Lux: Last episode we looked at the config format — YAML extension, JSON content, three mechanisms for determinism. Today we open the actual config files and read the parameter tables. Episode one fifty-nine — the major knobs.
Hex: From the recipe format to the actual recipe.
Lux: Exactly. And here's the key insight: six experiments, but most of the knobs are shared. The geometry paper uses a common configuration block across all four main exhibits.
Hex: Same control panel, different substrates.
Lux: Seven shared settings. Staging parameter tau equals five. Lens levels — a ladder from four up to one hundred twenty-eight macro states, doubling at each step. Number of eigenvectors for diffusion coordinates: six. Prototype choice: uniform. Symmetrization (SIM-eh-trih-ZAY-shun) method: weight average. Cost smoothing constant eta at ten to the minus twelve. Edge threshold epsilon at ten to the minus fifteen.
Hex: That's a lot of knobs. What does each one actually control?
Lux: Start with tau. Tau is the staging parameter — how many micro-steps the system takes before you repackage into macro labels. Think of it as integration time. How long do you let the dynamics run before you take a snapshot? Tau equals five means five steps of the micro kernel before each macro observation.
Hex: And if you change it?
Lux: The paper is explicit about this. Too small — you get under-mixed, noisy macro costs. The system hasn't had enough time to settle, so your cost estimates are jittery. Too large — prototype drift. The prototypes that define your macro points start drifting from the actual dynamics. You wash out the distinctions the geometry is supposed to capture.
Hex: The Goldilocks parameter (GOL-dee-loks). Not too hot, not too cold.
Lux: And there's no universal optimum. Different substrates might want different tau values. But for the canonical exhibits, five works. The diagnostics confirm it.
Hex: What about the lens levels?
Lux: The lens ladder defines how many macro states you construct at each resolution — four, eight, sixteen, thirty-two, sixty-four, one hundred twenty-eight. This is the packaging resolution. At four macro states, you're seeing the coarsest structure. At one hundred twenty-eight, the finest. The paper runs all levels and checks whether the geometry is consistent across the ladder.
Hex: Coarse to fine. Like zooming in on a map.
Lux: And n-eigs — six eigenvectors — controls how many dimensions of the diffusion embedding you use for clustering. More eigenvectors captures more structure but adds computational cost. Six is enough to separate the main geometric features across all four substrates.
Hex: What about the tiny numbers — eta and epsilon?
Lux: Eta is the cost smoothing constant. Remember, the directed cost between macro states is computed as negative log of the transition probability plus eta. Eta at ten to the minus twelve means you're adding an almost invisible floor — just enough to keep the logarithm from blowing up at zero transitions, without distorting the real costs.
Hex: And epsilon?
Lux: The edge threshold. If a transition probability is below ten to the minus fifteen, the pipeline treats it as zero — no edge in the cost graph. This prevents phantom connections from numerical noise.
Hex: That's the shared block. What changes between experiments?
Lux: The substrate. Exhibit one: a grid random walk. Twenty-five by twenty-five grid, six hundred twenty-five micro-states. Lazy parameter zero-point-five — meaning half the time the walker stays put. This gives you a flat, plane-like geometry.
Hex: Exhibit two?
Lux: A sphere point cloud. Five hundred points sampled on a sphere, connected by k-nearest neighbors with k equals ten. Gaussian weight falloff with sigma zero-point-five. A tiny self-loop at ten to the minus six for numerical stability. The key fact: coordinates are used only to generate the micro graph. Closure never accesses them. The geometry emerges from dynamics alone.
Hex: Exhibit three — the fractal.
Lux: Sierpinski (seer-PIN-skee) gasket at recursion level five. Three hundred sixty-six vertices. Lazy random walk, same as the grid. The fractal has perfect self-similarity at each scale, so the lens ladder has to capture structure that repeats differently at each resolution.
Hex: And exhibit four — the constrained case.
Lux: Same twenty-five by twenty-five grid, but with an anisotropic gate. Direction east, strength one-point-zero. This blocks or penalizes transitions in one direction. The constraint deforms the geometry — and the numbers show it. Delta rises from zero-point-three-two on the plain grid to zero-point-four-one under the gate. Distortion nearly doubles — six-point-four-nine to eleven-point-two-five.
Hex: Same substrate, same observer, but add a constraint and the geometry degrades.
