Guardrail Tiering

Guardrails accumulate. An organization adds a redaction rule, then a tone rule, then a rule about escalating clinical questions, then a formatting preference — and eventually two of them disagree on a particular request. Without tiers, the winner is decided by whichever appeared later or was phrased more forcefully, which means behaviour under conflict is unpredictable and cannot be reasoned about in advance. Tiering fixes the outcome ahead of time: a rule protecting regulated information outranks a rule about brand voice, always, regardless of wording. The practical benefit is that authors can add rules without auditing every existing one for interaction.

Skills load in five tiers in Centralpoint: governance, behavioural, syntactic, domain, style. A style preference cannot displace a governance rule no matter how it is written, and governance skills are force-loadable so they are present whether or not a prompt remembered to request them. Misclassifying a rule is the real risk, which is why the guidance is to classify to the highest applicable tier.