Lux: Today's mini-lab has two pieces of equipment on the bench, Hex. A joystick and a playlist. Both control something. But they work in fundamentally different ways — and the Throw paper uses both on purpose.
Hex: Control is control, Lux. Why do you need two different versions?
Lux: Because they measure different things. The joystick is feedback — you see what's happening and react in real time. The playlist is open-loop — you commit to a sequence in advance and measure what comes out. In the emergence calculus, the viability kernel is the joystick. Empowerment is the playlist. And today's experiment shows where they agree, where they split, and why the Six Birds program needs both.
Hex: Alright. Start with the joystick. What makes the viability kernel a feedback notion?
Lux: The viability kernel certifies that for each state in K, there exists a feasible action keeping all successors inside K. The key phrase is "for each state." The action you choose at state s can be different from the action you choose at state s-prime. You look at where you are, you assess what's available, and you pick the best move. Then the system transitions. You look again. Pick again.
Hex: Real-time steering. The joystick moves with you.
Lux: And the policy — the map from states to actions — is the essence of feedback control. You're not committing to a plan in advance. You're reacting to outcomes as they unfold. This is necessary for viability because the system is stochastic. The kernel flips a coin at every transition. If you committed to a fixed plan, a bad coin flip could take you outside K. Feedback lets you correct course.
Hex: So feedback handles surprises. What about the playlist?
Lux: Empowerment takes a different approach entirely. Fix a starting state s-zero. Fix a horizon H — say, two time steps. Now consider all possible action sequences of length H. Each sequence is a playlist: action a-zero, then action a-one. You press play and let the system run. At the end, you observe the output through a lens f — the outside macrostate, the position on the ring, whatever the output variable is.
Hex: And different playlists produce different outputs.
Lux: Sometimes. That's what empowerment measures. It asks: how many distinguishable outputs can I produce by choosing different playlists? Formally, it's channel capacity. The input is the action sequence. The output is the observed macrostate. The channel is the stochastic dynamics. Empowerment equals the maximum mutual information between input and output, optimized over the distribution of playlists.
Hex: That's information theory. Shannon's channel capacity.
Lux: Computed exactly for small channels via Blahut-Arimoto. The Throw paper adds one constraint: feasibility. Not every playlist is affordable. The total cost of the action sequence has to fit within the budget at the starting state. So feasible empowerment is the capacity of the restricted channel — only the playlists you can actually pay for.
Hex: So the joystick lets you react. The playlist makes you commit. Why can't empowerment use feedback too?
Lux: Because channel capacity is defined over a fixed input alphabet. If the inputs change depending on intermediate outcomes, the channel isn't a fixed object anymore. You'd be computing capacity over a tree of possible input-output histories, which is a different — and much more complex — quantity. The Throw paper deliberately keeps empowerment simple and auditable: fix the input alphabet, compute capacity, compare across states and conditions.
Hex: So it's a simplification. But a useful one?
Lux: Extremely useful. Because the paper knows what empowerment is for. It's not trying to measure everything. It's trying to measure difference-making — how much causal influence the agent's action channel has on the external world. And for that, open-loop channel capacity is a clean, well-understood proxy. The paper even names it explicitly: open-loop budgeted empowerment.
Hex: And acknowledges the limitations?
Lux: Directly. The paper notes that open-loop empowerment can over- or under-approximate stricter stepwise feasibility. It treats the measure as an operational proxy and includes calibrated nulls to guard against mistyped control channels. If the null regime shows high empowerment, something is wrong — the system is being credited with influence it doesn't have.
Hex: Now the fun part. The lab experiments. Where do the joystick and playlist agree?
Lux: Experiment one. Imagine the viability kernel is large — lots of states pass the survival test. Noise is low. All five actions are feasible at every state. In this regime, the joystick has plenty of room to steer. And the playlist has many songs to choose from, each producing a distinct output. Both measures are high. Viability says: the agent exists and can survive from many starting points. Empowerment says: the agent's choices make a big difference to the output.
Hex: No tension between the two.
Lux: But now turn the knobs. Experiment two, case A. The viability kernel is empty — K equals the empty set. No state has a feasible policy that guarantees survival. But suppose someone computes empowerment on the full state space anyway, ignoring viability. The empowerment number might look impressive — the channel has high capacity.
