Separate simulation warm-up from genuine campaign startup

Choose the target before inspecting the curve

Currie and Cheng distinguish steady-state analysis from a terminating simulation. An artificial empty start can contaminate the former; genuine campaign startup belongs in the latter. A cutoff is a modeling decision, not a way to improve an unattractive result.

Write the business question in a sentence before asking for a plot. A production owner investigating routine performance and a launch owner investigating the opening campaign are not necessarily commissioning the same analysis. Ask which experience the result must describe and whose decision depends on it. Retain that answer with the model specification.

Use a fictional display-kit processing station for this exercise. No customer or employee records are needed. The demonstration is intentionally simpler than a production system, so its useful outcome is a clearly explained interpretation rather than a recommendation about staffing or capacity.

Preserve the setup and both starting histories

The executed fixture has 64 replications of 400 steps. Each step adds a Poisson draw with stipulated mean 1.6 kits, then processes up to two kits: next queue = max(0, previous queue + arrivals − 2). Record the queue after processing. Compare initial queues of zero and 12, sharing arrivals within each pair. Replications use separate streams; the paired histories are not independent alternatives.

NumPy’s spawned streams are independent with very high probability, not certainty. Using NumPy 2.3.5, set root = 202611070081. Take SeedSequence(root).spawn(4)[0].spawn(64); each default_rng(child) draws poisson(1.6, size=400). Preserve every trace.

Prespecify cutoffs at 0, 40 and 100 steps and a separate 40-step opening campaign. These invented durations are diagnostic choices, not recommended burn-in settings.

A second implementation reconstructs each queue from cumulative net input and checks it against the step-by-step calculation. This verifies the saved fixture’s arithmetic, not the realism of its arrival mechanism. Keep those two forms of evidence separate when describing the exercise to a reviewer.

Calculate the retained-window comparison

Average each retained trace, then average the 64 replication means. With cutoffs 0, 40 and 100, empty-start results are 1.4357, 1.4806 and 1.4858 kits; loaded-start results are 1.8672, 1.4957 and 1.4858. The opening 40-step campaign from empty averages 1.0320 kits. Its startup observations must remain when that opening period is the target.

Close agreement here does not prove convergence or identify a universal cutoff. Steady-state output still fluctuates. Preserve the three comparisons instead of selecting whichever supports the preferred operational story.

Explain why the same saved history can support different questions without treating the questions as interchangeable. A launch report should not quietly substitute a later operating window simply because the later number looks more representative of normal activity. Conversely, a routine-performance report should disclose the artificial setup that influenced its early observations.

If the analyst cannot justify the chosen interpretation, leave the decision unresolved. Request a targeted diagnostic rather than lengthening the run until the chart looks persuasive. The additional work should answer a stated concern, with the earlier outputs retained for comparison.

Record a defensible disposition

Summarize which target was requested, which evidence was examined and what remains uncertain. The owner should be able to distinguish an arithmetic failure from a disagreement about what the report ought to describe. Those issues require different remedies, even when both prevent acceptance.

For a real application, ask the qualified analyst to explain the diagnostic strategy and the evidence supporting the starting conditions. Do not let an editor’s preferred story determine the retained observations. If the intended audience or operating question changes, reopen the interpretation rather than carrying forward the earlier label.

Keep the complete synthetic record, including the inconvenient opening period. Document any later authorized analytical revision as a separate version. The final deliverable is a reasoned answer to a specified question, with limitations visible; it is not permission to remove operational records or rewrite a campaign’s actual experience.

Make the reasoning concrete.

A fictional display team asks about its opening campaign, while an analyst brings a report prepared for routine operation.

The team keeps the opening experience visible and requests the appropriate analytical interpretation. It changes no actual operation.

Take the questions into your next brief.

Use these prompts in your own notes. Nothing is submitted from this page.

1. Question

   Whose decision must this result inform?

2. Target

   Which experience is the report supposed to describe?

3. Setup

   What was fixed before inspecting the outputs?

4. Evidence

   Can the saved fixture be independently reconstructed?

5. Disposition

   What remains unresolved after the comparison?

Watch for the wrong conclusion.

• Selecting observations to match a preferred story.
• Calling verified arithmetic evidence of operational realism.
• Replacing the commissioning question without telling its owner.

Sources & scope

Original decision reasoning and synthetic examples. Source-derived methods and their applications remain attributed. Fixtures were executed locally with NumPy 2.3.5 and independently checked arithmetic. No actual operational data, staffing decision, customer promise or configuration change is involved.

• A Practical Introduction to Analysis of Simulation Output Data →Primary methodological documentation · Supports: Supports the target distinction.
• Simulation Output Analysis →Primary methodological documentation · Supports: Supports initialization and observation scope.
• Parallel random number generation →Primary methodological documentation · Supports: Supports spawned stream construction.