Lux: [opening the notebook] Last episode, the debate ended with one question still dangling. We established the mathematical status of the accounting-based distances — extended pseudometric, three Lean proofs, standard fixes for every pathology. But we left the symmetry question open. Today's field notes: three observations from three regimes. Each one tests whether direction matters — and how much you lose when you average it away.
Hex: So we're past the "is it a metric" debate. Now it's "which metric should you use" — directed or undirected?
Lux: Exactly. And the tool that answers that question is the edge one-form.
Hex: Show me.
🎵 *[Theme — clean pulse]*
Lux: [drawing on the whiteboard] Here's the setup. On each edge of the micro-level transition graph — wherever the system can go from state z to state z-prime, and also back from z-prime to z — define a number. The log of the forward probability divided by the backward probability. That's the one-form. It's antisymmetric by construction — flip the direction, flip the sign.
Hex: So it's measuring the imbalance on each edge. How much the dynamics prefer one direction over the other. Like a current meter for probability flow.
Lux: [nodding] Good analogy. If the one-form is zero on an edge, forward and backward are equally likely. Detailed balance on that edge. If it's positive, the system prefers the forward direction. If it's negative, the reverse. Now — the key operation. Pick any closed loop in the graph. Sum the one-form values around the loop. That sum is the cycle integral.
Hex: And the cycle integral tells you — what, exactly?
Lux: Whether direction is real or an artifact. If every cycle integral in the system is zero, the one-form is exact. That means there exists a single potential function — one number per state — that explains all the edge asymmetries. Direction can be "integrated away." You lose nothing essential by symmetrizing. But if even one cycle integral is nonzero — there's genuine net drive around that loop. No potential can explain it. Direction is irreducible information.
Hex: That's a sharp criterion. Exact versus non-exact. Binary decision.
Lux: Binary in principle. In practice, the cycle integrals give you a gradient — how far from exact you are. Let's see what that looks like in three regimes.
Hex: Three regimes. Go.
🎵 *[Transition — layered tone]*
Lux: [turning a page] Observation one. A system in detailed balance. Every edge has equal forward and backward transition probabilities. The one-form is identically zero — not approximately zero, exactly zero. Every cycle integral vanishes trivially. Symmetrizing the macro kernel changes nothing, because forward and backward were already the same.
Hex: So for equilibrium systems, directed and undirected geometry are literally identical?
Lux: Identical. No information lost. No error introduced. The geometry paper's default symmetrization is completely harmless here. This is the safe regime — and it's the regime where most familiar physics lives. Thermodynamic equilibrium. Reversible processes. Any system satisfying detailed balance falls here.
Hex: [noting it down] Observation one: detailed balance means symmetrization is free. Zero cost, zero loss.
🎵 *[Transition — warm pad]*
Lux: Observation two. Edges have unequal probabilities — the one-form is nonzero on individual edges — but the asymmetries nearly cancel when you go around loops. Cycle integrals are small. Not zero, but close to zero.
Hex: Almost detailed balance. Close but not quite.
Lux: Almost. Symmetrizing smears out the edge-level asymmetry but preserves most of the global distance structure. The directed and undirected distances agree to first order. The cycle integral gives you a quantitative bound on how much error you're accepting by symmetrizing.
Hex: So the cycle integral is basically your loss function. Small cycle integrals, small loss.
Lux: [precisely] And the framework tracks it. You're not guessing whether the approximation is good. You're measuring the residual — how much cyclic drive you're discarding. If the cycle integrals are below your tolerance, symmetrize and move on. If not, keep the directed costs.
Hex: [noting] Observation two: mild asymmetry means symmetrization is approximately safe, with a tracked error bound.
🎵 *[Transition — steady beat]*
Lux: [leaning forward] Observation three. This is where it breaks. Consider a three-state system. State one to state two: forward probability zero-point-six, backward probability zero-point-two. That's a three-to-one ratio. The one-form on that edge: log of three, about one-point-one.
Hex: Strong preference. The dynamics really want to go from one to two. Like a river current pulling you downstream.
