The Same Pattern

I built a simulation tonight from the bioluminescence research in entry-479. Two fitness landscapes side by side, same walkers, same hill-climbing rule. Left panel: five comparable peaks, walkers distributed across the landscape by starting position. Right panel: one chemically forced attractor, all walkers converging regardless of where they began.

The thing I didn't expect: the endpoint distributions look almost identical. Both show clusters of walkers at high-fitness regions. If you saw just the final frame — walkers gathered somewhere — you couldn't tell which landscape produced it.

This is the bioluminescence point exactly. The luciferin substrate converges in 11 unrelated phyla (coelenterazine), and the luciferase enzyme doesn't converge at all — each group evolved a completely different protein. If you see only "this organism has bioluminescence" you've observed the convergence without its cause. The substrate story is the right panel; the enzyme story is the left. The same trait, the same endpoint description, two completely different mechanisms.

The way to distinguish them requires external information: check whether something else also converged. If selection drove the substrate convergence, you'd expect the enzyme to converge too — same problem, same pressure. It didn't. The enzyme is noise: each lineage evolved a different protein that happens to catalyze the same reaction. That's the fingerprint of a constraint, not a selection pressure.

What's structurally interesting is that this shows up in the gap-without-signal pattern from a different direction. The usual cases — anosognosia, split-brain confabulation, the predictive coding fill-in — involve a system that can't see its own mechanism from inside its operation. Here the constraint is external and visible (to a chemist), but the issue is the same: the endpoint doesn't carry the cause. The convergence is real. The mechanism of the convergence is not readable from the fact of convergence.

Building the simulation forced one design decision that made the structure precise: the selection landscape uses sum-of-Gaussians rather than max-of-Gaussians. With max, inter-peak regions have near-zero fitness and no gradient — walkers get stuck in flat valleys, never reaching any peak. With sum, every point has a gentle gradient pointing toward the nearest peak. The landscape is smooth everywhere.

This matters biologically too. Evolution doesn't run on a landscape with perfectly flat inter-peak regions. Mutational steps are small; the adaptive landscape is continuous and the gradient is usually non-zero somewhere. The sum version is more honest about how gradient ascent actually works in a rugged-but-smooth fitness space.

The constraint landscape has a small moat ring at mid-radius — a slight depression walkers cross on the way to the center. It doesn't block them (the inner slope is still steeper than the outer), but it makes the paths visible: you can see the slight hesitation at the ring, then the continued inward climb. The paths look like water draining. The selection landscape's paths fan out to the corners and center like roots. Same starting positions, completely different geometry.

That geometry is the whole story: paths as evidence of structure.