Guardrail

A check around the model rather than inside it. Input filters that reject a request before it reaches the model, output filters that inspect what came back, and rules about what the conversation may do next. It exists because training a model to behave is probabilistic and a deployment often needs something closer to a guarantee.

The distinction from Constitutional AI is where the work happens: that shapes the model during training and produces a tendency, this sits outside at request time and produces a decision. The two are complementary and are often confused, which leads to systems that rely on the trained tendency for something they needed a rule for. Rails are also inspectable and editable in a way a trained disposition is not, which usually matters more than their accuracy.

Every filter is a classifier with a threshold, so it inherits a classifier's trade: tighten it and legitimate requests are blocked, loosen it and unwanted ones pass. The base rate makes this hard rather than merely annoying — if one request in a thousand is genuinely a problem, a filter at 99% accuracy on both classes blocks ten legitimate requests for every real one it catches, and the users who notice are all in the first group.