Systems that answer
The code that holds a veto over what the model produces — rules that always run and always give the same answer to the same input.
Also called Rule layer · Guardrail · Детерминированный слой
The division of labour is the whole architecture: the model proposes, the rules dispose. Prices come from the price list rather than from the reply. Eligibility is evaluated by a function against the record. Anything the business would be held to is computed, and the model's role is to phrase what the computation returned.
This is what makes the probabilistic part safe to use, and it is this practice's stated invariant for the work: a probabilistic component is only safe inside a deterministic boundary. The boundary is not a filter reading the output for bad words — it is the layer that owns the facts, so the output has no opportunity to disagree with them.
It also fixes the failure that appears in production rather than in testing. A model that is right ninety-nine times in a hundred is wrong once, at scale, in a way nobody sees, on a claim somebody acts on. A rule is right every time on the cases it covers, and the cases it does not cover are enumerable — which is a different kind of risk, and the manageable one.
A filter that inspects the generated reply is checking a finished claim against patterns, and it will pass a confidently wrong price that happens to look like a price. The rules belong upstream, owning the values the answer is built from. If your system can produce a number the price list has never contained, the boundary is in the wrong place.
Knowing the definition is not the same as being able to check the figure. These are the procedures that do the second thing.
- Accepting an automated agent
- “The agent is trained on your data and ready to go live.” · 45 minutes, 6 questions.
The definitions are the easy part. Whether the figure on your dashboard was computed this way is a different question, and usually the more expensive one.