grenli

The learned confidence gate is the crucial piece for a 14MB action model. On ambiguous requests such as the HN example, what calibration target decides between abstaining locally and escalating to the cloud?

Live comparison is useful, especially when availability changes faster than list prices. How do you normalize effective cost across providers for interruption risk, storage and egress, and region?

Local extraction plus a 16ms target is compelling for agent loops. Does the 50-URL benchmark measure structural fidelity as well as speed—especially on JS-heavy pages and tables?

The nested-iframe support and context savings stand out, especially for browser agents running long workflows. How do you measure self-healing reliability separately from false-positive actions that appear to succeed?