Methodology
Methodology version: Beta 1 — 4 October 2026
What the check is
Mobile AI Check is a browser-based capability assessment. It combines signals that a browser legitimately exposes with conservative interpretation rules. It is not a laboratory benchmark and it cannot see every physical component in every phone.
Evidence labels
- Browser-reported: a value supplied by a browser API.
- Measured: an active capability probe completed during the scan, such as obtaining a WebGPU adapter.
- Verified specification: a value matched to maintained technical documentation. This tier is reserved for future evidence-backed device matching.
- User-confirmed: information explicitly supplied or confirmed by the user.
- Estimated / inferred: interpretation derived from available evidence.
- Unknown: the browser does not expose enough information to make a responsible claim.
Current Beta 1 signals
The engine may use logical CPU thread count, coarse device-memory class, mobile/OS hints, touch points, secure-context state, WebGPU API exposure and WebGPU adapter information where available. Browser support varies, so missing data is expected.
How recommendations are formed
Beta 1 uses conservative rules to classify a device as a basic, lightweight or stronger candidate for mobile AI workloads. Workload statements are estimates. They are affected by model size, quantisation, context length, runtime/backend, operating system support, memory pressure, thermal limits and implementation quality.
What we deliberately do not do
We do not turn a missing signal into an invented hardware specification. We do not claim that WebGPU alone proves a particular AI runtime will work. We do not treat a chipset's theoretical AI capability as proof that a specific phone exposes the required software/runtime.
Planned evidence system
Future versions are intended to combine published specifications, browser-observed evidence and deliberate real-device testing. Those evidence types will remain separate so users can see why a conclusion was reached.