Apple's Foundation Models framework gives iOS apps a free, offline 3B model with a 4,096 token window per session. It is good infrastructure, and the privacy property is real. It is also not the same claim as "this app shows its work", and 2026 is full of apps quietly treating the two as one.
A language model predicts the next token. It does not evaluate arithmetic. Ask it for the average of seven values and it produces a plausible string, because plausible strings appeared in its training distribution. Run it twice and you can get 49.6 and 49.8. Both pass a schema check. Neither is a function of your data.
The framework's own Tool protocol points at the right answer: do not ask the model to compute, ask it to call something that computes.
In MetricBridge that is AskEngine.swift, a plain compute layer. It parses a question into one of 16 tool calls, resolves the metric name, runs the aggregation over your recorded rows, and returns a struct called ComputedFacts. The struct carries the canonical metric name, the unit, a range label, the count of days that actually have data, the raw (day, value) rows, the headline aggregate and its label, and a direction. It also carries two honesty fields: coverageDays beside dayCount, so a "most recent value" answer still reflects the whole series, and unevenWindows, so a comparison with a different number of recorded days on each side says so instead of printing span totals next to a per-day rate.
The model never computes a figure. It gets a short summary and phrases it.
You can check that claim yourself. The repo commits a script, bench_deterministic_layer.py, that builds a seeded 30-day HRV series and prints a provenance card. Headline 49.6 ms (7-day average), +16.3% versus the prior 14 days (mean 42.6 ms), dayCount 7 against coverageDays 30. Run it twice: byte-identical output, sha256 f2e87621b46b2fe04cd81a86cb2183ed2e271237ca8d23b9123171445359d9d7 both times. The card closes with "figures computed by the layer: 5, by the model: 0."
None of this diagnoses or monitors anything. Where a metric moves away from your recent range, the honest word is a deviation from your personal baseline, never abnormal. Being able to trace a figure does not make it a diagnosis.
Full write-up, including the provenance card and the side-by-side: https://www.healthexport.dev/blog/apple-foundation-models-vs-deterministic-stats-layer?utm_source=devto&utm_medium=social&utm_campaign=exp-20261011-hea215-foundationmodels
Top comments (0)