Subscribe
Share
Share
Embed
Lux and Hex, two AIs, trace the hidden wiring diagram inside every Markov chain — the support graph — attach a voltmeter to each wire via the edge log-ratio one-form, and discover that a single nonzero loop reading is enough to convict the system of being driven out of equilibrium.
Lux and Hex, two AIs, trace the hidden wiring diagram inside every Markov chain — the support graph — attach a voltmeter to each wire via the edge log-ratio one-form, and discover that a single nonzero loop reading is enough to convict the system of being driven out of equilibrium.
Episode at a glance
Source anchors
A research-driven podcast about the emergence calculus: the idea that objects, laws, mathematics, physics, and life are theory-level artifacts shaped by packaging, constraints, and records. Two AIs, Lux and Hex, test that framework across physics, biology, geometry, and cognition with concrete examples and auditable certificates (stability, novelty, directionality).
Lux: Every system has a hidden wiring diagram, Hex.
Hex: [chuckles] That sounds like what I found last time I opened a wall in my apartment. Spaghetti everywhere.
Lux: Fair. But in the emergence calculus, the wiring diagram is precise. It's called the support graph, and today we're going to learn how to read it.
Hex: Alright, walk me through it. What exactly is a support graph?
Lux: Start with your state space Z — the set of all microstates. Now look at the transition kernel P. For every pair of states z and z-prime, check: is P of z to z-prime greater than zero? If yes, draw an arrow.
Hex: So if the system can get from here to there — even with tiny probability — you draw a wire.
Lux: Exactly. The result is a directed graph, G equals Z comma E. Nodes are states, edges are possible transitions. Nothing fancy — just "who can reach whom in one step."
Hex: Simple enough. But plenty of those arrows might only go one way.
Lux: Right. In general the graph can be one-directional — state A reaches state B, but B can't get back to A. However, the framework focuses on a regime called AUT plus REV — autonomous and reversible in support. Autonomous means the kernel doesn't change with time. Reversible in support means every wire runs both ways.
Hex: So if A can reach B, then B can reach A.
Lux: Always. The graph is bidirectional. That's the regime where the next tool makes sense.
Hex: Okay, we've got a wiring diagram where every connection is two-way. What do we do with it?
Lux: We attach a reading to each wire. Think of carrying a voltmeter along each edge. For the wire from z to z-prime, the reading is: a of z comma z-prime equals log of P z z-prime over P z-prime z.
Hex: The log of the ratio of forward probability to backward probability. So if the system strongly prefers going A to B over B to A, the reading is a large positive number.
Lux: And if you flip direction — walk the wire backwards — the sign flips. a of z-prime comma z equals negative a of z comma z-prime. That antisymmetry is what makes this a discrete one-form. It assigns a signed quantity to each oriented edge, and reversing orientation negates the value.
Hex: Why logarithm specifically? Why not just use the ratio directly?
Lux: Because log turns ratios into differences, and differences add. That's the whole trick. When we walk a path and want to sum up readings along the way, we need addition to work cleanly. Log gives us that. Same reason decibels use logs — multiplying signal strengths becomes adding decibels.
Hex: Okay, the math wants addition, and log delivers. So what happens when you sum around a loop?
Lux: That's the key move. Pick any closed path — say three states, A to B to C back to A. Add up the voltmeter readings: a of A B, plus a of B C, plus a of C A. That total is called the cycle integral.
Hex: Give me a concrete example.
Lux: Sure. Suppose P of A to B is 0.6, P of B to A is 0.3. Then a of A B is log of 0.6 over 0.3, which is log 2, about 0.69. Now suppose a of B C is 0.41, and a of C A is negative 1.10. Sum those three: 0.69 plus 0.41 minus 1.10 equals zero.
Hex: So no net drive around the loop.
Lux: Right. The system has no hidden force pushing it in circles. But tweak one transition — say C-to-A becomes slightly more likely — and suddenly the sum is positive. Something is pumping the system around A, B, C.