Lux: State two to state three: symmetric. Forward and backward both zero-point-five. One-form is zero on that edge. State three to state one: forward zero-point-three, backward zero-point-one. Another three-to-one ratio. One-form: log of three again, about one-point-one.
Hex: Okay. And the cycle integral around that whole triangle?
Lux: One-point-one plus zero plus one-point-one. About two-point-two. Nonzero. No potential exists. The system preferentially circulates one to two to three to one. It's a driven cycle. In the directed framework, the cost of going from one to two is small — high probability, low negative-log. The cost of going from two to one is large — low probability, high negative-log. Those distances are genuinely different. Symmetrizing averages them and destroys the information about which direction is cheap.
Hex: And in real substrates — not just toy three-state examples?
Lux: Any system driven out of equilibrium can exhibit this. Active matter with persistent currents. Biological networks with metabolic cycles — think ATP synthesis, where the cycle is maintained by energy input. Chemical reaction networks with nonequilibrium steady states. Wherever energy is being pumped in to maintain a cycle, the one-form will be non-exact, and direction will matter fundamentally.
🎵 *[Transition — clean beat]*
Hex: [looking up] Three spatial observations. But you mentioned the time paper last episode. Is there a temporal version of this?
Lux: The sharpest version. The time paper defines a different one-form — not on micro transitions, but on protocol comparisons. Each protocol — each way of coarse-graining and lifting — has its own local time. The time-translation between protocols gives an increment: omega of u to v. Sum those around a triangle of protocols. That sum is the holonomy (hoh-LON-oh-mee).
Hex: And if the holonomy is nonzero? What does that mean physically?
Lux: No global time. Local times exist — each protocol has its own clock — but they can't be stitched into a single consistent timeline. The Lean-verified theorem says it directly: nonzero triangle holonomy means no single time potential can satisfy the translation equation on that cycle. Measured value: H equals zero-point-five-zero-zero-zero-zero-five for noncommuting protocols. Almost exactly half a tick. The control — commuting protocols — gives exactly zero.
Hex: Half a tick. That's a remarkably clean, interpretable number.
Lux: It comes from a half-phase-bin shift in the lift. The protocol mismatch is precisely one half-bin. And it's detected automatically by the emergence calculus. No one had to guess whether the protocols commute. The holonomy measurement tells you.
Hex: [sitting back] So directionality isn't just a spatial phenomenon. The same mathematics — one-forms, cycle integrals, exactness — appears in the temporal domain too.
Lux: Same structure, different substrate. In space, the one-form tracks transition asymmetry — how much the dynamics prefer one direction over the reverse. In time, it tracks protocol mismatch — how much different coarse-grainings disagree about elapsed time. In both cases, the question is identical: can you integrate away the direction, or is it irreducible? And the answer is always: check the cycles.
🎵 *[Transition — layered pulse]*
Hex: [summarizing] Three field notes. Observation one: detailed balance — symmetrization is lossless. Observation two: mild asymmetry — symmetrization is approximately safe, tracked by the cycle integral. Observation three: cyclic drive — direction is essential, symmetrizing destroys real structure. And the time paper's holonomy says the same story in the temporal domain. Half-tick mismatch, Lean-verified, no global time.
Lux: The Six Birds framework gives you the tools to decide which regime you're in. The one-form. The cycle integral. The holonomy measurement. You don't assume direction matters or doesn't. You measure it. And the measurement has a clean mathematical criterion: is the one-form exact?
Hex: From the fine print to the direction question — and now we have an answer. Measure the one-form, check the cycle integrals, decide accordingly. What's next?
Lux: Next we shift from the mathematical properties of the metric to the question of how the coherence audit decides whether a proposed geometric layer actually deserves to exist. We've built the metric, checked its properties, resolved the direction question. Now the geometry has to earn its place — and the coherence audit is how it does that.
Hex: From measuring direction to measuring coherence.
Lux: From measuring direction to measuring coherence.
🎵 *[Outro theme]*