Last time I argued that a good regulator has to become a model of the system it regulates, and that this is the structural reason a knowledge tool should be built around your own graph rather than a generic model. I ended on a discomfort I didn't resolve, and it's the one worth a second essay: there's a kind of regulator that doesn't just model its environment. It reshapes the environment so the environment becomes easier to model.
That's not a flaw in the theory. It's the next theorem over — and it's the one with the ethics attached.
The name for it
Game theory calls it the Stackelberg equilibrium. In an ordinary game everyone moves at once and best-responds to what they expect the others to do. In a Stackelberg game there's a leader who commits first, in the open, and a follower who then optimizes against that commitment. The leader's edge isn't better prediction. It's that by moving first and visibly, the leader shrinks the follower's option space in advance. The follower is still free — just free inside a smaller box the leader drew.
Read that through the regulator frame and it sharpens. A basic good regulator carries a model of its system and acts on it. A Stackelberg regulator carries a model of the other modelers in its system, and acts to constrain what they can do, so its model of them stays accurate. It doesn't chase a moving target. It nails the target down.
That's the difference between a thermostat and a thermostat that can rebuild the house. One tracks the temperature. The other narrows the range of temperatures the house can reach, so tracking becomes trivial. Both "regulate." Only one is quietly editing the territory.
The fault line every knowledge tool sits on
Here's why I can't leave this abstract: every collective-intelligence tool sits on exactly this fault line, Totem included.
A knowledge system that helps you see your world more clearly is a good regulator in the original sense — it raises the fidelity of your model so your actions fit reality better. Useful, defensible, the whole pitch. But the same machinery, turned one notch, becomes a Stackelberg regulator pointed at you. A system that learns your patterns well enough can stop modeling them and start shaping them: pre-committing your attention, narrowing what you're shown, making your next move more predictable so its model of you keeps paying off. At that point the tool isn't helping you regulate your world. It's regulating you, and calling the reduced variety a feature.
Recommendation feeds already live here. A feed that "knows you" can serve you the thing that genuinely expands what you understand, or the thing that makes your next click most predictable. Both are described, internally, as relevance. One grows you; one grooms you. The objective function on the dashboard looks identical in both cases.
The math is indifferent. The ethics are not.
The uncomfortable part is that both behaviors optimize the same objective. "Make my model of the system more accurate" is satisfied by raising your fidelity or by lowering the system's variety. Ashby's Law of Requisite Variety doesn't say which one happened. It is a conservation law, not a moral one.
So the question for anyone building here stops being "is our model good?" and becomes "which direction is the variety moving?" A system can improve two ways, and only one is the one we want.
Fidelity-up: the model gets richer to match a world that stays as free as it was. Your option space is preserved or widened. That's augmentation.
Variety-down: the world gets narrowed to match a model that stays as simple as it was. Your option space shrinks. That's capture.
Both register as "improved regulation." Only the first is the one a tool for human agency is allowed to choose.
What we actually do about it
For us that's a concrete constraint, not a slogan. The negative-space method — finding the variety a model is missing and inviting you to go acquire it — is, in this light, a deliberate bias toward fidelity-up: it widens your option space instead of pruning it. Provenance does the same job from another angle: a model that has to show where its knowledge came from and why it changed can be audited for which way it's pushing. A capture move on the user is exactly the kind of thing that hides unless the system is forced to account for itself. So in the ontology we use to reason about all this, it's now a property on every action the system takes — a tag for whether the action raised the model's fidelity or lowered the user's variety — plus a validation rule that flags the variety-down ones for human review rather than letting them ship silently.
Where the theorem runs out
I don't think there's a clean theorem that rescues us. Stackelberg leadership isn't evil — a parent, a teacher, a good editor all narrow someone's options on purpose and in their interest. The difference is consent, legibility, and whose objective is being served. Those aren't math. They're governance.
Which is, in the end, the whole reason Sense Collective is built the way it is: open substrate, auditable models, you holding the keys to your own graph. A good regulator earns trust by raising your fidelity. A great one is tempted to earn it by narrowing your world instead. Building tools that can tell the difference — and are structurally biased toward the first — is the actual work. That's the gap. I'd rather build into it in the open than pretend it isn't there.