Hex: But the agent doesn't exist.
Lux: The joystick is disconnected. There's no viable state to start from. The empowerment number is meaningless because it's computed on a domain where the layer doesn't persist. This is a misdiagnosis. And it's why the paper computes empowerment only on the viable domain — median empowerment on K. If K is empty, empowerment is zero by convention.
Hex: So viability gates empowerment. You need the joystick to be connected before the playlist matters.
Lux: Case B is the mirror image. The viability kernel is full — every safe state survives. The joystick works perfectly. But empowerment is zero. Every playlist produces the same output. The agent can survive, but it can't make a causal difference. All playlists sound identical.
Hex: A joystick that keeps you alive but doesn't let you steer anywhere interesting.
Lux: The paper's interpretation: this is an agent with agenthood — maintained existence — but without agency — causal difference-making. The two notions split. And this is exactly why the paper separates them. Agenthood is the enablement claim: the layer exists and persists. Agency is the causal claim: interventions change outcomes.
Hex: Let me push on this. When does feedback actually outperform open-loop? Is there a concrete scenario?
Lux: Protocol holonomy. The Plot paper's E2 exhibit. On a curved substrate — think of a sphere rather than a flat grid — composing local moves around a small loop produces a residue. You go north, then east, then south, then west, and you don't end up where you started. The rotation angle — the holonomy — is nonzero.
Hex: Because the order of moves matters on a curved surface.
Lux: Exactly. P3 in the emergence calculus: protocol order matters. Now imagine an agent navigating this surface. A feedback controller — the joystick — can detect the holonomy as it accumulates and adjust mid-loop. It sees the drift and corrects. An open-loop controller — the playlist — committed to its sequence in advance, can't adapt. It runs the same moves regardless of what the curvature does.
Hex: So on curved substrates, feedback has a structural advantage.
Lux: The Plot paper measures this. Median holonomy on a plane-like substrate: 0.0479 radians — almost zero. Median holonomy on a sphere-like substrate: 0.5980 radians — more than twelve times higher. The curvature-induced order dependence is real and measurable. And it creates a gap between what feedback and open-loop control can achieve.
Hex: That's a clean lab result. But why does the Throw paper use both if feedback is more powerful?
Lux: Because they measure different things, and both measurements are needed. The Become paper's three-certificate loop makes this explicit. The first certificate is closure coherence — does the packaging stabilize? That maps to viability. The second is audit monotonicity — does coarse observation preserve distinguishability? That maps to the information-theoretic channel that empowerment measures. The third is route behavior — do different packaging routes commute?
Hex: Three certificates, not one. And feedback and open-loop map to different certificates.
Lux: Viability is the stability certificate: can the layer persist? Empowerment is a proxy for the audit certificate: does the action channel carry information? Neither subsumes the other. A system can be stable without being informative. A channel can carry information on a domain that doesn't persist. The paper calls them complementary proxies and uses them in pairs.
Hex: And the core Six Birds paper says this structure is unavoidable?
Lux: Section nine of the core paper argues that any layer satisfying minimal richness must exhibit all six primitives. The feedback-versus-open-loop distinction maps to the difference between P5 — closure and packaging, which requires the joystick-style maintenance — and P6 — accounting, which requires the playlist-style measurement of what resources flow through the interface. You can't collapse them because they govern different aspects of the layer.
Hex: The joystick keeps the layer alive. The playlist measures what the layer can do.
Lux: And in the Throw paper's ring-world, you can watch both gauges on the same dashboard. The noise-maintenance sweep plots viability kernel size and median feasible empowerment at every grid point. When both are high, the agent is robust and influential. When viability collapses, empowerment goes to zero by convention. When empowerment is zero but viability is positive, the agent exists but is impotent.
Hex: Two gauges, one dashboard. Joystick for survival, playlist for influence.
Lux: And the mini-lab result: you need both. Drop either gauge and you lose half the picture. The emergence calculus insists on paired diagnostics because a single number can't capture both persistence and difference-making.
Hex: Lab closed. Both instruments stay on the bench.