Examples

Real-world use cases showing how typed uncertainty changes decision-making. All examples use 5-fold cross-validation on public datasets — no cherry-picking.

Credit risk assessment

Loan approval/denial using the German Credit dataset (1,000 applications, 20 features). The highest-stakes example: a wrong approval costs $20K–$200K.

The problem

A Random Forest classifier achieves 78% accuracy. Sounds reasonable — until you look at the failures. 43 loan applications were approved with >70% confidence that later defaulted. The model was confident and wrong.

With typed uncertainty

import httpx

API = "https://api.example.com"
headers = {"X-API-Key": "sk_...", "Content-Type": "application/json"}

# Train on historical loan data
httpx.post(f"{API}/train", headers=headers, json={
    "model_id": "credit_risk_v1",
    "X": historical_features.tolist(),
    "y": historical_outcomes.tolist(),  # "good" or "bad"
})

# Assess a new application
result = httpx.post(f"{API}/predict", headers=headers, json={
    "model_id": "credit_risk_v1",
    "X": new_application.tolist(),
    "cost": 20,  # a wrong approval is 20x worse than a missed auto-approval
}).json()

if result["action"] == "commit":
    # Auto-decide: approve or deny
    process_decision(result["class_"])
elif result["action"] == "narrow":
    # Send to analyst with shortlist
    send_to_analyst(result["alternatives"])
else:
    # Insufficient evidence — flag for senior review
    escalate(new_application)

Results

Medical triage

Patient symptom classification into routine / urgent / emergency. The three output types map directly to clinical triage categories.

result = httpx.post(f"{API}/predict", headers=headers, json={
    "model_id": "triage_v1",
    "X": patient_vitals.tolist(),
    "cost": 50,  # missing an emergency is 50x worse than over-triaging
}).json()

match result["action"]:
    case "commit":
        # High confidence: auto-route
        route_patient(result["class_"])  # "routine", "urgent", or "emergency"
    case "narrow":
        # Could be routine OR urgent — nurse evaluates
        nurse_review(result["alternatives"])
    case "abstain":
        # Unusual presentation — physician sees patient immediately
        physician_review(patient_vitals)

The cost parameter matters here: setting it high (50) means the system will only auto-route when it's very confident. Ambiguous cases always go to a human. No emergency is ever auto-classified as routine.

Manufacturing quality control

Classifying products as pass / inspect / reject on a production line. Speed matters — but so does not shipping defects.

# Batch prediction: 100 items off the line
results = httpx.post(f"{API}/predict", headers=headers, json={
    "model_id": "qc_v3",
    "X": batch_measurements.tolist(),
    "cost": 5,
}).json()

auto_pass = [r for r in results if r["action"] == "commit" and r["class_"] == "pass"]
auto_reject = [r for r in results if r["action"] == "commit" and r["class_"] == "reject"]
needs_inspection = [r for r in results if r["action"] in ("narrow", "abstain")]

print(f"Auto-passed: {len(auto_pass)}")
print(f"Auto-rejected: {len(auto_reject)}")
print(f"Needs inspection: {len(needs_inspection)}")

Operational impact at scale

Common pattern: batch processing with routing

Most production use cases follow the same structure: predict a batch, then route each result by type.

results = httpx.post(f"{API}/predict", headers=headers, json={
    "model_id": model_id,
    "X": batch.tolist(),
    "cost": cost_ratio,
}).json()

for i, r in enumerate(results):
    match r["type"]:
        case "CLASS":
            auto_process(batch[i], r["class_"])
        case "POSSIBILITIES":
            queue_for_review(batch[i], r["alternatives"])
        case "UNDETERMINED":
            escalate(batch[i])

This pattern works for any domain. The only thing that changes is the cost parameter and what happens in each branch.