FoundationModels

A small SwiftUI demo app that moderates user-entered team names entirely on-device using Apple's Foundation Models framework.

You add friendly five-a-side football teams to a list. Before a name is accepted, an on-device language model checks it for profanity, slurs, sexual content, or hateful language — no network calls, nothing leaves the device.

How it works

• ContentView — the team list UI. Adding a name runs it through the moderator and either appends the team or shows a rejection reason.
• TeamNameModerator — wraps a LanguageModelSession and performs moderation. It uses two layers of protection:
  1. Apple's built-in safety guardrails (guardrailViolation) catch blatant content.
  2. A guided-generation classifier judges subtler cases.
• TeamNameVerdict — a @Generable struct used for guided generation. Fields are ordered so the model writes its assessment before committing to the isInappropriate boolean, which improves accuracy.
• Team — a simple Identifiable model.

Design notes

• Fails open. If the model isn't available on the device, or a check can't complete, names are allowed rather than blocking the user.
• Prewarming. The session is warmed up when the text field gains focus to reduce first-check latency.
• Prompt safety. Moderation instructions deliberately avoid quoting example profanity, since the guardrail scans the entire prompt.

Requirements

• Xcode 16 or later
• A device/OS version that supports the Foundation Models framework with Apple Intelligence enabled
• Swift 5.0+ language mode

Note: The Swift module is named TeamNamesApp (set via PRODUCT_MODULE_NAME) so it doesn't collide with Apple's FoundationModels framework, while the app product is still named FoundationModels.

Running

1. Open FoundationModels.xcodeproj in Xcode.
2. Select a supported device (or simulator with Apple Intelligence available).
3. Build and run.

On unsupported devices the app still runs — moderation simply allows all names.