When Apple Intelligence entered the project, the existing extractor was not discarded. Phase 7 first introduced an extraction protocol and coordinator, then placed structured Foundation Models output behind that boundary.
Stable actions such as Stop could attempt model-assisted extraction on supported, ready devices. Candidates were validated before becoming tasks, and empty, noisy, duplicate, ungrounded, or unsafe details could be rejected.
If the model was unavailable, timed out, failed, or returned no acceptable result, the coordinator returned to the rule-based extractor. Live and provisional processing remained heuristic-only for predictable feedback.
The architecture keeps the basic capture loop available without presenting every speech or intelligence operation as offline. Model assistance can improve interpretation without becoming the only route from speech to tasks.
← Back to SayList