Dual-process architecture in production

Input → System 1 (Decision Model) → Simple? → Typed answer (ms)
                               ↓
                           Complex? → System 2 (LLM) → Deep reasoning

Safety gate pattern

Put the Decision Model in front of tool calls. If the typed decision denies, never invoke the LLM or the tool.

Model routing

Mature 2026 stacks route 70–90% of traffic to light classifiers / SLMs and reserve frontier LLMs for the rest.

FAQ

Do I need a Jev waitlist for dual-process?

No. The pattern works with any typed classifier; Jev is the first public System One option.

Where should the gate sit?

Before expensive or dangerous tool calls, and before calling the frontier LLM.