Write cost, latency, and quality rules

PCC is not a credit-card invoice in your app, but it is still a larger, slower, networked call. On-device is cheaper in battery when the prompt is short. A custom Core AI model is worth it when quality on your job beats the system model, not because the file is large.

enum TaskClass: String, Sendable {
    case classify, extract, draft, plan, customStyle
}

enum PrivacyBound: String, Sendable {
    case mustStayOnDevice
    case applePCCAllowed
}

struct RouteRule: Sendable {
    var task: TaskClass
    var privacy: PrivacyBound
    var maxMilliseconds: Int
    var needsDeepReasoning: Bool
    var prefersCustomCoreAI: Bool
}

func choose(_ rule: RouteRule, available: Set<String>) -> String {
    if rule.privacy == .mustStayOnDevice {
        if rule.prefersCustomCoreAI, available.contains("coreai") { return "coreai" }
        return "system"
    }
    if rule.needsDeepReasoning, available.contains("pcc") { return "pcc" }
    if rule.maxMilliseconds <= 800 { return "system" }
    if rule.prefersCustomCoreAI, available.contains("coreai") { return "coreai" }
    return available.contains("system") ? "system" : "pcc"
}

Start with four task classes:

• classify / extract: on-device first.
• draft: on-device, PCC if the user asked for a long rewrite and privacy allows it.
• plan: PCC with a deeper reasoning level when the SDK offers .reasoningLevel.
• customStyle: Core AI when you baked that voice.

If privacy is mustStayOnDevice, PCC is illegal for that request even if quality would rise. That is a product rule, not a model score.

Next, attach the choice to a Profile modifier.