What is the recommended 'safe' memory threshold for a local model to ensure it isn't purged by the OS when the app moves to the background?

Since we cannot pin model versions, does Apple provide an API to trigger the Evaluations framework automatically upon an OS update to verify prompt stability before the user interacts with the app?

When using the continued-processing.gpu entitlement for background tasks, is the rate-limiting behavior different for third-party adapters (MLX/Core AI) compared to the system’s AFM models?

yesterday: Eric (Foundation Models team) gave a "spoiler alert" explaining that "the weight list applies only to Siri" and not to the Private Cloud Compute (PCC) language model or other on-device Siri features . So I wanted to understand the nuance: the term... appears to refer specifically to the set of authorized model weights or versions that the Siri engine is allowed to load or route to for system-wide voice features