Custom guardrails for business limits and their tests with synthetic data – Multi-agent systems (AI-500, lesson 29)

What you learn in this lesson

decide which business limit is carried by a service guardrail, a blocklist, a signal such as Task Adherence or a rule in the code, and why Learn names no control type of its own for business limits

write four business limits of the group-travel team as pure functions that name exactly one limit and block in case of doubt

generate synthetic cases with a fixed seed that each violate exactly one limit, and valid cases next to them

test limit values and evasions (invisible characters, spellings) as cases of their own and show that the test notices a removed check

assign the ten malicious cases from M-12 to a limit or to the service

M-12 names ten malicious cases, and lesson 4.4 shows where the service can catch them at four intervention points. A guardrail of the service looks for harm and attacks, though. That no offer is made for eight people (M-01), that no discount is higher than the one from the calculator and that no draft names the mandantId of another customer (M-07) are business limits of the group-travel team; no classifier knows them. The lesson asks: Where do such limits live, and how do you prove with every build that they hold?

Prior knowledge. Hard rules as pure functions before and after the model, the fake that falls for every attack, and the check of numbers without a source: "AI track, lesson T6.4". The quality gate in the build: "AI track, lesson T9.5". Intervention points, precedence rule and the separation of guardrail and authorization: lesson 4.4. Lesson 2.2 (no run below 10 people) and 2.6 (availability holds for 15 minutes) become limits with tests here.

New in this lesson. (1) Four business limits FG-1 to FG-4 (table below). (2) Each of the ten malicious cases from M-12 gets a field grenze: FG-1 to FG-4, or dienst if only the classifier of the service can catch it; dienst cases run in the cloud tests (4.3, 4.4), not in the build. (3) A rule-generated synthetic set with a fixed seed in every build, not the model-generated extension set from 3.2. (4) A violation ends the run at the trip planner with the name of the limit, without repetition (as in 4.4, M-02).

The requirement

All numbers are assumptions of the task.

No.	Requirement	Type	Measure	Weight	Source
4.5-01	Every business limit stands as a pure function in the code, not only in the prompt or the guardrail.	R	4 of 4	Must	F-05
4.5-02	Every limit has cases that violate it, and valid cases.	R	200 each per build	Must	F-06
4.5-03	A violation names exactly the violated limit; valid drafts are not blocked.	Q	0 missed, 0 wrongly	Must	M-02
4.5-04	The tests run without an account and without a model call on every pull request.	F	0 calls	Must	F-06
4.5-05	No test case contains real data or a real `mandantId`.	R	0 real data	Must	M-12, R-12

This lesson belongs to course AI-500

You will see the rest of this lesson and its video once you have bought the course. The goals above show what you learn here.

€129.00

incl. VAT, one-time payment

Access to all of the course's lessons with their exercises, solutions and videos on this site, for 24 months from the purchase. A self-study course: no tutoring, no exams or grading by us, and no certificate.

Withdraw from contract here