- Protocols: one-button Speed check, Trust check, and Heat and battery tests that check the phone is ready first (cool, power, battery) and record exactly how they ran. - Rigor levels for experiments: Standard alternates the models, discards a warmup run and lets the phone cool between repeats; results lead with the effect size and its 95% interval. - Parameter sweep: time one model across settings such as flash attention, GPU layers and KV cache type. - Blind arena: compare two models on your own prompts without knowing which wrote which, and build your own ranking. - Simpler model settings: a sampling preset and a load profile (Auto, Max speed, Low memory) on top, a list of what differs from the defaults, and every control still under All controls. - LoRA adapters, a chat template viewer, and quick comparison of the same model at different quantizations. - Siri and Shortcuts: ask your model, run a quick benchmark, or open the Lab. - Every run records the engine that produced it and whether the GPU was used; evaluations draw a random seeded sample of questions and show intervals for calibration. - Experiments let you choose the reply length and turn thinking off or on. - Scrolling up while a reply streams no longer pulls you back down; swipe down to hide the keyboard mid-reply. - A clear message when a chat fills the model's context window.

- Evaluations: knowledge, math, tool calling, and structured JSON output, with side-by-side model comparison. - Long-context recall and context scaling experiments. - Vision: ask a vision model about a photo, on the device. - Apple's on-device model on iPhones with Apple Intelligence. - Signed reports you can verify, and share cards for runs and comparisons. - Energy-per-token estimate and a throttling comparison chart. - Four themes, including High Contrast, and a colorblind-safe confidence scale. - Continue a reply that stopped at the length limit. - Session folders; swipe to delete with Undo.