Apple's Foundation Models framework puts a roughly 3-billion-parameter model on the phone. That is the part everyone reports. The part that matters for health data is that it is still a generative model: it predicts tokens, it does not compute.
So when it tells you your HRV trended up, there is no sample count, no date range, and no reproducible arithmetic behind that sentence. And a 4,096-token session window means you cannot just hand it a year of readings and ask.
The split that actually works: the model routes and phrases, a deterministic stats layer computes. @Generable and the Tool protocol are designed for this. The model opens a tool, the tool does the math on your real samples, and the result is byte-identical across runs. A real provenance card from our benchmark: 49.6 ms seven-day average, dayCount 7 against coverageDays 30, +16.3% versus the prior 14 days, sha256 f2e87621.
On-device is not the same as deterministic. One predicts a plausible number. The other shows you the work.
Top comments (0)