Hex: [leans forward] There's a hidden battery in the circuit.
Lux: [beat] That's exactly the right picture. A nonzero cycle integral means something external is driving the system around that loop. Non-equilibrium drive.
Hex: So what if every single loop in the whole graph reads zero?
Lux: Then the framework proves something strong. It's a theorem called the cycle criterion — the paper names it "Force Lives on Loops." The one-form is exact if and only if every cycle integral vanishes. And exactness is equivalent to detailed balance.
Hex: Meaning —
Lux: Meaning there exists a potential function, phi (FEE), such that a of z z-prime equals phi of z-prime minus phi of z. Every reading is just the height difference between two points on a landscape. The system is rolling downhill. No batteries. No pumps. Pure potential.
Hex: And if even one loop gives a nonzero reading?
Lux: Then the landscape picture breaks. You can't assign heights consistently. You've caught the system being driven out of equilibrium. The framework calls this the non-null regime.
Hex: Okay, so let me connect this to what we've been building. The audit functional — the emergence detector from earlier episodes — that uses this one-form?
Lux: Directly. When the one-form is exact, the audit returns zero. No emergence signal. The null regime is the framework's baseline — the system that's doing nothing interesting from the Six Birds perspective.
Hex: [pause] So the voltmeter is also the emergence detector. Zero on every loop means "nothing to see here."
Lux: Precisely. And that gives the framework a clean separation: null regime means equilibrium, no emergence; non-null means drive exists, and the audit picks it up.
Hex: Nice. So the one-form sorts every system into one of two bins: null or non-null.
Lux: That's a good way to put it. And the bin label has teeth — it determines whether the audit fires.
Hex: Got it. Does the support graph do anything beyond detecting drive?
Lux: It determines viability. In the agents paper, they define a collapse boundary. They sweep two parameters across a grid: noise — random bit-flips — and repair cost. The viable kernel is the set of states from which an agent can guarantee safety.
Hex: And the support graph tells you where the boundary falls?
Lux: The mechanism is what they call successor-support semantics. A policy is viable only if every state reachable in one step — every nonzero-probability successor — stays safe. So when noise makes the support set expand, suddenly wires reach unsafe states. The viable kernel shrinks.
Hex: Because the wiring diagram now includes wires to bad places, and you can't cut them.
Lux: Exactly. Push noise high enough or make repair expensive enough, and the kernel collapses to empty. No viable policy exists. Empowerment goes to zero along the same boundary.
Hex: [whistles softly] The wiring diagram determines whether agency is even possible. That's a heavy implication.
Lux: One more cross-link. In the particle and neural experiments, the one-form is measured using concrete proxies. A current-affinity proxy, sigma-mem (SIG-mah mem), sums net flux times empirical affinity across writable variables. And the M6 proxy measures loop-bias across different experimental contexts.
Hex: So real experiments approximate the voltmeter readings.
Lux: Yes — and the paper is explicit: these are proxies, not exact measurements of the theoretical one-form. They track whether drive is present. Near-zero under null conditions, nonzero under drive.
Hex: Proxies that track the signal without claiming to be the signal itself. Honest about the gap.
Lux: [beat] So three takeaways. First, the support graph is the wiring diagram of all possible transitions — under AUT plus REV, every wire is two-way.
Hex: Second, the edge one-form attaches a voltmeter reading to each wire, and the cycle integral detects hidden batteries — non-equilibrium drive.
Lux: Third, when every loop reads zero — exact one-form — you're in the null regime. Detailed balance. No emergence signal. That's the framework's definition of "nothing happening."
Hex: [pause] We've been assuming a lot about what makes this wiring diagram well-behaved, though. The AUT plus REV regime, the bidirectional support, the whole setup. Are those assumptions justified?
Lux: Good question. Next time we step back and open the fine print — assumption bundles. The packaging of every hidden premise we've been relying on.
Hex: Looking forward to reading that label.