Lux: Which is exactly what the emergence calculus framework predicts. Constraints don't just change the geometry — they stress it. The diagnostics measure the stress.
Hex: What about the holonomy (hoh-LON-oh-mee) and Pythagorean configs? Those are separate?
Lux: Separate config files with specialized parameters. The holonomy demo uses the sphere substrate but adds loop-sampling settings. Neighborhood size k-neigh equals twenty-four. Expansion hops: one. Minimum overlap: four. Maximum eight hundred loops. Seeds zero and one for plane and sphere respectively.
Hex: Those parameters control how you estimate curvature?
Lux: Exactly. K-neigh defines the local patch. Overlap threshold ensures neighboring patches share enough structure for meaningful comparison. And the loop count determines statistical power — eight hundred triangles gives you enough to distinguish holonomy (hoh-LON-oh-mee) of zero-point-zero-four-eight on the plane from zero-point-five-nine-eight on the sphere.
Hex: Two orders of magnitude apart. That's a clean signal.
Lux: The Pythagorean config is different again. A two-dimensional torus random walk with five hundred twelve micro-states. Tau ranges from four through one hundred twenty-eight — not a single value but a sweep. Displacement window parameters: D-factor three, D-max thirty. Two thousand right-triangle samples. The question: does the emergent metric satisfy the Pythagorean theorem? And the answer depends on tau. At tau equals four, the median Pythagorean residual is thirty-three-point-two. At tau one hundred twenty-eight, it drops to zero-point-zero-six.
Hex: From wildly non-Pythagorean to essentially Euclidean. Same substrate, different staging.
Lux: And this knob vocabulary isn't unique to the geometry paper.
Hex: The other papers have their own knobs?
Lux: The agency paper has a beautiful example. A discrete skill variable theta — takes values zero, one, or two. This is a direct P-one knob. It changes the effective transition kernel — higher theta means fewer unintended outcomes from left-right moves. With action costs set to zero, empowerment goes from zero-point-seven-three-three bits at theta zero to one-point-three-four-two bits at theta two.
Hex: So turning one knob — the skill variable — nearly doubles the agent's effective control.
Lux: And the life paper has substrate knobs for the neural instantiation. Per-site symbol variables. Auxiliary bounded fields. Two classes of operator tokens — within-layer tokens W with a global budget constraint, and inter-layer tokens K with a per-site budget constraint. Different vocabulary, same idea: each knob corresponds to a Six Birds primitive being instantiated at a specific value.
Hex: Same mixing desk, different genres of music.
Lux: Which brings us to the last question. You've set all the knobs. How do you know if the combination works?
Hex: How do you know when the knobs are set wrong?
Lux: The diagnostics. Delta — the idempotence (eye-dem-POH-tence) defect — tells you how far the closure operator is from being a true projection. Stability measures whether small perturbations in the micro-states produce large jumps in macro labels. Distortion compares emerged distances to reference distances. Connectivity counts unreachable pairs. These are your warning lights.
Hex: A cockpit instrument panel. Everything green means the geometry is coherent.
Lux: And when a knob is wrong, a light goes amber. Tau too small — delta spikes. Resolution too fine — distortion climbs. Gate too strong — connectivity drops. The paper doesn't just report the good settings. It documents what happens at the boundaries.
Hex: That's the honest part. Not just "here's what works" but "here's where it breaks."
Lux: And the foundations paper grounds all of this. The self-generation theorem says: given a process soup, an interface lens, a refinement family, and a bounded-interface assumption, the six primitives appear canonically. Each knob in the config is one of those instantiation choices. There's no "optimal" setting in some universal sense. There's only: does this particular combination produce a coherent closure?
Hex: [quiet] Not hyperparameters. Instantiation choices. That's a fundamentally different framing.
Lux: Hyperparameters are tuned against a loss function. Instantiation choices define which layer of the emergence calculus you're constructing. The diagnostics don't say "good" or "bad." They say "coherent" or "not coherent at this resolution."
Hex: [pause] The honest instrument panel. It doesn't judge. It just reports what the knobs produced.
Lux: Canonical configuration snapshot. The major knobs of emergent geometry.
Hex: Next time?
Lux: Episode one sixty. How this paper was built. The Plot paper's development story — from first substrate to final exhibit. Not the science, but the engineering journey. How the pipeline grew, what failed, and what survived.
Hex: From the knobs to the workshop. See you there.