Every production system eventually hits the same wall. Business logic that once fit comfortably inside a few hundred lines of if-statements now has to read free text, interpret JSON payloads, weigh risk, and look at images — and it has to do all of that thousands of times a minute without drifting. The two usual answers are unsatisfying. Hand-written rules break the moment reality stops being enumerable, and general-purpose chat LLMs are flexible but return prose, vary between identical runs, and often cost more than the decision is worth. A decision engine occupies the space between those extremes. Instead of generating text and hoping a parser survives it, a decision…