Predictive Outreach 2D: AI Risk-Scoring Scatter Plot

AI Outreach Controls

Flagging threshold 0.50

Staff capacity 30%

Outreach effectiveness 55%

Simulation speed 1.5×

Live Model Performance

Recall

—

Precision

—

Missed rate (AI outreach)

—

Missed rate (no program)

—

Cases resolved

Slots in use

0 / 0

Waiting, not yet contacted

Under active outreach

Resolved on time

Missed the deadline

How it works

Every simulated citizen has two hidden features — payment-history score h and income-volatility score v, both in [0,1] — that determine a ground-truth risk via a logistic model, the same kind of model real government risk-scoring pipelines use:

trueRisk = σ(2.2·h + 1.8·v − 1.5),  σ(x) = 1 / (1 + e⁻ˣ)
groundTruthPositive = trueRisk > 0.5

The AI never sees the ground truth directly — it produces a noisy estimate, aiScore = clamp(trueRisk + 𝒩(0, 0.15), 0, 1). Citizens with aiScore ≥ threshold are flagged for proactive outreach, but only up to the available staff capacity (a hard cap on concurrent cases, just like a real caseworker team). A contacted citizen's chance of missing the deadline is reduced by the outreach's effectiveness:

P(missed | contacted)     = trueRisk × (1 − effectiveness)
P(missed | not contacted) = trueRisk

Each resolved case is scored against the classic confusion matrix — comparing the model's flag (predicted positive) to whether the citizen was truly high-risk (actual positive):

Recall    = TP / (TP + FN)   — share of truly at-risk citizens the model catches
Precision = TP / (TP + FP)   — share of flagged citizens who were truly at-risk

Raising the threshold flags fewer citizens (higher precision, lower recall); lowering it catches more real risk but burns capacity on false positives. The main plot places every citizen by AI score (x) and days left until the deadline (y, counting down toward the red wall at the bottom); the blue line marks the current threshold. Drag to pan and scroll to zoom that plane; the two panels below track the live confusion matrix and a rolling history of recall/precision/missed-rate.

AI score → days to deadline ↓ (far → near)

drag to pan · scroll to zoom

Confusion matrix

Recall / precision / missed-rate history