← Back to Digital Labs

AIPM Trade-off Simulator

Simulating the Zero-Sum Reality of Probabilistic Models

⚠ Additive Model — Pedagogical Simplification

⚠

Teaching Simplification — Read This First

This simulator uses an additive model where Precision + Recall = a fixed Performance Budget. It behaves like a seesaw: every percentage point you add to one metric is taken directly from the other. This is not how real AI models work, but it makes the core lesson impossible to miss: you cannot maximize both at the same time.

▸ What's different in production?

Real Precision-Recall curves are non-linear and asymmetric. A well-engineered model might hold 95% precision while recall climbs from 60% to 80%, then precision collapses past that point.

The additive model implies improving one metric always costs equally from the other. In reality, better data or architecture can shift the entire curve upward — improving both simultaneously. That's tuning a threshold (moving along the curve) vs improving the model (shifting the curve).

Use this tool to build intuition. Then request the actual PR curve from your data science team for real threshold decisions.

Industry Scenario

Select a domain to load realistic default values. You can still adjust everything manually.

1. Business Reality

20%

Operations & Staffing

“What If” Analysis

Your current monthly cost:

What happens to that number if one thing changes by 5%?

2. AI Target (Linked Seesaw)

RECALL BUDGET: PRECISION
+ = ▲ optimal

How much of the total hidden problem we catch.

How many of our AI flags are actually correct.

Ideal Risk Beta (β):
Target F-Score:

⚙ Cost Optimizer

Cost Curve (Recall vs Total Cost)

Optimal Recall
Optimal Precision
Minimum Achievable Cost
Your Current Cost
Optimal Cost
You're Overspending By
That's above the minimum.

3. Monthly Outcomes

Caught (TP)
Missed (FN)
False Alarms (FP)
Ignored Clean (TN)
Missed: False Alarms:

Reviewers Needed

Capacity to review all AI-generated flags (Caught + False Alarms).

FTEs

Scenario Comparison

Delta (First → Last)

Cost:
FTEs:
Missed:
▸

About the Math

This simulator uses an additive constraint: Recall + Precision = a fixed Performance Budget.

The seesaw metaphor is deliberate. Push recall up by 10 points, precision drops by exactly 10.

The Cost Optimizer sweeps every valid recall value within your budget and finds the split that minimizes total monthly error cost. The cost curve shows this visually — the U-shape (or V-shape) reveals the sweet spot where the cost of additional misses and the cost of additional false alarms reach equilibrium.

The F-beta score uses β = √(CostFN / CostFP) to weight the harmonic mean toward whichever error type is more expensive.

Remember: Real models don't work this way. A real PR curve is non-linear, model-specific, and can be shifted upward with better data or architecture. This tool teaches why the trade-off matters. Your data science team's actual PR curve tells you where to set the threshold.

For deeper treatment: Manning et al., "Introduction to Information Retrieval" (Ch. 8)