On-device, on Apple silicon.
Inference runs locally, on the chip in your iPhone. No data leaves the device. No server path. No "fallback to cloud" loophole. If your device doesn't support it, a rule-based summary ships — same words, no model.
← Apple stack On-device intelligence
Three sentences. Sunday morning. Validated.
Apple's Foundation Models run locally on Apple silicon. Rule-based diagnostics produce the numbers; the on-device model paraphrases under a strict no-new-numbers rule; a validator extracts every number from the model's output and confirms each one came from the diagnostics. If validation fails, the plain-text rule-based summary ships instead. The AI never invents a stat.
FRAMEWORKS
Foundation ModelsOn-device intelligenceApple silicon What this actually does, in flexRep.
No new numbers — validated.
The prompt instructs the model to paraphrase without introducing numbers not present in the diagnostics. A validator parses the output and confirms. If a number appears that the validator can't source, the verdict card falls back to the plain-text version.
A specific action, not vague advice.
The verdict comes back as one of five specific actions — add a set, drop volume, deload, see a clinician, stay the course — with a target, a reason, and a confidence number. Low confidence is shown as low confidence, never dressed up as certainty.
Three sentences.
The card is short by design. One observation. One context. One named pattern. Sunday morning, gym closed, coffee in your hand — read it once and move on.
Four refusals on this surface.
- We don't send your data to a model provider.
- We don't tell you what to do tomorrow.
- We don't adjust your loads.
- We don't cite numbers we can't verify.