`varietyDirection` is a tag on every action a system takes that records which way the variety moved — did the model get richer to fit a world that stayed free (good), or did the world get narrower to fit a model that stayed simple (bad). It exists because the control-theory math underneath treats those two as identical, and you need a place in the data model where they stop being identical.
What it distinguishes
Every action gets sorted into one of three buckets. Fidelity-up (augmentation): the model's variety rose; the world stayed at least as free; your option space is preserved or widened. Variety-down (capture): the system's variety fell to fit a simple model; your option space shrank — the move that looks like "improved relevance" on every dashboard. Neutral: no net change. The crucial, counterintuitive point: both of the first two register as "better regulation." That's the whole reason the tag has to exist — you cannot infer direction from "did the metric improve?", because the metric improves either way.
Why it's an action-level tag
Instead of judging the system as a whole — "is this tool good?" — you judge discrete acts. That's what makes the ethics tractable: a question per action, not a verdict on the product. Each action also carries who did it and to whom, because variety-down on a market is strategy, but variety-down on the person you serve is the thing to flag.
The governance posture, encoded
Three rules carry the stance. Completeness: you may not have an unlabeled action — silence is the failure mode capture relies on, so the schema makes "we never decided" illegal. Capture guard: a variety-down action on a user isn't auto-rejected, because sometimes it's legitimate; it's flagged for a consent-and-legibility review. The system can't decide capture-versus-care on its own, so it escalates to a human rather than silently allowing or silently blocking. Consistency: if the measured change says the world shrank but the label says otherwise, that's the signature of a mislabel — flagged. It's the defense against the obvious gaming move of tagging everything as augmentation.
What it is not
It's a governance scaffold, not a solved problem. The tag surfaces the decision; it doesn't make it. Whether a given narrowing is care or capture still comes down to consent, legibility, and whose objective is served — human judgments the schema routes to a person rather than resolves. The measure of "variety" is a proxy and an honest one. And it's gameable if labeling is adversarial — which is why direction is cross-checked against the measured change, and why the completeness rule is a hard requirement. The point isn't to make capture impossible. It's to make it un-silent — to deny it the "we never noticed which way the variety was moving" defense.