Brands as configuration
A content engine that runs any number of brands from one codebase, where a brand is a config file and a voice guide, and the controls that stop it publishing the wrong thing are enforced in code.
Client: HW Legacy Group.
One pipeline: sense sources, generate candidates, put them through a quality gate, queue, schedule, publish, log, measure. A brand is not a fork of that pipeline. It is a configuration file and a voice guide.
The quality gate has two layers. The deterministic layer is regular expressions and entity lists: banned terms, house style, duplicate detection against what has already gone out. The second layer is a model scoring against a rubric. Anything touching a refused category goes to a review queue with the reason attached rather than being silently dropped, because a filter that discards input without saying what it discarded turns a content problem into a mystery.
About a third of the gate's test suite is negative controls, cases that must pass. A filter that blocks everything scores full marks on every positive test and teaches the reviewer to stop reading reasons.
The part I would keep in any system like this: no channel can publish until a smoke test for that specific channel has been run and recorded, and that requirement is enforced by the config validator in CI rather than by intention. A rule written down and not enforced somewhere is not a rule.