Lux: Picture this, Hex. You're sitting in a car — steering wheel, pedals, gearshift, the whole setup. Looks like you're in the driver's seat.
Hex: I mean, if I've got the wheel, I'm driving.
Lux: Except the car is on a flatbed truck being towed down the highway. You can spin that wheel all day. Crank it left, crank it right — nothing changes about where you end up.
Hex: Okay, that's a cruel image. The controls are real but useless.
Lux: And that's exactly the puzzle a recent preprint in emergence calculus digs into. Not whether controls exist — but whether using them actually makes a difference.
Hex: So we're talking about agency.
Lux: Agency in this framework is causation inside a fixed layer. You freeze the stage — the variable set, the dynamics, the constraints, the budget. Then you ask one clean question: does changing the action variable change the future?
Hex: Wait — the stage has to already be set? Like, the road has to already exist?
Lux: Right. That's the agenthood side — the enablement question we unpacked last time. The road, the lane markings, the speed limits, all of it being there and stable. Agency comes after. Once the road exists, agency is whether turning the steering wheel actually steers.
Hex: So agenthood builds the stage, and agency is the performance on that stage.
Lux: Exactly. And the framework has a specific tool for measuring it.
Hex: What's the tool?
Lux: Feasible empowerment. It's a channel capacity — borrowed straight from information theory. You look at the induced channel from budget-respecting action sequences to outside macrostates, and you measure how many distinguishable futures your actions can reach.
Hex: Channel capacity. So — high empowerment means your actions fan out into many different possible futures?
Lux: Many distinct, distinguishable futures. And zero empowerment means the outcome is the same no matter what you do. Your steering wheel is disconnected.
Hex: So you're measuring the width of your steering range.
Lux: The width — not the direction. Empowerment doesn't care where you want to go. It just measures how many places you could go. Think of it like a combination lock — if you have three dials, each with ten digits, your empowerment is a thousand reachable combinations. Whether any of those combinations is the one you want? Different question entirely.
Hex: A thousand possible futures, and the framework doesn't rank them.
Lux: It refuses to rank them. On purpose.
Hex: That's a deliberate choice, right? Not importing goals?
Lux: Very deliberate. We'll come back to that. But first — the trap.
Hex: The trap?
Lux: The painted-on steering wheel. In the ring-world exhibits, there's a null regime where the agent's action toggles follow a fixed exogenous schedule. Every time step, certain toggles flip. The outcomes change. It looks like the agent is choosing.
Hex: But the flipping isn't caused by the agent's policy.
Lux: It's baked into the environment's schedule. The agent's supposed action variable is just tracking what would have happened anyway. Empowerment looks high because different toggle patterns produce different outside positions — but the agent didn't cause the pattern.
Hex: The painted-on steering wheel. You can see it turn, the car changes direction, but your hands aren't doing anything — the tow truck is.
Lux: The paper calls it an enablement mistake leaking into the causal layer. You smuggle an environment feature into the action column and then claim you're driving.
Hex: So how do you catch that?
Lux: You run the null audit. Compare empowerment under the agent's actual policy against empowerment under the null — the regime where the schedule runs regardless. If they match, the "agency" was fabricated.
Hex: And in the ring-world, they literally ran that comparison?
Lux: They did. Under the null regime, empowerment stayed high even though no genuine policy choice was happening. The toggles flipped on schedule, outside positions changed, and the channel looked rich — but the richness came from the environment, not the agent. The null audit exposed it.
Hex: Okay, so agency is strictly a within-layer thing. But the Notch paper — that made a sharper distinction, right? Two kinds of time?
Lux: Causation-time and enablement-time. Causation-time is the within-layer successor — same variables, same laws, you just advance the clock. Given the stage, what happens next?
Hex: That's where agency lives.
Lux: That's where agency lives. Enablement-time is different. That's when the layer itself rewrites — new variables appear, old ones lose meaning, the closure changes. The theory gets reborn.
Hex: So causation-time is driving on the road. Enablement-time is building a new road entirely.
Lux: And the diagnostic is clean. Ask: does the variable set change? If no — same variables, same intervention semantics — you're in causation-time. If yes — new variables needed for closure to work — you're watching enablement.
Hex: And agency never crosses that boundary?
Lux: Agency is defined strictly within a fixed layer. The moment you need to invoke new variables or rewrite the closure to explain what happened, you've left the agency story and entered the enablement story. They're different logical levels.
Hex: Like — explaining a chess move is causation-time. Explaining why chess exists as a game is enablement-time.
Lux: Perfect example. The chess rules are the layer. Moving a knight is agency — your choice changes the board state. But if the rules themselves changed mid-game — new pieces, new squares — that's enablement rewriting the layer.
Hex: But here's what bugs me. If agency depends on a layer, and the layer depends on which variables you package and which lens you use — isn't agency just... relative?
Lux: It is relative. And the framework owns that. The Quantum Theory paper's OI–EI principle — ontological identity of empirical indiscernibles — puts it sharply: don't build into a layer's ontology distinctions that the layer itself cannot stably witness.
Hex: So if the layer can't tell two outcomes apart, you can't claim your action caused the difference between them.
Lux: You can't claim difference-making through a distinction that's invisible to the layer's own instruments. Agency is real — but real as seen through the layer's lens. Switch the lens, the empowerment number changes. The framework mitigates this by using explicit, stable lenses and by including null regimes that catch obvious mis-modeling.
Hex: The null regime again. It keeps coming back as the honesty check.
Lux: It's the most important single audit in the whole paper. Without it, you can't separate genuine causation from smuggled-in enablement.
Hex: Alright, so — big guardrails moment. What empowerment is not.
Lux: Empowerment is not a goal theory. High empowerment does not mean good outcomes. An agent can have enormous difference-making capacity and still be completely aimless — lots of futures it could reach, no preference among them.
Hex: And flip side — a system can be laser-focused on a single objective but have tiny empowerment?
Lux: Narrow channel, one clear target, low capacity. Think of a thermostat — it has a goal, it pursues it effectively, but its action channel is one bit wide. Barely any empowerment. The framework deliberately avoids importing goals as primitives because it wants to separate "how much can you steer" from "do you have somewhere to go."
Hex: That's a real discipline. Most agency definitions smuggle in goals or preferences right at the start.
Lux: Which is exactly why Six Birds theory keeps them apart. Empowerment is structural. Goals are a separate layer of description. You might eventually want both — but you don't confuse the measuring stick with the destination.
Hex: Steering range versus heading.
Lux: And one more thing the paper is careful about — the interface itself. The action alphabet, the set of moves you can make, is assumed as part of the layer. Discovering what your controls even are? That's a packaging problem — an enablement question. Agency starts once the dashboard is in place.
Hex: So the framework doesn't pretend to solve everything. It draws a clear line and works inside it.
Lux: So here's the picture. Agency in this framework is causation inside a fixed layer — difference-making measured by feasible empowerment, audited against the null, constrained by what the layer can actually witness. The road exists because of enablement. Whether the wheel is connected is agency.
Hex: And the schedule trap is the audit for fakes — check whether the steering actually causes the turn, or whether the tow truck was doing the work all along.
Lux: Next time someone says "this system has agency," three questions. Inside which layer? Through which lens? And did anyone check the null?
Hex: That's a checklist worth keeping.