Why should you care?

AIPredict production needs one playbook: train → gate → canary challenger → monitor → promote or rollback.

See it live — copy this example

Use a .NET console or Web API project with Microsoft.ML. Run dotnet run after pasting.

// CI: fail if challenger AUC < champion - 0.01
var champion = registry.Get("fraud", "champion");
var challenger = registry.Get("fraud", "challenger");
if (challenger.Metrics.Auc < champion.Metrics.Auc - 0.01) throw new InvalidOperationException("CI blocked");
// deploy: 10% traffic to challenger, watch drift + latency
services.AddPredictionEnginePool<TxRow, FraudPred>().FromFile(
    Environment.GetEnvironmentVariable("FRAUD_MODEL_ROLE") == "challenger" ? challenger.Path : champion.Path);
monitor.CompareDaily(champion.Metrics, liveSampleEval);

What happened?

• Champion stays live until challenger wins CI, canary, and drift checks.
• Monitoring and observability close the loop to retrain.

Practice next

1. CI compare challenger vs champion AUC.
2. Canary env FRAUD_MODEL_ROLE.
3. Daily drift CompareDaily.
4. Auto-promote after 7-day canary win.
5. Rollback script swaps registry pointer.

Remember

Champion/challenger. CI + canary gates. Monitor + drift loop.

AIPredict enterprise MLOps

Challenger v14 passes CI, canaries 10%, no drift.

Outcome: Promoted to champion; v13 archived for rollback.