01Use the simplest mechanism that fitsStored answers need lookup; bounded policy needs rules; predictable sequence needs workflow. Reasoning is for recurring judgment with multiple feasible options.
02Treat reasoning as a controlled systemCase context, deterministic logic, analytics or optimization, bounded AI assistance, authority and a decision trace should work together.
03Make recommendations inspectableShow evidence, assumptions, excluded options, trade-offs, authority and expected outcome. Insufficient evidence should be allowed to produce no-decision.
04Earn wider authority with evidenceProgress from observe to prepare, recommend and approval routing before considering bounded execution with monitoring and rollback.