What is held constant
The goal, expected labels, observed accessibility state, action menu, local freshness checks, input execution, and final verification are shared. Each attempt restarts the app; stored application data is preserved. Starting labels and a stable accessibility snapshot are checked. Any initial-state mismatch excludes the cohort from speed comparison.
Jev chooses from categorical answers. The baseline generates a structured decision using the same offered operations and targets. Their transports and output formats differ.
How to read these numbers
Full-run times include model calls and local device work. Request latency is measured at the client and includes network and provider processing. Its median is calculated over available call timings, including failed attempts; it is a separate measure from full-run speed.
Only pairs where both runs verified and shared the same initial state contribute speed ratios. This small local experiment measures these models, this scenario, and this machine. It does not establish a general winner. Costs are estimates from recorded usage and provider pricing, not billing readback.