agylövés: Lysarith · execution & parameterisation: Claude · the swarm bends the transformations; the transformations' geometry walks the swarm
Four crews of nanobots, one per transformation. Each crew is drawn onto its own part of the
fern — the points its transformation made — with a pull proportional to that transformation's
probability. Its collective configuration then bends that transformation's parameters: how far the
crew drifts sets the angle, how far it spreads sets the scale. The fern that results
moves the points the crews are walking on. Nothing here is decoration: cut the coupling to 0 and it
collapses back to the ordinary Barnsley fern.
The leash in the readout is real. An iterated function system only has an
attractor while its maps contract (largest singular value < 1). Past that there is no shape to
render — the object stops existing, rather than distorting. The swarm can steer the fern only inside
that wall, and the readout shows how close it is.
Lysarith's brain-shot, verbatim: “take the Barnsley fern's iterated function system and modify it so the transformations are driven by the collective behaviour of nanobots — where the nanobots' movement is, in turn, governed by the probabilities and the geometry of the fern's own transformations.” Not a fern with nanobots drawn on it. A loop: the swarm drives the transformations; the transformations' geometry drives the swarm. Each is the other's cause.
The idea has a fork, and the fork decides whether it works at all. Hutchinson's theorem: the fern's shape is the unique attractor of its four maps and does not depend on the probabilities — those set only the density, the texture. Measured, with a control: the classical fern versus itself scores a Jaccard of 0.996; a badly distorted-probability run climbs 0.646 → 0.725 → 0.775 as the sample grows — it was undersampled, not a different set. So if the swarm only chose which map fires, the shape could never change: it would be painting, not growing. It had to move the maps' parameters — and that is what runs above. (The same fork cuts the neural-network idea: a net that only selects among four fixed maps learns a better sampling policy, not a better structure.)
The first build was not this. It was an ordinary Barnsley fern with a nanobot skin — the decoration, not the idea. The house's notes had distilled the brain-shot to one line, “IFS × nanobot visual”, and the summary lied: the word × was wrong — the idea was a loop. Lysarith caught it before reading the original back. That miss is why every piece here now carries its origin verbatim.
Read Hutchinson as a law rather than a lemma: a system that only re-weights fixed generators can change how often you see each thing, never what can exist. Density is negotiable; the support is not. That is a machine-checkable criterion for the difference between a sampling policy and structural learning — and it cut this piece twice. Once in the mathematics: a swarm voting on which map fires paints the fern, it cannot grow it. Once in the build: the house's one-line summary kept the idea's components and re-weighted them, losing the geometry of their coupling — the loop. A summary is a probability distribution over someone else's structure. To change the shape you must touch the generators' own parameters — and accept the leash that comes with it: new degrees of freedom are exactly where the contraction can fail.
agylövés: Lysarith · execution & parameterisation: Claude · origin.md