register was supplied, so the drivers are the activities’ own duration ranges, not named risk events.

Holding 19 working days of contingency behind the forecast finish covers 80% of the simulated runs; 15 working days covers half of them.

The P10 to P90 range is 9 working days wide: the spread the plan carries on these inputs, not a series of separate alarms.

Each activity can run up to 30% longer than planned but only 10% shorter, so the simulated finishes lean later than the forecast.

The figures come from 5,000 simulated runs of the schedule’s own logic, with one illustrative duration range on every activity: the result shows how the plan responds to uncertainty, not a forecast calibrated to this project.

How the Run Was Made

Each run draws the remaining duration of every incomplete activity from a Triangular distribution between 90% and 130% of its planned duration, most likely 100%, and recalculates the schedule forward through its logic. The percentile dates are actual runs from the 5,000, not interpolated values. The run used random seed 42, so it reproduces exactly.

Where the Inputs Came From

The range is one illustrative range applied to every activity alike: it is not calibrated to this project, and no activity-specific estimates were supplied.

Durations are drawn independently for each activity. A shared cause that delays several activities at once, such as weather across trades, is not modelled, and independent sampling understates that kind of linked delay, so the true spread may be wider than shown.

Convergence and the Anchor Check

No convergence diagnostics were run, so whether 5,000 runs settle the P80 to the day is not tested; that is one reason the P80 is a planning scenario.

Anchor check: run without uncertainty, the simulation network reproduces the 3 September 2026 forecast exactly, so every date starts from the schedule’s own logic.

What the Model Cannot See

The simulation network carries logic, durations and constraint floors only, with no resource levelling or progress override modelling. Completed work is held at its actual dates. Each activity’s duration is read on its own calendar, but the pass does not step it through that calendar’s days off, so on a schedule that mixes calendars the percentile dates are approximate.

Against the five maturity levels AACE RP 122R-22 names, the run reads at Level 2 of 5, Ad-hoc; exhibit A-02 sets out what the next level needs.

How the Schedule Was Read

Basis of Measurement

The exhibits behind these findings follow in full, each under its drawing code:

• A-01Headline Risk Metrics
• A-02QRAMM-Aligned Maturity Reading
• A-03Probabilistic Completion Percentiles
• A-04Completion-day Distribution
• A-05Sensitivity, Tornado
• A-06Ranked Sensitivity Drivers
• A-07Schedule Recovery, Baseline vs Mitigated
• A-08Methodology
• A-09Reproducibility Manifest

Δ is measured against the single-point deterministic CPM finish. A later percentile (red) is downside schedule risk, an earlier one upside. The color flags direction relative to the deterministic estimate. The P10–P90 spread itself is expected uncertainty, not a per-percentile alarm.

At 5,000 iterations, a |correlation| below the two-sided 95% significance threshold of 1.96/√n ≈ 0.028 is not statistically distinguishable from zero. 196 of 252 drivers fall below the threshold and are shaded; they remain plotted because the register is never truncated, but their order within the shaded band is noise, not ranking.

Ranked by |correlation|, the same order as the tornado chart; planned duration and impact are in working days. Impact = |correlation| × planned duration. Critical-path activities flagged CP. † At 5,000 iterations, entries with |correlation| below 1.96/√n ≈ 0.028 are not statistically distinguishable from zero; 196 of 252 listed drivers carry the marker. Their rank order is noise, not signal.

A mitigated re-run was not in scope for this analysis; CPP runs this comparison when a recovery scenario is defined.

Forecast completion dates are produced by Monte Carlo simulation over the activity duration distribution implied by the optimistic / most-likely / pessimistic percentages applied to each incomplete activity's planned duration. A full CPM forward pass is executed per iteration. Percentiles are derived from the completion-day vector using numpy.percentile(method='nearest'), which reports an actual simulation trial rather than an interpolated value, so each percentile traces back to a real, reproducible iteration. Sensitivity (tornado) is the Pearson correlation between each activity's sampled-duration vector and the final completion-day vector. Impact = |corr| × planned duration on the legacy PERT/Triangular path, or |corr| × posterior σ on the Bayesian-posterior path (when historical_snapshots was supplied). The active prior shape is reported per row in the sensitivity CSV's prior_method column. Statistical significance of the correlations: at n = 5,000 iterations, the two-sided 95% critical value for a Pearson correlation differing from zero is approximately 1.96/√n ≈ 0.028. Drivers whose |correlation| falls below that threshold are statistically indistinguishable from zero; they are shaded in the tornado (A-05) and marked † in the ranking (A-06) but remain fully listed.

Distribution: Triangular · Iterations: 5,000 · Random seed: 42 · Day basis: working days from data date. Working-day basis counts only the project’s own working days, read from the calendar loaded with this XER, so its workweek and its holiday exceptions are both honoured. The corresponding completion dates are the true calendar dates on which those working-day offsets fall. Each activity's duration is read on its own P6 calendar, but the pass does not step it through that calendar's own days off, so on a schedule that mixes calendars the percentile dates are approximate.

Prepared by the schedule-risk-analysis skill in the CPP Scheduling Suite. This dashboard is self-contained. All assets are inlined and it renders offline.