Measure resource use and preserve user expectations under load.
1. Cold and warm runs
Separate first-use initialization and download time from warm inference latency. Benchmark on representative supported hardware rather than extrapolating from one flagship phone.
2. Memory and thermal behavior
Long sessions can increase memory pressure and heat. Bound concurrency and test repeated requests. A low median latency can hide severe slowdown after sustained use.
3. Fallback contract
An unsupported or overheated device can offer manual editing or a delayed retry. Cloud fallback requires a separate consent and policy decision; it is not an automatic reliability fix.
Worked scenario
Ten consecutive rewrites perform worse than the first. The feature disables parallel requests and keeps the editor responsive.
Apply it
Record cold and warm latency, repeated-run behavior and failure states on a supported device. Include an unsupported-device demonstration.
Check your understanding
Your performance report states the tested hardware and the limits of its conclusions. Explain the decision and show evidence from your implementation or design. If you cannot demonstrate it yet, revisit the relevant section before continuing.