reasoning-sandwich

Structured model allocation wraps implementation actions with dedicated reasoning phases, ensuring thorough analysis before execution and verification after completion. Apply this pattern to complex tasks requiring rigorous planning, traceable execution, and strict verification.

To implement pre-reasoning and post-verification, allocate dedicated models to analyze requirements and define success criteria before acting, then use a post-verification model to check outcomes against criteria and regressions after action completion.

Yes, the reasoning-sandwich pattern handles edge cases by assigning a pre-reasoning model to explicitly analyze requirements and identify edge cases before the action model executes the solution, mitigating potential safety issues.

The best way to enforce traceable planning is applying a reasoning-sandwich pattern that separates pre-reasoning, action, and post-verification models, explicitly documenting the plan and verification criteria within the AI workflow.

Avoid using multi-model allocation for simple tasks that do not require rigorous planning, edge-case analysis, or post-verification, as the overhead of dedicated pre-reasoning and post-verification models outweighs the safety benefits.